Hindawi Publishing CorporationJournal of Medical EngineeringVolume 2013 Article ID 104684 13 pageshttpdxdoiorg1011552013104684

Research ArticleHybrid Discrete Wavelet Transform and Gabor Filter BanksProcessing for Features Extraction from Biomedical Images

Salim Lahmiri and Mounir Boukadoum

Department of Computer Science University of Quebec at Montreal 201 President-Kennedy Local PK-4150Montreal QC Canada H2X 3Y7

Correspondence should be addressed to Salim Lahmiri lahmirisalimcourrieruqamca

Received 13 December 2012 Revised 12 March 2013 Accepted 27 March 2013

Academic Editor Ying Zhuge

Copyright copy 2013 S Lahmiri and M Boukadoum This is an open access article distributed under the Creative CommonsAttribution License which permits unrestricted use distribution and reproduction in any medium provided the original work isproperly cited

A new methodology for automatic feature extraction from biomedical images and subsequent classification is presented Theapproach exploits the spatial orientation of high-frequency textural features of the processed image as determined by a two-stepprocess First the two-dimensional discrete wavelet transform (DWT) is applied to obtain the HH high-frequency subband imageThen a Gabor filter bank is applied to the latter at different frequencies and spatial orientations to obtain new Gabor-filteredimage whose entropy and uniformity are computed Finally the obtained statistics are fed to a support vector machine (SVM)binary classifier The approach was validated on mammograms retina and brain magnetic resonance (MR) images The obtainedclassification accuracies show better performance in comparison to common approaches that use only the DWT or Gabor filterbanks for feature extraction

1 Introduction

Computer-aided diagnosis (CAD) has been the subject of alot of research as a tool to help health professionals inmedicaldecision making As a result many CAD systems integrateimage processing computer vision and intelligent and sta-tistical machine learning methods to aid radiologists in theinterpretation ofmedical images and ultimately help improvediagnostic accuracy These systems have been employed toanalyze and classify various types of digitized biomedicalimages including retina [1 2] mammograms [3ndash5] brainmagnetic resonance images [6ndash8] skin cancer images [9 10]lung images [11 12] and ulcer detection in endoscopy images[13 14] just to name a few

The typical CAD process starts with a segmentation stageto identify one or more regions of interest (ROI) in theimage of interest Then the ROI(s) is processed for imageenhancement andor feature extraction before classificationBecause the segmentation step requires prior knowledge ofdiscriminant image features and its implementation typicallycalls for numerous parameter settings recent works have

attempted to eliminate it These approaches realize featurespace reduction by applying one or more transforms to thewhole image and extracting the feature vector to classify fromone or more of the obtained components [3 5 7ndash14]

Texture analysis has played an important role in the char-acterization of biomedical images Texture analysis methodscan be categorized as statistical geometrical and signalprocessing types [14] Statisticalmethods aremainly based onthe spatial distribution of pixel gray values while geometricalapproaches depend on the geometric properties of textureprimitives As for signal processingmethods they use texturefiltering in the spatial or frequency domain to extract relevantfeatures

Multiresolution analysis is the most widely employedsignal processing technique for characterizing biomedicalimages due to its capability to obtain high time-frequencyresolutions The wavelet-transform family methods are typi-cal examples ofmultiresolution analysis techniquesThe basicwavelet transform [15 16] starts with a basis function themother wavelet and decomposes a signal into componentsof different time and frequency scales longer time intervals

2 Journal of Medical Engineering

are used to obtain low-frequency information and shorterintervals are used to obtain high-frequency information

The most commonly used wavelet transform in biomedi-cal image processing is the discrete wavelet transform (DWT)[14] whose discrete time shifting and stretching variables leadto a sparse and efficient representation The DWT takes aninput image and decomposes into four subimage componentsthat characterize it for different orientations in the horizontaland vertical frequency axesThe process can be repeated withone or more subimages if needed More precisely the DWTdecomposition yields the approximation subband (LL) thehorizontal detail subband (LH) the vertical detail subband(HL) and the diagonal detail subband (HH) These describerespectively the low-frequency components in the horizontaland vertical directions the low-frequency components inthe horizontal direction and high-frequency components inthe vertical direction the high-frequency components inthe horizontal direction and low-frequency components inthe vertical direction and the high-frequency componentsin both directions Thus in essence the standard DWTalgorithm yields horizontal vertical and diagonal directionalinformation about the frequency spectrum of an imageHowever these three directions may not be sufficient toexpress all the directional information in digital images par-ticularly biomedical images [4 14] In an attempt to expressthe directional features more efficiently several directionalwavelet systems have been proposedThese include theGaborwavelets [17] the dual-tree complex wavelet transform (DT-CWT) [18] the ridgelet [19] the curvelet [20] and thecontourlet [21] There exist also reports on biomedical appli-cations of Gabor filter banks [22] DT-CWT [4] ridgelets[23] curvelets [24] and contourlets [5]

The two-dimensional (2D) Gabor filter decomposes animage into components corresponding to different scalesand orientations As a result it captures visual propertiessuch as spatial localization orientation selectivity and spatialfrequency The 2D Gabor filter has real and imaginary partsand is highly flexible in its representation as its parameterscan be adapted to the structure of the patterns that one wantsto analyze in the image It is however difficult to find theoptimal set of parameters to characterize a given image Incomparison the DT-CWT transform provides directionalselectivity shift invariant features and complex imagesHowever it suffers from limited orientation selectivity [25]and redundancy of information [26] The ridgelet transformis appropriate to capture radial directional details in thefrequency domain in particular it is optimal for representingstraight-line singularities However those structures are notdominant in medical images and are rarely observed inreal world images This limits the suitability of the ridgelettransform to characterize the texture of real images [27] Thecurvelet transform is an extension of the ridgelet transformfor detecting image edges and singularities along curveswhileanalyzing images at multiple scales locations and desiredorientations It is particularly suitable for image featureswith discontinuities across straight lines Unfortunately thecurvelet transform is highly redundant [28] and only fewchoices of mother functions are available for the curveletsas opposed to the many choices available for the standard

wavelet transform [29] Finally the contourlet transformcan capture directional details and smooth contours in agiven image In particular it is suitable in the analysis ofimages containing textures and oscillatory patterns Its maindrawback is the high degree of information redundancy andoccurrence of artefacts [30 31]

In past works we proposed several transform-basedapproaches to account for directional features in classifyingbiomedical images For instance in the case of brainmagneticresonance images we proposed a simple methodology in[32 33] where features are extracted from the LH and HLcomponents of the DWT instead of the more common LLor image approximation component We found that the LHand HL coefficients are efficient at characterizing changesin the biological tissue and help distinguish normal andabnormal image textures For mammograms we investigatedin [34] a hybrid processing system that sequentially usesthe discrete cosine transform (DCT) to obtain the high-frequency component of the mammogram and then appliesthe Radon transform (RT) to the result in order to extract itsdirectional featuresThe validation results showed that the RThelps improve the recognition rate of the detection system Insubsequent work we combined the DWT and RT transforms[35] The approach targeted the HH component of the DWTdecomposition and improved classification accuracy whencompared to using the DWT or RT alone or the DCT-RT used in [34] Our previous works clearly showed thatdirectional information helps improve classification accu-racy In addition the DWT-RT detection system was moreefficient for classifying normal and abnormal images than theDCT-RT possibly because of the multiresolution capabilityof the DWT and the fact that it leads to a sparser signalrepresentation than the DCT Still the RT cannot capturespatial frequency a potential feature to improve further theclassification accuracy

In this paper we describe a hybrid biomedical imageprocessing and classification system that uses both the DWTand Gabor filter as directional transforms and statisticalfeatures derived from them for the classification task whichis accomplished by support vector machines (SVMs) [36]As stated before the DWT is powerful at providing sparseand efficient image representations [14] However exceptfor the LH and HL subbands whose coefficients depend onimage row and column information respectively (an effectof the subband coding used by the algorithm) the standardDWT is essentially an image compression tool and it cannotperform directional analysis at arbitrary directions On theother hand the Gabor filter can process images in termsof preferred orientations at arbitrary spatial frequenciesMoreover it provides nonredundant information and canoffer high directional selectivity Thus combining DWT andGabor filter banks in sequence may lead to improved featureextraction from biomedical images and better classificationof normal versus abnormal images in comparison to usingDWT or Gabor filter banks alone In this hybrid processingscheme theDWTacts both as high-frequency filter to extractabrupt changes in image texture and image compressionengine to reduce image dimensionality and a Gabor filterbank extracts the directional information

Journal of Medical Engineering 3

In a preliminary work [37] the previously mentionedDWT-Gabor hybrid systemwas successfully applied tomam-mograms to extract features that allow discriminating normaland cancer images More specifically the goal was to detectthe presence of malign microcalcifications (specs of calciumin the breast tissue that appear in the mammogram as smallbright spots that are scattered or grouped in clusters) whoseearly detection is important for cancer screening [38 39]The results showed the superiority of the approach oversimply using the DWT alone In the present work wewiden our study to retina digital images and brain magneticresonance images to investigate the effectiveness of theDWT-Gabor approach across application domains with similarimage features Indeed the images of some pathologiesrelated to brain retina and breast present similar contrastfeatures characterized by abrupt changes in image texturewith directional properties (see examples in Figures 1 2and 3) For instance breast cancer is characterized by denseconcentration of contrast cells in the biological tissue cancerin brain magnetic resonance images is often characterized bylarge cells with high contrast and many forms of retinopathyare characterized by the presence of spots on the retina orcovering the macula As a result the DWT-Gabor hybridsystem we have used in our previous work [37] to detectcancer in mammograms could potentially also be applied tobrain magnetic resonance images and retina digital imageswith similar properties Next is a brief description of thepathologies that were studied in this work

Circinate retinopathy is a retinal degeneration charac-terized by a circle of white spots encircling the maculathat causes complete foveal blindness [40] retinal microa-neurysms are due to a swelling of the capillaries caused bya weakening of the vessel wall [41] and are considered tobe the earliest sign of diabetic retinopathy among othersMagnetic resonance imaging (MRI) is a noninvasive imagingmodality largely used for brain imaging to detect diseasessuch as Alzheimerrsquos and multiple sclerosis [6 8] Alzheimerrsquosdisease is the most frequent cause of age-related dementiaand multiple sclerosis is a progressive neurological disorderthat can result in various dysfunctions [42] Additional brainpathologies that can be detected fromMR images and that areinvestigated in this work include glioma herpes encephalitisand metastatic bronchogenic carcinoma (Figure 1) All ofthese are characterized by large cells with high contrasthence the interest in being able to detect them with the samealgorithm

The contribution of our work can be summarized asfollows First we propose a relatively simple and fast approachto biomedical image characterization that relies on the direc-tional properties of high-frequency componentsTheDWT isapplied first to extract high-frequency components that char-acterize abrupt changes in the biological tissue and then theGabor filter is applied to the obtainedHH subimage to extractdirectional features Second the statistical features extractedfrom the hybrid DWT-Gabor transform are processed by anSVM for classification This statistical binary classifier hasproven its efficiency [4ndash6 32ndash35 37] and ease of tuning incomparison to alternatives such as artificial neural networksAnother desirable feature is its scalability and ability to avoid

local minima [36] Third contrary to alternatives that focuson ROIs or specific image details the proposed methodologyis ofmore general reach as three different types of images usedfor validation show

The paper is organized as follows Section 2 reviewsprevious works related to the automatic classification ofnormal versus abnormal images in the context of brain mag-netic (MR) resonance imaging mammograms and retinadigital images Section 3 describes our proposed approach fordirectional features extraction from biomedical images usingdiscrete wavelet transform followed by Gabor filter banksand support vector machines classifier Section 4 presentsexperimental results Finally Section 5 draws the conclusionsand gives future work to be done

2 Related Works

Mammograms retina and MR images are the subject ofmany research efforts on feature extraction and subsequentclassification Next is a summary of some recent worksrelated to DWTs andor Gabor filters In the problem ofautomatic classification of mammograms the authors in [3]used Gabor filter banks to process images and 119896 nearestneighbour (119896-NN) algorithm as classifier The obtained clas-sification rate was 80 In [4] the dual-tree complex wavelettransform (DT-CWT) and support vector machine (SVM)were employed to classify benign and malignant images Theexperimental result achieved 8864 classification accuracyThe authors in [5] employed the contourlet transform andsuccessive enhancement learning (SEL) weighted SVM toobtain 966 correct classification rate The previous studiesall used images of size 1024 times 1024 pixels

In the problem of retina digital image classification theauthors in [1] employed the Belkynsrsquos shift-invariant DWTto classify normal against abnormal retina images of size700 times 605 pixels The pathologies of the abnormal imagesincluded exudates large drusen fine drusen choroidal neo-vascularization central vein and artery occlusion histoplas-mosis arteriosclerotic retinopathy hemicentral retinal veinocclusion and more In order to capture texture directionalfeatures they employed normalized gray level cooccurrencematrices (GLCMs)The obtained classification accuracy withlinear discriminant analysis (LDA) was 822 The authorsin [2] employed the probabilistic boosting algorithm andmorphological scale space analysis and GLCM to extracttexture features The purpose was to classify normal imagesversus drusen images with various texture complexities Thedetection accuracy of normal images varied between 813and 922 and that of abnormal images varied between717 and 852 depending on texture complexity (gradeof pathology) The authors in [43] used four approaches toextract features from retina digital images of size 300 times 300pixels to automatically classify glaucoma images The first setof features is obtained by taking the pixel intensities as inputto principal component analysis The second features areobtained from Gabor texture filter responses The third set offeatures is computed from the coefficients of the fast Fouriertransform and the fourth set of features is obtained from

4 Journal of Medical Engineering

(a) Normal (b) Alzheimerrsquos disease (c) Glioma

(d) Herpes encephalitis (e) Metastatic bronchogenic carcinoma (f) Multiple sclerosis

Figure 1 Examples of brain MR images

(a) Normal (b) Microaneurysms (c) Circinate

Figure 2 Examples of retina images

the histogram of the intensity distribution of the imageFinally support vector machines were employed for the clas-sification task The performance of the classifications usingone feature set only was 73 with the histogram features76with the fast Fourier transform coefficients 80with theGabor textures and 83 with the pixel intensities

Finally in the problem of brain MRI classification theauthors in [6] used the wavelet coefficients as input to asupport vector machine to classify normal and abnormal

Alzheimerrsquos disease images of size 256 times 256 pixels Theclassification accuracy was 98 using SVM with a radialbasis kernel More recently the authors in [42] used voxelsto represent each brain MRI of size 512 times 512 pixels Usingcross-validated tests the obtained correct classification ratesof normal and Alzheimer images were 90 92 and78 respectively when using classification by SVM naıveBayes classifier and voting feature intervals (VFIs) Stillmore recently the authors in [8] employed the DWT to

Journal of Medical Engineering 5

(a) Normal (b) Abnormal

Figure 3 Examples of mammograms

extract features from brain magnetic resonance images ofsize 256 times 256 pixels and then principal component analysiswas used to reduce the dimensions of the features space Theabnormal images included gliomameningiomaAlzheimerrsquosAlzheimerrsquos plus visual agnosia Pickrsquos disease sarcomaand Huntingtonrsquos disease The classification accuracies usingbackpropagation neural network (BPNN) were 100 usinglearning and testing sets of 33 images each

In this work we are interested in how a DWT-Gabor-based approach for feature extraction may provide betterclassification results than those reported in the previousworks particularly those based on the DWT alone The nextsection provides the details of our methodology

3 Methodology

The overall methodology proceeds as follows First theDWT is applied to the biomedical image to obtain its high-frequency image component since it often contains mostof the desired information about the biological tissue [39]Indeed sudden changes in the texture of the image are typicalindicators of the presence of abnormal biological tissueSecond a bank of Gabor filters with different scales andorientations is applied to the high-frequency image to obtainGabor-filtered images along different spatial orientationsThird statistical features are extracted from the Gabor-filtered images Finally the SVM is used to classify theresulting feature vector for final diagnosisThe block diagramof the DWT-Gabor system is shown in Figure 4 Figure 5summarizes the DWT approach in comparison

31 Discrete Wavelet Transform The two-dimensional dis-crete wavelet transform (2D-DWT) [14ndash16] performs asubband coding of an image in terms of spectral spa-tialfrequency components using an iterative and recursiveprocess Figure 6 illustrates the case of two-level decompo-sition The image is first represented by LH HL and HHsubbands that encode the image details in three directionsand an LL subband which provides an approximation of it

Source image DWTExtract DWT

high frequency

Gabor filter

Extract Gabor filtered images

Compute statistics

SVM Classification

Figure 4 Schematic diagram of the DWT-Gabor approach

Mammogram DWT Extract DWT high frequency

Compute statistics

SVMClassification

Figure 5 Schematic diagram of the DWT approach

The obtained detail or approximation images can be decom-posed again to obtain second-level detail and approximationimages and the process can be repeated for finer analysis aseach iteration doubles the image scale

The computation of the 2D-DWT proceeds from that ofthe 1D-DWT the discrete version of the one-dimensionalcontinuous wavelet transform The one-dimensional contin-uous wavelet transform of a signal 119909(119905) is defined by [7 8]

119882120595 (119886 119887) = int

+infin

minusinfin

119909 (119905) 120595lowast

119886119887(119905) 119889119905 (1)

120595119886119887 (119905) =1

radic|119886|120595(

119905 minus 119887

119886) (2)

where 120595119886119887(sdot) stands for a given wavelet function and 119886 and 119887are the scale and translation parameters respectivelyThe 1D-DWT is obtained by sampling 119886 and 119887 so that (1) becomes that

6 Journal of Medical Engineering

Image

LL

LH

HL

HH

darr

darr

darr

darr

darr

darr

119892(119899)

119892(119899)

119892(119899)

ℎ(119899)

ℎ(119899)ℎ(119899)

Figure 6 2D-DWT decomposition of an image

of a sequence In dyadic sampling 119886 and 119887 are respectively apower of 2 and multiples thereof and the sequence elements(wavelet coefficients) are given by

119888119895119896 = 119882120595 (2minus119895 2minus119895119896) (3a)

where 119895 represents the discrete scale factor and 119896 the discretetranslation factor In other words 119886 and 119887 in (1) are replacedby 2119895 and 2119895119896 respectively

The one-dimensional wavelet decomposition is extendedto an image by applying it to the row variable first and thento the column variable of the obtained result [44] At eachstep two subimages are created with half the number ofpixels of the row or column that was processed In the endan 119872 times 119873 image is decomposed into 4 subimages eachwith 1198722 times 1198732 resolution and preserved scale However(1) has only theoretical merit due to the infinite ranges of119886 and 119887 For a practical implementation the fact that (1)is essentially a measure of correlation between a signal andvarious wavelets derived from a mother is exploited and theDWT decomposition is turned into a filtering operation witha sequence of high-pass and low-pass filters [45] Followingthe notation in [7 8] the discrete form of (1) can then bewritten as

119888119886119895119896 [119909 (119905)] = DS [sum119909 (119905) 119892lowast

119895(119905 minus 2

119895119896)]

119888119889119895119896 [119909 (119905)] = DS [sum119909 (119905) ℎlowast

119895(119905 minus 2

119895119896)]

(4)

where coefficients 119888119886119895119896 and 119888119889119895119896 specify approximationand details components provided by the 119892(119899) low-pass andℎ(119899) high-pass impulse responses respectively and the DSoperator performs downsampling by a factor of 2 The one-dimensional wavelet decomposition is extended to two-dimensional objects by using row and column decomposi-tions as shown in Figure 6 In our work the most frequentlyused wavelet (Daubechies-4) [25] is considered to extract theHH image component

32 Gabor Filter The two-dimensional (2D) Gabor filterdecomposes an image into components corresponding todifferent scales and orientations [22] thus capturing visualproperties such as spatial localization orientation selectivityand spatial frequency The 2D Gabor filter consists of acomplex exponential centered at a given frequency andmodulated by a Gaussian envelope Because of the complex

exponential the filter has both real and imaginary parts Thegeneral form of the real part is defined as follows

119866(119909 119910 120590119909 120590119910 119891 120579) = exp[minus12((

1199091015840

120590119909

)

2

+ (1199101015840

120590119910

)

2

)]

times cos (21205871198911199091015840)

(5)

where

1199091015840= 119909 cos (120579) + 119910 sin (120579)

1199101015840= 119910 cos (120579) minus 119909 sin (120579)

(6)

and where 120590119909 and 120590119910 are the standard deviations of theGaussian envelope along the 119909 and 119910 axes The parameters 119891and 120579 are respectively the central frequency and the rotationof the Gabor filter To obtain the Gabor-filtered image119891(119909 119910)of a given image 119894(119909 119910) the 2D convolution operation (lowast) isperformed

119891 (119909 119910) = 119866 (119909 119910 120590119909 120590119910 119891 120579) lowast 119894 (119909 119910) (7)

The selection of parameters 120590119909 120590119910 119891 and 120579 plays animportant role in the filterrsquos operation However no formaltechnique exists for choosing them and experience-guidedintuition trial and error or heuristic searchmust be used Forretina digital images and brain MR images 120590119909 and 120590119910 werearbitrarily set to unity In the case of mammograms 120590119909 and120590119910 were set to the values used in [46] which were determinedempirically Consequently we used 120590119909 = 120591235 where 120591 isthe full width at half-maximumof theGaussian and120590119910 = 8120590119909No values of120590119909 and120590119910 other than the previous oneswere triedsince optimalitywas not the primary concern of thiswork andwe obtained satisfactory results with these values

Four orientations 120579 = 0 1205874 1205872 and 31205874 were used asin [22 33] These values seemed reasonable as a first try sincethey covered both image axis directions and the forward andbackward diagonals Finally the central frequency 119891 was setto 2 25 and 3 Given that the Gabor filter is modulated bythe cosine of 119891 large values of 119891 lead to a compressed cosineand consequently the filter output is more likely to showfast or frequent changes in biological tissue texture This inturn would help verify our hypothesis that abnormal imagesare characterized by sudden and frequent variations in imagetexture In the end the application of the Gabor filter bank tothe HH image component obtained with the 2D-DWT leadsto twelve Gabor-filtered HH images components for eachchoice of 119891 and 120579

33 Feature Extraction Statistical measures are employed toextract features from both the DWTHH subband image andthe real Gabor-filtered HH image components More pre-cisely the entropy (119864) and uniformity (119880) of the coefficientsof each one are computed Entropy and uniformity wereselected as features because previousworks onmammogramshave shown that uniformity is correlated with suspiciousmalignancy [47] and that entropy can successfully charac-terize breast biological tissue [48] In this study the entropy

Journal of Medical Engineering 7

and uniformity statistics are hypothesized to also characterizeretina andbrainMR imageswith similar contrast information(ie abrupt andor frequent variations in texture) Entropy(119864) and uniformity (119880) are defined by [49]

119864 = minussum119901 (119911) times log (119901 (119911))

119880 = sum1199012(119911)

(8)

where 119911 is a random variable that represents a coefficientin the Gabor filtered image and 119901(119911) is its probability ofoccurrence as estimated by its relative frequency

To investigate the performance of the previous approach-es the image features were extracted from HH at bothlevel-one DWT decomposition (HH1) and level-two DWTdecomposition (HH2) with and without filtering by a Gaborfilter bank We also applied the Gabor filter directly to theoriginal image without the DWT for comparison purposeFor eachDWTHH subband image the feature vector is givenby

119881DWT119886 = [119864119886 119880119886] (9)

where 119886 is the level of wavelet analysis (decomposition)Similarly for each of the twelve outputs generated by theGabor filter bank (4 angles times 3 central frequencies) theentropy and uniformity are computed and a twenty-fourcomponent feature vector is formed to represent the initialimage We thus have

119881Gabor119886 = [1198641119886 1198642119886 11986412119886 1198801119886 1198802119886 11988012119886] (10)

Either feature vector is subsequently fed to the SVM toclassify normal versus pathological images

34 The Support Vector Machine Classifier Introduced byVapnik [36] the support vector machine (SVM) classifieris based on statistical learning theory It implements theprinciple of structural risk minimization and has excellentgeneralization ability as a result even when the data sample issmall Moreover SVM can tolerate high-dimensional andorincomplete data [50] It has been used with great success invarious applications including speech emotion recognition[51] card-sharing traffic detection [52] fault diagnosis [53]cardiac decision making [54] Parkinsonrsquos disease diagnosis[55] and Alzheimerrsquos disease detection [56]

The support vector machine performs classification tasksby constructing an optimal separating hyperplane that maxi-mizes themargin between the two nearest data points belong-ing to two separate classes Given a training set (119909119894 119910119894) 119894 =1 2 119898 where the input 119909119894 isin 119877

119889 and class labels 119910119894 isin+1 minus1 the separation hyperplane for a linearly separablebinary classification problem is given by

119891 (119909) = ⟨119908 sdot 119909⟩ + 119887 (11)

where 119908 is a weight vector and 119887 is a bias The optimalseparation hyperplane is found by solving the followingoptimization problem

minimize119908119887120585

1

2⟨119908 sdot 119908⟩ + 119862

119898

sum

119894=1

120585119894

Subject to 119910119894 (⟨119908 sdot 119909119894⟩ + 119887) + 120585119894 minus 1 ge 0 120585119894 ge 0

(12)

where 119862 is a penalty parameter that controls the tradeoffbetween the complexity of the decision function and thenumber of misclassified training examples and 120585 is a positiveslack variableTheprevious optimizationmodel can be solvedby introducing Lagrange multipliers and using the Karush-Kuhn-Tucker theorem of optimization to obtain the solutionas

119908 =

119898

sum

119894=1

120572119894119910119894119909119894 (13)

The 119909119894 values corresponding to positive Lagrange multipliers120572119894 are called support vectors which define the decisionboundary The 119909119894 values corresponding to zero 120572119894 are irrel-evant Once the optimal solution 120572

lowast

119894is found the optimal

hyperplane parameters 119908lowast and 119887lowast are determined Then thediscriminant function of the SVM for a linearly separablebinary classification problem is [32]

119892 (119909) = sign(119898

sum

119894=1

119910119894120572lowast

119894⟨119909119894 sdot 119909⟩ + 119887

lowast) (14)

In the nonlinearly separable case the SVM classifier nonlin-early maps the training points to a high-dimensional featurespace using a kernel function Φ where linear separation canbe possible The scalar product ⟨Φ(119909119894) sdot Φ(119909119895)⟩ is computedby Mercer kernel function 119870 as 119870(119909119894 119909119895) = ⟨Φ(119909119894) sdot Φ(119909119895)⟩Then the nonlinear SVM classifier has the following form

119892 (119909) = sign(119898

sum

119894=1

119910119894120572lowast

119894119870⟨119909 119909119894⟩ + 119887

lowast) (15)

In this study a polynomial kernel of degree 2 was used forthe SVM As a global kernel it allows data points that are faraway from each other to also have an influence on the kernelvalues The general polynomial kernel is given by

119870(119909 119909119894) = ((119909119894 sdot 119909) + 1)119889 (16)

where 119889 is the degree of the polynomial to be used

4 Experimental Results

Asmentioned previouslymammograms and retina and brainMR images corresponding to given pathologies are consid-ered in this work and the aim is to classify normal versusabnormal images for each image category To do so onehundred digitalmammograms (171times 364 pixels) consisting offifty normal images and fifty cancer images were taken fromThe Digital Database for Screening Mammography (DDSM)

8 Journal of Medical Engineering

Table 1 Average SVM classification accuracy as a function of feature extraction method and level of DWT decompositionlowast

DWT DWT-Gabor DWT DWT-GaborDecomposition level One One Two TwoMammograms 9598 (plusmn004) 9667 (plusmn005) 8913 (plusmn001) 9109 (plusmn005)Retina 7469 (plusmn005) 100 9098 (plusmn003) 100Brain MRI 8780 (plusmn000) 9736 (plusmn002) 8576 (plusmn000) 9118 (plusmn004)lowastTenfold cross-validation used for mammograms and retina images leave-one-out used for brain MRIs

Table 2 SVM classification specificity as a function of feature extraction method and level of DWT decompositionlowast

DWT DWT-Gabor DWT DWT-GaborDecomposition level One One Two TwoMammograms 9781 (plusmn002) 100 8885 (plusmn005) 9209 (plusmn006)Retina 681 (plusmn004) 100 232 (plusmn010) 100Brain MRI 2155 (plusmn001) 9958 (plusmn001) 0 9724 (plusmn001)lowastTenfold cross-validation used for mammograms and retina images leave-one-out used for brain MRIs

[57] For retina a set of 69 color images (150 times 130 pixels)from the STARE [58] database were employed including23 normal images 24 with microaneurysms and 22 withcircinate Finally a collection of 56 axial T2-weighted andMR brain images (256 times 256 pixels) were taken from theAANLIB database [59] of the Harvard Medical School Theyconsisted of 7 normal images 9 with Alzheimerrsquos disease 13with glioma 8 with Herpes encephalitis 8 with metastaticbronchogenic carcinoma and 14 with multiple sclerosis Itis unfortunate that the number of images was not constantacross pathologies but we had no control over this and usedwhat was available with tenfold cross-validation or leave-one-out cross-validation of the results depending on samplesize All experiments were based on a binary classificationapproach of normal versus abnormal images Many kinds ofbiomedical images could be considered for our experimentswe focused on mammograms retina and brain magneticresonance images mainly because of public availability Anexample of the processing of a normal retina and a retina withcircinate is illustrated in Figures 7 and 8 respectively

For each image type the average and standard deviationof the correct classification rate (CCR) sensitivity and speci-ficity were computed to evaluate the performance featureextraction techniquewhen used in conjunctionwith the SVMclassifier The three performance measures are defined by

CCR =Classified Samples

Total Number of Samples

Sensitivity =Correctly Classified Positive Samples

True Positive Samples

Specificity =Correctly Classified Negative Samples

True Negative Samples

(17)

where positive samples and negative samples are respectivelyabnormal and normal images

Finally all experiments were performed with tenfoldcross-validation except those for MR images which usedleave-one-out cross-validation due to the small sample sizeof each brain image category

Table 1 shows the obtained average results for the threetypes of images that were investigated The performance ofthe SVM classifier improved for all types of images and alllevels of HH decomposition by the DWT At level one theaverage correct classification rate increased by respectively069 2531 and 956 percentage points for mammogramsretina and brain magnetic resonance images when usingthe DWT-Gabor approach At level-two decomposition theimprovement was respectively 196 902 and 542 percent-age points

Tables 2 and 3 provide the average results for classifiersensitivity and specificity At level-one DWT decompositionthe DWT-Gabor approach improved classification specificityfor mammograms and retina images to make it reach 100while it improved it by 7803 percentage points for brain MRimages At level-two DWT decomposition the improvementwas 324 percentage points for mammograms 9768 percent-age points for retina images (100 specificity) and 9724percentage points for MR images Regarding sensitivity theresults were mixed

At level-one DWT decomposition the values were aboutthe same for the DWT-Gabor and DWT-only approachesfor mammograms and retina images with respectivelyminus081 and 0 percentage points differences but there was adegradation of minus1133 and minus4616 percentage points for brainMR images at level one and level two of decompositionrespectively

Following the same cross-validation protocol we alsoconducted classification experiments with features extractedfrom aGabor filtered image of the original biomedical imageThe purpose was to check whether Gabor-based featuresalone help characterize images better than DWT or DWT-Gabor-based features The results are given in Table 4

Journal of Medical Engineering 9

20 40 60 80 100 120 140

20

40

60

80

100

120020406080100120140160180200

(a) Original normal retina image

10 20 30 40 50 60 70

10

20

30

40

50

60 minus60

minus40

minus20

0

20

40

60

(b) Normal retina HH1 subimage

0

5

10

15

20

25

30

35

20406080100

minus100

minus80

minus60

minus40

minus20

5 10 15 20 25 30 35 40(c) Normal retina HH2 subimage

10 20 30 40 50 60 70

10

20

30

40

50

60 10

20

30

40

50

60

70

(d) Gabor filtered HH1 at 119891 = 2 and 120579 = 1205874

5

10

15

20

25

30

355 10 15 20 25 30 35 40

102030405060708090100

(e) Gabor filtered HH2 119891 = 2 and 120579 = 1205874

Figure 7 Analysis of a normal retina

20 40 60 80 100 120 140

20

40

60

80

100

120 20406080100120140160180200220

(a) Original retina image with circinate

10 20 30 40 50 60 70

10

20

30

40

50

60minus20

minus10

0

10

20

30

(b) HH1 subimage

5

10

15

20

25

30

355 10 15 20 25 30 35 40

minus80

minus60

minus40

minus20020406080100

(c) HH2 subimage

10 20 30 40 50 60 70

10

20

30

40

50

600

5

10

15

20

25

30

35

(d) Gabor filtered HH1 at 119891 = 2 and 120579 = 1205874

102030405060708090100

5

10

15

20

25

30

355 10 15 20 25 30 35 40

(e) Gabor filtered HH2 at 119891 = 2 and 120579 = 1205874

Figure 8 Analysis of a retina with circinate

10 Journal of Medical Engineering

Table 3 SVM classification sensitivity as a function of feature extraction method and level of DWT decompositionlowast

DWT DWT-Gabor DWT DWT-GaborDecomposition level One One Two TwoMammograms 9414 (plusmn006) 9333 (plusmn006) 9029 (plusmn0039) 8978 (plusmn004)Retina 100 100 100 100Brain MRI 9384 (plusmn000) 8251 (plusmn016) 100 5384 (plusmn023)lowastTenfold cross-validation used for mammograms and retina images leave-one-out used for brain MRIs

Table 4 SVM classification performance measures obtained withGabor-based features

Accuracy Specificity SensitivityMammograms 6803 (plusmn001) 100 0Retina 5000 (plusmn000) 100 0Brain MRI 8661 (plusmn003) 100 0

The obtained correct classification rate of mammogramsretina and brain magnetic resonance images is respectively6803 (plusmn001) 5000 (plusmn000) and 8661 (plusmn003) Theaverage results for classifier specificity and sensitivity for allimages are 100 and 0 This finding indicates that Gabor-based features are suitable to detect pathological imagesbut fails to detect normal images In sum the results showthat Gabor-based features do not perform better than DWTand DWT-Gabor-based features These findings confirm thesuperiority of combining theDWT andGabor filter banks forfeature extraction

Based on the previous results it appears that the DWT-Gabor approach for feature extraction is effective for detect-ing the abrupt changes in biological tissue that characterizethe pathological patterns that were investigated and it yieldsbetter classification accuracy and specificity than the DWT-only approach It also offers equal of better sensitivity exceptfor brain MRIs For brain MRIs the obtained specificityand sensitivity results with the DWT-Gabor approach showimproved true negative detection but lower true positiveperformance Finally the obtained results reveal also thatlevel-one DWT decomposition is preferable to level-twodecomposition

Finally Table 5 compares the results obtained with theDWT-Gabor approach to other work that we surveyed Inmany cases the DWT-Gabor method yields higher clas-sification rates particularly for mammograms and retinaFor the problem of brain MRI classification our obtainedperformance is better than the results of [38] but less thanwhat is reported in [6 8] However these comparisons shouldbe viewed with caution as not all the results stem froma common image database and the different authors usedifferent sample and image sizesMoreovermany authors useno cross-validation and simply performa single arbitrary splitof their data into training and test sets to obtain their accuracyresults Obviously one cannot generalize or draw definiteconclusion from such efforts and comparisons between

works cannot be made other than in general terms In thisrespect it can only be concluded from our results that theDWT-Gabor for feature extraction is effective for obtaininghigh image classification accuracy by an SVM and that itmay outperform other feature extraction and classificationtechniques reported in the literature at least those basedon DWT-only image decomposition Unfortunately a moredefinite conclusion is impossiblewithout gaining access to theimage databases used by the other authors

5 Computational Complexity

Finally the computational complexity of the DWT Gaborcontourlet and curvelet for an 119873 times 119873 image is respectively119874(119873) 119874(1198732 times 119872

2) with 119872 being the width of Gabor

(Gaussian) mask filter119874(1198732) and119874(1198732 log(119873)) As a resultthe computational complexity of the combination of theDWT and Gabor filter is 119874(119873) + 119874(1198732 times 1198722) In terms offeatures extraction processing time the average time requiredto process a brain a mammogram and a retina imagewith the DWT approach (DWT-Gabor) was respectively019 (031) 017 (032) and 015 seconds (035) using MatlabR2009a on a 15 GHz Core2 Duo processor

6 Conclusion

We proposed a supervised system for biomedical imagesclassification that uses statistical features obtained from thecombination of the discrete wavelet transform and Gaborfilter to classify normal images versus cancer images usingsupport vector machines as classifiers Our experimentalresults show that such a hybrid processing model achieveshigher accuracy in comparison to using DWT or Gaborfilter banks alone Therefore the proposed image processingand features extraction approach seem to be very promis-ing for the detection of certain pathologies in biomedicalimages

For future works it is recommended to consider a largerset of features and a selection process to identify the mostdiscriminant ones In addition the Gabor parameters willbe adjusted for each type of image separately to improvethe accuracy Furthermore the DWT-Gabor will be directlycompared to the dual-tree complex wavelet curvelet andcontourlet using the same databases and images in orderto draw general conclusions Also multilabels classificationswill be considered in future works to investigate the discrim-inative power of our approach for each type of pathology

Journal of Medical Engineering 11

Table 5 Comparison with the literature

Features Classifier Accuracylowast

Mammograms[3] Gabor k-NN 80[4] DT-CWT SVM 8864[5] Contourlet SVM 966Our approach DWT-Gabor SVM 9667 (plusmn005)

Retina[1] DWT + GLCM LDA 822

[2] Morphological + GLCM Probabilistic boosting algorithm 813ndash922717ndash852

[39] Gabor SVM 83Our approach DWT-Gabor SVM 100

Brain[6] DWT SVM 98[8] DWT + PCA BPNN 100

SVM 90[38] Voxels Bayes 92

VFI 78Our approach DWT-Gabor SVM 9736 (plusmn002)

lowastCorrect classification rate

Finally more experiments on the effect of kernel choice andits parameter on classification accuracy will be investigated

References

[1] A Khademi and S Krishnan ldquoShift-invariant discrete wavelettransform analysis for retinal image classificationrdquoMedical andBiological Engineering and Computing vol 45 no 12 pp 1211ndash1222 2007

[2] N Lee A F Laine andT R Smith ldquoLearning non-homogenoustextures and the unlearning problemwith application to drusendetection in retinal imagesrdquo in Proceedings of the 5th IEEEInternational Symposium on Biomedical Imaging From Nano toMacro (ISBI rsquo08) pp 1215ndash1218 Paris France May 2008

[3] A Dong and B Wang ldquoFeature selection and analysis onmammogram classificationrdquo in Proceedings of the IEEE PacificRim Conference on Communications Computers and SignalProcessing (PACRIM rsquo09) pp 731ndash735 Victoria BC CanadaAugust 2009

[4] A Tirtajaya and D D Santika ldquoClassification of microcalcifi-cation using dual-tree complex wavelet transform and supportvector machinerdquo in Proceedings of the 2nd International Confer-ence onAdvances in Computing Control and TelecommunicationTechnologies (ACT rsquo10) pp 164ndash166 Jakarta Indonesia Decem-ber 2010

[5] F Moayedi Z Azimifar R Boostani and S Katebi ldquoContour-let-based mammography mass classification using the SVMfamilyrdquo Computers in Biology and Medicine vol 40 no 4 pp373ndash383 2010

[6] S Chaplot LM Patnaik andN R Jagannathan ldquoClassificationof magnetic resonance brain images using wavelets as input tosupport vector machine and neural networkrdquo Biomedical SignalProcessing and Control vol 1 no 1 pp 86ndash92 2006

[7] Y Zhang S Wang and L Wu ldquoA novel method for magneticresonance brain image classification based on adaptive chaoticPSOrdquo Progress in Electromagnetics Research vol 109 pp 325ndash343 2010

[8] Y Zhang Z Dong L Wu and S Wang ldquoA hybrid method forMRI brain image classificationrdquo Expert Systems with Applica-tions vol 38 no 8 pp 10049ndash10053 2011

[9] M E Celebi H Iyatomi G Schaefer and W V StoeckerldquoLesion border detection in dermoscopy imagesrdquoComputerizedMedical Imaging and Graphics vol 33 no 2 pp 148ndash153 2009

[10] Q AbbasM E Celebi and I F Garcıa ldquoSkin tumor area extrac-tion using an improved dynamic programming approachrdquo SkinResearch and Technology vol 18 pp 133ndash142 2012

[11] Q Li F Li and K Doi ldquoComputerized detection of lungnodules in thin-section CT images by use of selective enhance-ment filters and an automated rule-based classifierrdquo AcademicRadiology vol 15 no 2 pp 165ndash175 2008

[12] A El-Bazl M Nitzken E Vanbogaertl G Gimelrsquojarb RFalfi and M Abo El-Ghar ldquoA novel shaped-based diagnosticapproach for early diagnosis of lung nodulesrdquo in Proceedingsof the IEEE International Symposium in Biomedical Imaging(ISBI rsquo11) pp 137ndash140 Chicago Ill USA 2011

[13] M T Coimbra and J P S Cunha ldquoMPEG-7 visual descriptorsmdashcontributions for automated feature extraction in capsuleendoscopyrdquo IEEE Transactions on Circuits and Systems for VideoTechnology vol 16 no 5 pp 628ndash636 2006

[14] B Li and M Q H Meng ldquoTexture analysis for ulcer detectionin capsule endoscopy imagesrdquo Image andVision Computing vol27 no 9 pp 1336ndash1342 2009

[15] C K Chui An Introduction to Wavelets Academic Press SanDiego Calif USA 1992

[16] M Vetterli andCHerley ldquoWavelets and filter banks theory anddesignrdquo IEEE Transactions on Signal Processing vol 40 no 9pp 2207ndash2232 1992

12 Journal of Medical Engineering

[17] J G Daugman ldquoUncertainty relation for resolution inspace spatial frequency and orientation optimized by two-dimensional visual cortical filtersrdquo Journal of the Optical Societyof America A vol 2 no 7 pp 1160ndash1169 1985

[18] I W Selesnick R G Baraniuk and N G Kingsbury ldquoThedual-tree complex wavelet transformrdquo IEEE Signal ProcessingMagazine vol 22 no 6 pp 123ndash151 2005

[19] E Candes and D Donoho ldquoRidgelets a key to higher-dimensional intermittencyrdquo Philosophical Transactions of theLondon Royal Society vol 357 pp 2495ndash2509 1999

[20] E J Candes and D L Donoho ldquoContinuous curvelettransformmdashI Resolution of the wavefront setrdquo Applied andComputational Harmonic Analysis vol 19 no 2 pp 162ndash1972005

[21] M N Do and M Vetterli ldquoThe contourlet transform an effi-cient directional multiresolution image representationrdquo IEEETransactions on Image Processing vol 14 no 12 pp 2091ndash21062005

[22] R J Ferrari R M Rangayyan J E L Desautels and A FFrere ldquoAnalysis of asymmetry in mammograms via directionalfiltering with Gabor waveletsrdquo IEEE Transactions on MedicalImaging vol 20 no 9 pp 953ndash964 2001

[23] Z Cui and G Zhang ldquoA novel medical image dynamic fuzzyclassification model based on ridgelet transformrdquo Journal ofSoftware vol 5 no 5 pp 458ndash465 2010

[24] T Geback and P Koumoutsakos ldquoEdge detection inmicroscopyimages using curveletsrdquo BMC Bioinformatics vol 10 article 752009

[25] J Ma andG Plonka ldquoThe curvelet transform a review of recentapplicationsrdquo IEEE Signal ProcessingMagazine vol 27 no 2 pp118ndash133 2010

[26] N Kingsbury ldquoComplex wavelets and shift invariancerdquo inProceedings of the IEEE Seminar on Time-Scale and Time-Frequency Analysis and Applications pp 501ndash510 London UK2000

[27] Y L Qiao C Y Song and C H Zhao ldquoM-band ridgelettransform based texture classificationrdquo Pattern RecognitionLetters vol 31 no 3 pp 244ndash249 2010

[28] F Gomez and E Romero ldquoTexture characterization using acurvelet based descriptorrdquo Lecture Notes in Computer Sciencevol 5856 pp 113ndash120 2009

[29] H Shan and J Ma ldquoCurvelet-based geodesic snakes for imagesegmentation with multiple objectsrdquo Pattern Recognition Let-ters vol 31 no 5 pp 355ndash360 2010

[30] R Eslami and H Radha ldquoNew image transforms using hybridwavelets and directional filter banks analysis and designrdquo inProceedings of the IEEE International Conference on ImageProcessing (ICIP rsquo05) pp 733ndash736 Genova Italy September2005

[31] O O V Villegas and V G C Sanchez ldquoThe wavelet basedcontourlet transform and its application to feature preservingimage codingrdquo Lecture Notes in Computer Science vol 4827 pp590ndash600 2007

[32] S Lahmir and M Boukadoum ldquoClassification of brain MRIusing the LH and HL wavelet transform sub-bandsrdquo in Pro-ceedings of the IEEE International Symposium on Circuits andSystems (ISCAS rsquo11) pp 1025ndash1028 Rio de Janeiro Brazil May2009 2011

[33] S Lahmir and M Boukadoum ldquoBrain MRI classificationusing an ensemble system and LH and HL wavelet Sub-bandsFeaturesrdquo in Proceedings of the IEEE Symposium Series on

Computational Intelligence (SSCI rsquo11) pp 1ndash7 Paris FranceApril2011

[34] S Lahmir and M Boukadoum ldquoHybrid Cosine and RadonTransform-based processing for Digital Mammogram FeatureExtraction and Classification with SVMrdquo in Proceedings of the33rd IEEE Annual International Conference on Engineering inMedecine and Biology Society (EMBS rsquo11) pp 5104ndash5107 BostonMass USA 2011

[35] S Lahmir andM Boukadoum ldquoDWT and RT-Based Approachfor Feature Extraction and classification of Mammograms withSVMrdquo in Proceedings of the IEEE Biomedical Circuits andSystems Conference (BioCAS rsquo11) pp 412ndash415 San Diego CalifUSA November 2011

[36] VN VapnikTheNature of Statistical LearningTheory Springer1995

[37] S Lahmiri and M Boukadoum ldquoHybrid discret wavelet trans-form and Gabor filter banks processing for mammogramfeatures extractionrdquo in Proceedings of the IEEE New Circuits andSystems (NEWCAS rsquo11) pp 53ndash56 Bordeaux France June 2011

[38] L M Bruce and N Shanmugam ldquoUsing neural networkswith wavelet transforms for an automated mammographicmass classifierrdquo in Proceedings of the 22nd Annual InternationalConference of the IEEE Engineering in Medicine and BiologySociety pp 985ndash987 Chicago Ill USA July 2000

[39] S M H Jamarani G Rezai-rad and H Behnam ldquoA novelmethod for breast cancer prognosis using wavelet packet basedneural networkrdquo in Proceedings of the 27th Annual InternationalConference of the Engineering in Medicine and Biology Society(IEEE-EMBS rsquo05) pp 3414ndash3417 Shanghai China September2005

[40] httpmedical-dictionarythefreedictionarycomcircinate+retinopathy

[41] C I O Martins F N S Medeiros R M S Veras F N Bezerraand R M Cesar Jr ldquoEvaluation of retinal vessel segmentationmethods for microaneurysms detectionrdquo in Proceedings of theIEEE International Conference on Image Processing (ICIP rsquo09)pp 3365ndash3368 Cairo Egypt November 2009

[42] C Plant S J Teipel A Oswald et al ldquoAutomated detectionof brain atrophy patterns based on MRI for the prediction ofAlzheimerrsquos diseaserdquo NeuroImage vol 50 no 1 pp 162ndash1742010

[43] J Meier R Bock L G Nyul and G Michelson ldquoEye fun-dus image processing system for automated glaucoma clas-sificationrdquo in Proceedings of the 52nd Internationales Wis-senschaftliches Kolloquium Technische Universitat Ilmenau2007

[44] E Sakka A Prentza I E Lamprinos and D KoutsourisldquoMicrocalcification detection using multiresolution analysisbased on wavelet transformrdquo in Proceedings of the IEEE Inter-national Special Topic Conference on Information Technology inBiomedicine Ioannina Greece October 2006

[45] S G Mallat ldquoTheory for multiresolution signal decompositionthe wavelet representationrdquo IEEE Transactions on Pattern Anal-ysis and Machine Intelligence vol 11 no 7 pp 674ndash693 1989

[46] J Suckling J Parker and D R Dance ldquoThe mammographicimage analysis society digital mammogram databaserdquo in Pro-ceedings of the the 2nd International Workshop on DigitalMammography A G Gale S M Astley D D Dance and AY Cairns Eds pp 375ndash378 Elsevier York UK 1994

[47] H J Chiou C Y Chen T C Liu et al ldquoComputer-aideddiagnosis of peripheral soft tissue masses based on ultrasound

Journal of Medical Engineering 13

imagingrdquo Computerized Medical Imaging and Graphics vol 33no 5 pp 408ndash413 2009

[48] J K Kim J M Park K S Song and H W Park ldquoAdaptivemammographic image enhancement using first derivative andlocal statisticsrdquo IEEE Transactions on Medical Imaging vol 16no 5 pp 495ndash502 1997

[49] H S Sheshadri andAKandaswamy ldquoBreast tissue classificationusing statistical feature extraction of mammogramsrdquo MedicalImaging and Information Sciences vol 23 no 3 pp 105ndash1072006

[50] N Cristianini and J Shawe-Taylor Introduction to SupportVector Machines and Other Kernel-Based Learning MethodsCambridge University Press Cambridge UK 2000

[51] L Chen X Mao Y Xue and L L Cheng ldquoSpeech emotionrecognition features and classification modelsrdquo Digital SignalProcessing vol 22 pp 1154ndash1160 2012

[52] F Palmieri U Fiore A Castiglione and A De Santis ldquoOn thedetection of card-sharing traffic through wavelet analysis andSupport Vector Machinesrdquo Applied Soft Computing vol 13 no1 pp 615ndash627 2013

[53] A Azadeh M Saberi A Kazem V Ebrahimipour A Nour-mohammadzadeh and Z Saberi ldquoA flexible algorithm forfault diagnosis in a centrifugal pump with corrupted data andnoise based on ANN and support vector machine with hyper-parameters optimizationrdquoApplied Soft Computing vol 13 no 3pp 1478ndash1485 2013

[54] R J Martis U R Acharya K M Mandana A K Ray andC Chakraborty ldquoCardiac decision making using higher orderspectrardquo Biomedical Signal Processing and Control vol 8 pp193ndash203 2013

[55] M R Mohammad ldquoChi-square distance kernel of the gaitsfor the diagnosis of Parkinsonrsquos diseaserdquo Biomedical SignalProcessing and Control vol 8 pp 66ndash70 2013

[56] R Vandenberghe N Nelissen E Salmon et al ldquoBinary clas-sification of 18F-flutemetamol PET using machine learningcomparison with visual reads and structuralMRIrdquoNeuroImagevol 64 no 1 pp 517ndash525 2013

[57] httpmarathoncseeusfeduMammographyDatabasehtml[58] httpwwwcesclemsonedusimahooverstare[59] httpwwwmedharvardeduaanlib

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

2 Journal of Medical Engineering

are used to obtain low-frequency information and shorterintervals are used to obtain high-frequency information

The most commonly used wavelet transform in biomedi-cal image processing is the discrete wavelet transform (DWT)[14] whose discrete time shifting and stretching variables leadto a sparse and efficient representation The DWT takes aninput image and decomposes into four subimage componentsthat characterize it for different orientations in the horizontaland vertical frequency axesThe process can be repeated withone or more subimages if needed More precisely the DWTdecomposition yields the approximation subband (LL) thehorizontal detail subband (LH) the vertical detail subband(HL) and the diagonal detail subband (HH) These describerespectively the low-frequency components in the horizontaland vertical directions the low-frequency components inthe horizontal direction and high-frequency components inthe vertical direction the high-frequency components inthe horizontal direction and low-frequency components inthe vertical direction and the high-frequency componentsin both directions Thus in essence the standard DWTalgorithm yields horizontal vertical and diagonal directionalinformation about the frequency spectrum of an imageHowever these three directions may not be sufficient toexpress all the directional information in digital images par-ticularly biomedical images [4 14] In an attempt to expressthe directional features more efficiently several directionalwavelet systems have been proposedThese include theGaborwavelets [17] the dual-tree complex wavelet transform (DT-CWT) [18] the ridgelet [19] the curvelet [20] and thecontourlet [21] There exist also reports on biomedical appli-cations of Gabor filter banks [22] DT-CWT [4] ridgelets[23] curvelets [24] and contourlets [5]

The two-dimensional (2D) Gabor filter decomposes animage into components corresponding to different scalesand orientations As a result it captures visual propertiessuch as spatial localization orientation selectivity and spatialfrequency The 2D Gabor filter has real and imaginary partsand is highly flexible in its representation as its parameterscan be adapted to the structure of the patterns that one wantsto analyze in the image It is however difficult to find theoptimal set of parameters to characterize a given image Incomparison the DT-CWT transform provides directionalselectivity shift invariant features and complex imagesHowever it suffers from limited orientation selectivity [25]and redundancy of information [26] The ridgelet transformis appropriate to capture radial directional details in thefrequency domain in particular it is optimal for representingstraight-line singularities However those structures are notdominant in medical images and are rarely observed inreal world images This limits the suitability of the ridgelettransform to characterize the texture of real images [27] Thecurvelet transform is an extension of the ridgelet transformfor detecting image edges and singularities along curveswhileanalyzing images at multiple scales locations and desiredorientations It is particularly suitable for image featureswith discontinuities across straight lines Unfortunately thecurvelet transform is highly redundant [28] and only fewchoices of mother functions are available for the curveletsas opposed to the many choices available for the standard

wavelet transform [29] Finally the contourlet transformcan capture directional details and smooth contours in agiven image In particular it is suitable in the analysis ofimages containing textures and oscillatory patterns Its maindrawback is the high degree of information redundancy andoccurrence of artefacts [30 31]

In past works we proposed several transform-basedapproaches to account for directional features in classifyingbiomedical images For instance in the case of brainmagneticresonance images we proposed a simple methodology in[32 33] where features are extracted from the LH and HLcomponents of the DWT instead of the more common LLor image approximation component We found that the LHand HL coefficients are efficient at characterizing changesin the biological tissue and help distinguish normal andabnormal image textures For mammograms we investigatedin [34] a hybrid processing system that sequentially usesthe discrete cosine transform (DCT) to obtain the high-frequency component of the mammogram and then appliesthe Radon transform (RT) to the result in order to extract itsdirectional featuresThe validation results showed that the RThelps improve the recognition rate of the detection system Insubsequent work we combined the DWT and RT transforms[35] The approach targeted the HH component of the DWTdecomposition and improved classification accuracy whencompared to using the DWT or RT alone or the DCT-RT used in [34] Our previous works clearly showed thatdirectional information helps improve classification accu-racy In addition the DWT-RT detection system was moreefficient for classifying normal and abnormal images than theDCT-RT possibly because of the multiresolution capabilityof the DWT and the fact that it leads to a sparser signalrepresentation than the DCT Still the RT cannot capturespatial frequency a potential feature to improve further theclassification accuracy

In this paper we describe a hybrid biomedical imageprocessing and classification system that uses both the DWTand Gabor filter as directional transforms and statisticalfeatures derived from them for the classification task whichis accomplished by support vector machines (SVMs) [36]As stated before the DWT is powerful at providing sparseand efficient image representations [14] However exceptfor the LH and HL subbands whose coefficients depend onimage row and column information respectively (an effectof the subband coding used by the algorithm) the standardDWT is essentially an image compression tool and it cannotperform directional analysis at arbitrary directions On theother hand the Gabor filter can process images in termsof preferred orientations at arbitrary spatial frequenciesMoreover it provides nonredundant information and canoffer high directional selectivity Thus combining DWT andGabor filter banks in sequence may lead to improved featureextraction from biomedical images and better classificationof normal versus abnormal images in comparison to usingDWT or Gabor filter banks alone In this hybrid processingscheme theDWTacts both as high-frequency filter to extractabrupt changes in image texture and image compressionengine to reduce image dimensionality and a Gabor filterbank extracts the directional information

Journal of Medical Engineering 3

In a preliminary work [37] the previously mentionedDWT-Gabor hybrid systemwas successfully applied tomam-mograms to extract features that allow discriminating normaland cancer images More specifically the goal was to detectthe presence of malign microcalcifications (specs of calciumin the breast tissue that appear in the mammogram as smallbright spots that are scattered or grouped in clusters) whoseearly detection is important for cancer screening [38 39]The results showed the superiority of the approach oversimply using the DWT alone In the present work wewiden our study to retina digital images and brain magneticresonance images to investigate the effectiveness of theDWT-Gabor approach across application domains with similarimage features Indeed the images of some pathologiesrelated to brain retina and breast present similar contrastfeatures characterized by abrupt changes in image texturewith directional properties (see examples in Figures 1 2and 3) For instance breast cancer is characterized by denseconcentration of contrast cells in the biological tissue cancerin brain magnetic resonance images is often characterized bylarge cells with high contrast and many forms of retinopathyare characterized by the presence of spots on the retina orcovering the macula As a result the DWT-Gabor hybridsystem we have used in our previous work [37] to detectcancer in mammograms could potentially also be applied tobrain magnetic resonance images and retina digital imageswith similar properties Next is a brief description of thepathologies that were studied in this work

Circinate retinopathy is a retinal degeneration charac-terized by a circle of white spots encircling the maculathat causes complete foveal blindness [40] retinal microa-neurysms are due to a swelling of the capillaries caused bya weakening of the vessel wall [41] and are considered tobe the earliest sign of diabetic retinopathy among othersMagnetic resonance imaging (MRI) is a noninvasive imagingmodality largely used for brain imaging to detect diseasessuch as Alzheimerrsquos and multiple sclerosis [6 8] Alzheimerrsquosdisease is the most frequent cause of age-related dementiaand multiple sclerosis is a progressive neurological disorderthat can result in various dysfunctions [42] Additional brainpathologies that can be detected fromMR images and that areinvestigated in this work include glioma herpes encephalitisand metastatic bronchogenic carcinoma (Figure 1) All ofthese are characterized by large cells with high contrasthence the interest in being able to detect them with the samealgorithm

The contribution of our work can be summarized asfollows First we propose a relatively simple and fast approachto biomedical image characterization that relies on the direc-tional properties of high-frequency componentsTheDWT isapplied first to extract high-frequency components that char-acterize abrupt changes in the biological tissue and then theGabor filter is applied to the obtainedHH subimage to extractdirectional features Second the statistical features extractedfrom the hybrid DWT-Gabor transform are processed by anSVM for classification This statistical binary classifier hasproven its efficiency [4ndash6 32ndash35 37] and ease of tuning incomparison to alternatives such as artificial neural networksAnother desirable feature is its scalability and ability to avoid

local minima [36] Third contrary to alternatives that focuson ROIs or specific image details the proposed methodologyis ofmore general reach as three different types of images usedfor validation show

The paper is organized as follows Section 2 reviewsprevious works related to the automatic classification ofnormal versus abnormal images in the context of brain mag-netic (MR) resonance imaging mammograms and retinadigital images Section 3 describes our proposed approach fordirectional features extraction from biomedical images usingdiscrete wavelet transform followed by Gabor filter banksand support vector machines classifier Section 4 presentsexperimental results Finally Section 5 draws the conclusionsand gives future work to be done

2 Related Works

Mammograms retina and MR images are the subject ofmany research efforts on feature extraction and subsequentclassification Next is a summary of some recent worksrelated to DWTs andor Gabor filters In the problem ofautomatic classification of mammograms the authors in [3]used Gabor filter banks to process images and 119896 nearestneighbour (119896-NN) algorithm as classifier The obtained clas-sification rate was 80 In [4] the dual-tree complex wavelettransform (DT-CWT) and support vector machine (SVM)were employed to classify benign and malignant images Theexperimental result achieved 8864 classification accuracyThe authors in [5] employed the contourlet transform andsuccessive enhancement learning (SEL) weighted SVM toobtain 966 correct classification rate The previous studiesall used images of size 1024 times 1024 pixels

In the problem of retina digital image classification theauthors in [1] employed the Belkynsrsquos shift-invariant DWTto classify normal against abnormal retina images of size700 times 605 pixels The pathologies of the abnormal imagesincluded exudates large drusen fine drusen choroidal neo-vascularization central vein and artery occlusion histoplas-mosis arteriosclerotic retinopathy hemicentral retinal veinocclusion and more In order to capture texture directionalfeatures they employed normalized gray level cooccurrencematrices (GLCMs)The obtained classification accuracy withlinear discriminant analysis (LDA) was 822 The authorsin [2] employed the probabilistic boosting algorithm andmorphological scale space analysis and GLCM to extracttexture features The purpose was to classify normal imagesversus drusen images with various texture complexities Thedetection accuracy of normal images varied between 813and 922 and that of abnormal images varied between717 and 852 depending on texture complexity (gradeof pathology) The authors in [43] used four approaches toextract features from retina digital images of size 300 times 300pixels to automatically classify glaucoma images The first setof features is obtained by taking the pixel intensities as inputto principal component analysis The second features areobtained from Gabor texture filter responses The third set offeatures is computed from the coefficients of the fast Fouriertransform and the fourth set of features is obtained from

4 Journal of Medical Engineering

(a) Normal (b) Alzheimerrsquos disease (c) Glioma

(d) Herpes encephalitis (e) Metastatic bronchogenic carcinoma (f) Multiple sclerosis

Figure 1 Examples of brain MR images

(a) Normal (b) Microaneurysms (c) Circinate

Figure 2 Examples of retina images

the histogram of the intensity distribution of the imageFinally support vector machines were employed for the clas-sification task The performance of the classifications usingone feature set only was 73 with the histogram features76with the fast Fourier transform coefficients 80with theGabor textures and 83 with the pixel intensities

Finally in the problem of brain MRI classification theauthors in [6] used the wavelet coefficients as input to asupport vector machine to classify normal and abnormal

Alzheimerrsquos disease images of size 256 times 256 pixels Theclassification accuracy was 98 using SVM with a radialbasis kernel More recently the authors in [42] used voxelsto represent each brain MRI of size 512 times 512 pixels Usingcross-validated tests the obtained correct classification ratesof normal and Alzheimer images were 90 92 and78 respectively when using classification by SVM naıveBayes classifier and voting feature intervals (VFIs) Stillmore recently the authors in [8] employed the DWT to

Journal of Medical Engineering 5

(a) Normal (b) Abnormal

Figure 3 Examples of mammograms

extract features from brain magnetic resonance images ofsize 256 times 256 pixels and then principal component analysiswas used to reduce the dimensions of the features space Theabnormal images included gliomameningiomaAlzheimerrsquosAlzheimerrsquos plus visual agnosia Pickrsquos disease sarcomaand Huntingtonrsquos disease The classification accuracies usingbackpropagation neural network (BPNN) were 100 usinglearning and testing sets of 33 images each

In this work we are interested in how a DWT-Gabor-based approach for feature extraction may provide betterclassification results than those reported in the previousworks particularly those based on the DWT alone The nextsection provides the details of our methodology

3 Methodology

The overall methodology proceeds as follows First theDWT is applied to the biomedical image to obtain its high-frequency image component since it often contains mostof the desired information about the biological tissue [39]Indeed sudden changes in the texture of the image are typicalindicators of the presence of abnormal biological tissueSecond a bank of Gabor filters with different scales andorientations is applied to the high-frequency image to obtainGabor-filtered images along different spatial orientationsThird statistical features are extracted from the Gabor-filtered images Finally the SVM is used to classify theresulting feature vector for final diagnosisThe block diagramof the DWT-Gabor system is shown in Figure 4 Figure 5summarizes the DWT approach in comparison

31 Discrete Wavelet Transform The two-dimensional dis-crete wavelet transform (2D-DWT) [14ndash16] performs asubband coding of an image in terms of spectral spa-tialfrequency components using an iterative and recursiveprocess Figure 6 illustrates the case of two-level decompo-sition The image is first represented by LH HL and HHsubbands that encode the image details in three directionsand an LL subband which provides an approximation of it

Source image DWTExtract DWT

high frequency

Gabor filter

Extract Gabor filtered images

Compute statistics

SVM Classification

Figure 4 Schematic diagram of the DWT-Gabor approach

Mammogram DWT Extract DWT high frequency

Compute statistics

SVMClassification

Figure 5 Schematic diagram of the DWT approach

The obtained detail or approximation images can be decom-posed again to obtain second-level detail and approximationimages and the process can be repeated for finer analysis aseach iteration doubles the image scale

The computation of the 2D-DWT proceeds from that ofthe 1D-DWT the discrete version of the one-dimensionalcontinuous wavelet transform The one-dimensional contin-uous wavelet transform of a signal 119909(119905) is defined by [7 8]

119882120595 (119886 119887) = int

+infin

minusinfin

119909 (119905) 120595lowast

119886119887(119905) 119889119905 (1)

120595119886119887 (119905) =1

radic|119886|120595(

119905 minus 119887

119886) (2)

where 120595119886119887(sdot) stands for a given wavelet function and 119886 and 119887are the scale and translation parameters respectivelyThe 1D-DWT is obtained by sampling 119886 and 119887 so that (1) becomes that

6 Journal of Medical Engineering

Image

LL

LH

HL

HH

darr

darr

darr

darr

darr

darr

119892(119899)

119892(119899)

119892(119899)

ℎ(119899)

ℎ(119899)ℎ(119899)

Figure 6 2D-DWT decomposition of an image

of a sequence In dyadic sampling 119886 and 119887 are respectively apower of 2 and multiples thereof and the sequence elements(wavelet coefficients) are given by

119888119895119896 = 119882120595 (2minus119895 2minus119895119896) (3a)

where 119895 represents the discrete scale factor and 119896 the discretetranslation factor In other words 119886 and 119887 in (1) are replacedby 2119895 and 2119895119896 respectively

The one-dimensional wavelet decomposition is extendedto an image by applying it to the row variable first and thento the column variable of the obtained result [44] At eachstep two subimages are created with half the number ofpixels of the row or column that was processed In the endan 119872 times 119873 image is decomposed into 4 subimages eachwith 1198722 times 1198732 resolution and preserved scale However(1) has only theoretical merit due to the infinite ranges of119886 and 119887 For a practical implementation the fact that (1)is essentially a measure of correlation between a signal andvarious wavelets derived from a mother is exploited and theDWT decomposition is turned into a filtering operation witha sequence of high-pass and low-pass filters [45] Followingthe notation in [7 8] the discrete form of (1) can then bewritten as

119888119886119895119896 [119909 (119905)] = DS [sum119909 (119905) 119892lowast

119895(119905 minus 2

119895119896)]

119888119889119895119896 [119909 (119905)] = DS [sum119909 (119905) ℎlowast

119895(119905 minus 2

119895119896)]

(4)

where coefficients 119888119886119895119896 and 119888119889119895119896 specify approximationand details components provided by the 119892(119899) low-pass andℎ(119899) high-pass impulse responses respectively and the DSoperator performs downsampling by a factor of 2 The one-dimensional wavelet decomposition is extended to two-dimensional objects by using row and column decomposi-tions as shown in Figure 6 In our work the most frequentlyused wavelet (Daubechies-4) [25] is considered to extract theHH image component

32 Gabor Filter The two-dimensional (2D) Gabor filterdecomposes an image into components corresponding todifferent scales and orientations [22] thus capturing visualproperties such as spatial localization orientation selectivityand spatial frequency The 2D Gabor filter consists of acomplex exponential centered at a given frequency andmodulated by a Gaussian envelope Because of the complex

exponential the filter has both real and imaginary parts Thegeneral form of the real part is defined as follows

119866(119909 119910 120590119909 120590119910 119891 120579) = exp[minus12((

1199091015840

120590119909

)

2

+ (1199101015840

120590119910

)

2

)]

times cos (21205871198911199091015840)

(5)

where

1199091015840= 119909 cos (120579) + 119910 sin (120579)

1199101015840= 119910 cos (120579) minus 119909 sin (120579)

(6)

and where 120590119909 and 120590119910 are the standard deviations of theGaussian envelope along the 119909 and 119910 axes The parameters 119891and 120579 are respectively the central frequency and the rotationof the Gabor filter To obtain the Gabor-filtered image119891(119909 119910)of a given image 119894(119909 119910) the 2D convolution operation (lowast) isperformed

119891 (119909 119910) = 119866 (119909 119910 120590119909 120590119910 119891 120579) lowast 119894 (119909 119910) (7)

The selection of parameters 120590119909 120590119910 119891 and 120579 plays animportant role in the filterrsquos operation However no formaltechnique exists for choosing them and experience-guidedintuition trial and error or heuristic searchmust be used Forretina digital images and brain MR images 120590119909 and 120590119910 werearbitrarily set to unity In the case of mammograms 120590119909 and120590119910 were set to the values used in [46] which were determinedempirically Consequently we used 120590119909 = 120591235 where 120591 isthe full width at half-maximumof theGaussian and120590119910 = 8120590119909No values of120590119909 and120590119910 other than the previous oneswere triedsince optimalitywas not the primary concern of thiswork andwe obtained satisfactory results with these values

Four orientations 120579 = 0 1205874 1205872 and 31205874 were used asin [22 33] These values seemed reasonable as a first try sincethey covered both image axis directions and the forward andbackward diagonals Finally the central frequency 119891 was setto 2 25 and 3 Given that the Gabor filter is modulated bythe cosine of 119891 large values of 119891 lead to a compressed cosineand consequently the filter output is more likely to showfast or frequent changes in biological tissue texture This inturn would help verify our hypothesis that abnormal imagesare characterized by sudden and frequent variations in imagetexture In the end the application of the Gabor filter bank tothe HH image component obtained with the 2D-DWT leadsto twelve Gabor-filtered HH images components for eachchoice of 119891 and 120579

33 Feature Extraction Statistical measures are employed toextract features from both the DWTHH subband image andthe real Gabor-filtered HH image components More pre-cisely the entropy (119864) and uniformity (119880) of the coefficientsof each one are computed Entropy and uniformity wereselected as features because previousworks onmammogramshave shown that uniformity is correlated with suspiciousmalignancy [47] and that entropy can successfully charac-terize breast biological tissue [48] In this study the entropy

Journal of Medical Engineering 7

and uniformity statistics are hypothesized to also characterizeretina andbrainMR imageswith similar contrast information(ie abrupt andor frequent variations in texture) Entropy(119864) and uniformity (119880) are defined by [49]

119864 = minussum119901 (119911) times log (119901 (119911))

119880 = sum1199012(119911)

(8)

where 119911 is a random variable that represents a coefficientin the Gabor filtered image and 119901(119911) is its probability ofoccurrence as estimated by its relative frequency

To investigate the performance of the previous approach-es the image features were extracted from HH at bothlevel-one DWT decomposition (HH1) and level-two DWTdecomposition (HH2) with and without filtering by a Gaborfilter bank We also applied the Gabor filter directly to theoriginal image without the DWT for comparison purposeFor eachDWTHH subband image the feature vector is givenby

119881DWT119886 = [119864119886 119880119886] (9)

where 119886 is the level of wavelet analysis (decomposition)Similarly for each of the twelve outputs generated by theGabor filter bank (4 angles times 3 central frequencies) theentropy and uniformity are computed and a twenty-fourcomponent feature vector is formed to represent the initialimage We thus have

119881Gabor119886 = [1198641119886 1198642119886 11986412119886 1198801119886 1198802119886 11988012119886] (10)

Either feature vector is subsequently fed to the SVM toclassify normal versus pathological images

34 The Support Vector Machine Classifier Introduced byVapnik [36] the support vector machine (SVM) classifieris based on statistical learning theory It implements theprinciple of structural risk minimization and has excellentgeneralization ability as a result even when the data sample issmall Moreover SVM can tolerate high-dimensional andorincomplete data [50] It has been used with great success invarious applications including speech emotion recognition[51] card-sharing traffic detection [52] fault diagnosis [53]cardiac decision making [54] Parkinsonrsquos disease diagnosis[55] and Alzheimerrsquos disease detection [56]

The support vector machine performs classification tasksby constructing an optimal separating hyperplane that maxi-mizes themargin between the two nearest data points belong-ing to two separate classes Given a training set (119909119894 119910119894) 119894 =1 2 119898 where the input 119909119894 isin 119877

119889 and class labels 119910119894 isin+1 minus1 the separation hyperplane for a linearly separablebinary classification problem is given by

119891 (119909) = ⟨119908 sdot 119909⟩ + 119887 (11)

where 119908 is a weight vector and 119887 is a bias The optimalseparation hyperplane is found by solving the followingoptimization problem

minimize119908119887120585

1

2⟨119908 sdot 119908⟩ + 119862

119898

sum

119894=1

120585119894

Subject to 119910119894 (⟨119908 sdot 119909119894⟩ + 119887) + 120585119894 minus 1 ge 0 120585119894 ge 0

(12)

where 119862 is a penalty parameter that controls the tradeoffbetween the complexity of the decision function and thenumber of misclassified training examples and 120585 is a positiveslack variableTheprevious optimizationmodel can be solvedby introducing Lagrange multipliers and using the Karush-Kuhn-Tucker theorem of optimization to obtain the solutionas

119908 =

119898

sum

119894=1

120572119894119910119894119909119894 (13)

The 119909119894 values corresponding to positive Lagrange multipliers120572119894 are called support vectors which define the decisionboundary The 119909119894 values corresponding to zero 120572119894 are irrel-evant Once the optimal solution 120572

lowast

119894is found the optimal

hyperplane parameters 119908lowast and 119887lowast are determined Then thediscriminant function of the SVM for a linearly separablebinary classification problem is [32]

119892 (119909) = sign(119898

sum

119894=1

119910119894120572lowast

119894⟨119909119894 sdot 119909⟩ + 119887

lowast) (14)

In the nonlinearly separable case the SVM classifier nonlin-early maps the training points to a high-dimensional featurespace using a kernel function Φ where linear separation canbe possible The scalar product ⟨Φ(119909119894) sdot Φ(119909119895)⟩ is computedby Mercer kernel function 119870 as 119870(119909119894 119909119895) = ⟨Φ(119909119894) sdot Φ(119909119895)⟩Then the nonlinear SVM classifier has the following form

119892 (119909) = sign(119898

sum

119894=1

119910119894120572lowast

119894119870⟨119909 119909119894⟩ + 119887

lowast) (15)

In this study a polynomial kernel of degree 2 was used forthe SVM As a global kernel it allows data points that are faraway from each other to also have an influence on the kernelvalues The general polynomial kernel is given by

119870(119909 119909119894) = ((119909119894 sdot 119909) + 1)119889 (16)

where 119889 is the degree of the polynomial to be used

4 Experimental Results

Asmentioned previouslymammograms and retina and brainMR images corresponding to given pathologies are consid-ered in this work and the aim is to classify normal versusabnormal images for each image category To do so onehundred digitalmammograms (171times 364 pixels) consisting offifty normal images and fifty cancer images were taken fromThe Digital Database for Screening Mammography (DDSM)

8 Journal of Medical Engineering

Table 1 Average SVM classification accuracy as a function of feature extraction method and level of DWT decompositionlowast

DWT DWT-Gabor DWT DWT-GaborDecomposition level One One Two TwoMammograms 9598 (plusmn004) 9667 (plusmn005) 8913 (plusmn001) 9109 (plusmn005)Retina 7469 (plusmn005) 100 9098 (plusmn003) 100Brain MRI 8780 (plusmn000) 9736 (plusmn002) 8576 (plusmn000) 9118 (plusmn004)lowastTenfold cross-validation used for mammograms and retina images leave-one-out used for brain MRIs

Table 2 SVM classification specificity as a function of feature extraction method and level of DWT decompositionlowast

DWT DWT-Gabor DWT DWT-GaborDecomposition level One One Two TwoMammograms 9781 (plusmn002) 100 8885 (plusmn005) 9209 (plusmn006)Retina 681 (plusmn004) 100 232 (plusmn010) 100Brain MRI 2155 (plusmn001) 9958 (plusmn001) 0 9724 (plusmn001)lowastTenfold cross-validation used for mammograms and retina images leave-one-out used for brain MRIs

[57] For retina a set of 69 color images (150 times 130 pixels)from the STARE [58] database were employed including23 normal images 24 with microaneurysms and 22 withcircinate Finally a collection of 56 axial T2-weighted andMR brain images (256 times 256 pixels) were taken from theAANLIB database [59] of the Harvard Medical School Theyconsisted of 7 normal images 9 with Alzheimerrsquos disease 13with glioma 8 with Herpes encephalitis 8 with metastaticbronchogenic carcinoma and 14 with multiple sclerosis Itis unfortunate that the number of images was not constantacross pathologies but we had no control over this and usedwhat was available with tenfold cross-validation or leave-one-out cross-validation of the results depending on samplesize All experiments were based on a binary classificationapproach of normal versus abnormal images Many kinds ofbiomedical images could be considered for our experimentswe focused on mammograms retina and brain magneticresonance images mainly because of public availability Anexample of the processing of a normal retina and a retina withcircinate is illustrated in Figures 7 and 8 respectively

For each image type the average and standard deviationof the correct classification rate (CCR) sensitivity and speci-ficity were computed to evaluate the performance featureextraction techniquewhen used in conjunctionwith the SVMclassifier The three performance measures are defined by

CCR =Classified Samples

Total Number of Samples

Sensitivity =Correctly Classified Positive Samples

True Positive Samples

Specificity =Correctly Classified Negative Samples

True Negative Samples

(17)

where positive samples and negative samples are respectivelyabnormal and normal images

Finally all experiments were performed with tenfoldcross-validation except those for MR images which usedleave-one-out cross-validation due to the small sample sizeof each brain image category

Table 1 shows the obtained average results for the threetypes of images that were investigated The performance ofthe SVM classifier improved for all types of images and alllevels of HH decomposition by the DWT At level one theaverage correct classification rate increased by respectively069 2531 and 956 percentage points for mammogramsretina and brain magnetic resonance images when usingthe DWT-Gabor approach At level-two decomposition theimprovement was respectively 196 902 and 542 percent-age points

Tables 2 and 3 provide the average results for classifiersensitivity and specificity At level-one DWT decompositionthe DWT-Gabor approach improved classification specificityfor mammograms and retina images to make it reach 100while it improved it by 7803 percentage points for brain MRimages At level-two DWT decomposition the improvementwas 324 percentage points for mammograms 9768 percent-age points for retina images (100 specificity) and 9724percentage points for MR images Regarding sensitivity theresults were mixed

At level-one DWT decomposition the values were aboutthe same for the DWT-Gabor and DWT-only approachesfor mammograms and retina images with respectivelyminus081 and 0 percentage points differences but there was adegradation of minus1133 and minus4616 percentage points for brainMR images at level one and level two of decompositionrespectively

Following the same cross-validation protocol we alsoconducted classification experiments with features extractedfrom aGabor filtered image of the original biomedical imageThe purpose was to check whether Gabor-based featuresalone help characterize images better than DWT or DWT-Gabor-based features The results are given in Table 4

Journal of Medical Engineering 9

20 40 60 80 100 120 140

20

40

60

80

100

120020406080100120140160180200

(a) Original normal retina image

10 20 30 40 50 60 70

10

20

30

40

50

60 minus60

minus40

minus20

0

20

40

60

(b) Normal retina HH1 subimage

0

5

10

15

20

25

30

35

20406080100

minus100

minus80

minus60

minus40

minus20

5 10 15 20 25 30 35 40(c) Normal retina HH2 subimage

10 20 30 40 50 60 70

10

20

30

40

50

60 10

20

30

40

50

60

70

(d) Gabor filtered HH1 at 119891 = 2 and 120579 = 1205874

5

10

15

20

25

30

355 10 15 20 25 30 35 40

102030405060708090100

(e) Gabor filtered HH2 119891 = 2 and 120579 = 1205874

Figure 7 Analysis of a normal retina

20 40 60 80 100 120 140

20

40

60

80

100

120 20406080100120140160180200220

(a) Original retina image with circinate

10 20 30 40 50 60 70

10

20

30

40

50

60minus20

minus10

0

10

20

30

(b) HH1 subimage

5

10

15

20

25

30

355 10 15 20 25 30 35 40

minus80

minus60

minus40

minus20020406080100

(c) HH2 subimage

10 20 30 40 50 60 70

10

20

30

40

50

600

5

10

15

20

25

30

35

(d) Gabor filtered HH1 at 119891 = 2 and 120579 = 1205874

102030405060708090100

5

10

15

20

25

30

355 10 15 20 25 30 35 40

(e) Gabor filtered HH2 at 119891 = 2 and 120579 = 1205874

Figure 8 Analysis of a retina with circinate

10 Journal of Medical Engineering

Table 3 SVM classification sensitivity as a function of feature extraction method and level of DWT decompositionlowast

DWT DWT-Gabor DWT DWT-GaborDecomposition level One One Two TwoMammograms 9414 (plusmn006) 9333 (plusmn006) 9029 (plusmn0039) 8978 (plusmn004)Retina 100 100 100 100Brain MRI 9384 (plusmn000) 8251 (plusmn016) 100 5384 (plusmn023)lowastTenfold cross-validation used for mammograms and retina images leave-one-out used for brain MRIs

Table 4 SVM classification performance measures obtained withGabor-based features

Accuracy Specificity SensitivityMammograms 6803 (plusmn001) 100 0Retina 5000 (plusmn000) 100 0Brain MRI 8661 (plusmn003) 100 0

The obtained correct classification rate of mammogramsretina and brain magnetic resonance images is respectively6803 (plusmn001) 5000 (plusmn000) and 8661 (plusmn003) Theaverage results for classifier specificity and sensitivity for allimages are 100 and 0 This finding indicates that Gabor-based features are suitable to detect pathological imagesbut fails to detect normal images In sum the results showthat Gabor-based features do not perform better than DWTand DWT-Gabor-based features These findings confirm thesuperiority of combining theDWT andGabor filter banks forfeature extraction

Based on the previous results it appears that the DWT-Gabor approach for feature extraction is effective for detect-ing the abrupt changes in biological tissue that characterizethe pathological patterns that were investigated and it yieldsbetter classification accuracy and specificity than the DWT-only approach It also offers equal of better sensitivity exceptfor brain MRIs For brain MRIs the obtained specificityand sensitivity results with the DWT-Gabor approach showimproved true negative detection but lower true positiveperformance Finally the obtained results reveal also thatlevel-one DWT decomposition is preferable to level-twodecomposition

Finally Table 5 compares the results obtained with theDWT-Gabor approach to other work that we surveyed Inmany cases the DWT-Gabor method yields higher clas-sification rates particularly for mammograms and retinaFor the problem of brain MRI classification our obtainedperformance is better than the results of [38] but less thanwhat is reported in [6 8] However these comparisons shouldbe viewed with caution as not all the results stem froma common image database and the different authors usedifferent sample and image sizesMoreovermany authors useno cross-validation and simply performa single arbitrary splitof their data into training and test sets to obtain their accuracyresults Obviously one cannot generalize or draw definiteconclusion from such efforts and comparisons between

works cannot be made other than in general terms In thisrespect it can only be concluded from our results that theDWT-Gabor for feature extraction is effective for obtaininghigh image classification accuracy by an SVM and that itmay outperform other feature extraction and classificationtechniques reported in the literature at least those basedon DWT-only image decomposition Unfortunately a moredefinite conclusion is impossiblewithout gaining access to theimage databases used by the other authors

5 Computational Complexity

Finally the computational complexity of the DWT Gaborcontourlet and curvelet for an 119873 times 119873 image is respectively119874(119873) 119874(1198732 times 119872

2) with 119872 being the width of Gabor

(Gaussian) mask filter119874(1198732) and119874(1198732 log(119873)) As a resultthe computational complexity of the combination of theDWT and Gabor filter is 119874(119873) + 119874(1198732 times 1198722) In terms offeatures extraction processing time the average time requiredto process a brain a mammogram and a retina imagewith the DWT approach (DWT-Gabor) was respectively019 (031) 017 (032) and 015 seconds (035) using MatlabR2009a on a 15 GHz Core2 Duo processor

6 Conclusion

We proposed a supervised system for biomedical imagesclassification that uses statistical features obtained from thecombination of the discrete wavelet transform and Gaborfilter to classify normal images versus cancer images usingsupport vector machines as classifiers Our experimentalresults show that such a hybrid processing model achieveshigher accuracy in comparison to using DWT or Gaborfilter banks alone Therefore the proposed image processingand features extraction approach seem to be very promis-ing for the detection of certain pathologies in biomedicalimages

For future works it is recommended to consider a largerset of features and a selection process to identify the mostdiscriminant ones In addition the Gabor parameters willbe adjusted for each type of image separately to improvethe accuracy Furthermore the DWT-Gabor will be directlycompared to the dual-tree complex wavelet curvelet andcontourlet using the same databases and images in orderto draw general conclusions Also multilabels classificationswill be considered in future works to investigate the discrim-inative power of our approach for each type of pathology

Journal of Medical Engineering 11

Table 5 Comparison with the literature

Features Classifier Accuracylowast

Mammograms[3] Gabor k-NN 80[4] DT-CWT SVM 8864[5] Contourlet SVM 966Our approach DWT-Gabor SVM 9667 (plusmn005)

Retina[1] DWT + GLCM LDA 822

[2] Morphological + GLCM Probabilistic boosting algorithm 813ndash922717ndash852

[39] Gabor SVM 83Our approach DWT-Gabor SVM 100

Brain[6] DWT SVM 98[8] DWT + PCA BPNN 100

SVM 90[38] Voxels Bayes 92

VFI 78Our approach DWT-Gabor SVM 9736 (plusmn002)

lowastCorrect classification rate

Finally more experiments on the effect of kernel choice andits parameter on classification accuracy will be investigated

References

[1] A Khademi and S Krishnan ldquoShift-invariant discrete wavelettransform analysis for retinal image classificationrdquoMedical andBiological Engineering and Computing vol 45 no 12 pp 1211ndash1222 2007

[2] N Lee A F Laine andT R Smith ldquoLearning non-homogenoustextures and the unlearning problemwith application to drusendetection in retinal imagesrdquo in Proceedings of the 5th IEEEInternational Symposium on Biomedical Imaging From Nano toMacro (ISBI rsquo08) pp 1215ndash1218 Paris France May 2008

[3] A Dong and B Wang ldquoFeature selection and analysis onmammogram classificationrdquo in Proceedings of the IEEE PacificRim Conference on Communications Computers and SignalProcessing (PACRIM rsquo09) pp 731ndash735 Victoria BC CanadaAugust 2009

[4] A Tirtajaya and D D Santika ldquoClassification of microcalcifi-cation using dual-tree complex wavelet transform and supportvector machinerdquo in Proceedings of the 2nd International Confer-ence onAdvances in Computing Control and TelecommunicationTechnologies (ACT rsquo10) pp 164ndash166 Jakarta Indonesia Decem-ber 2010

[5] F Moayedi Z Azimifar R Boostani and S Katebi ldquoContour-let-based mammography mass classification using the SVMfamilyrdquo Computers in Biology and Medicine vol 40 no 4 pp373ndash383 2010

[6] S Chaplot LM Patnaik andN R Jagannathan ldquoClassificationof magnetic resonance brain images using wavelets as input tosupport vector machine and neural networkrdquo Biomedical SignalProcessing and Control vol 1 no 1 pp 86ndash92 2006

[7] Y Zhang S Wang and L Wu ldquoA novel method for magneticresonance brain image classification based on adaptive chaoticPSOrdquo Progress in Electromagnetics Research vol 109 pp 325ndash343 2010

[8] Y Zhang Z Dong L Wu and S Wang ldquoA hybrid method forMRI brain image classificationrdquo Expert Systems with Applica-tions vol 38 no 8 pp 10049ndash10053 2011

[9] M E Celebi H Iyatomi G Schaefer and W V StoeckerldquoLesion border detection in dermoscopy imagesrdquoComputerizedMedical Imaging and Graphics vol 33 no 2 pp 148ndash153 2009

[10] Q AbbasM E Celebi and I F Garcıa ldquoSkin tumor area extrac-tion using an improved dynamic programming approachrdquo SkinResearch and Technology vol 18 pp 133ndash142 2012

[11] Q Li F Li and K Doi ldquoComputerized detection of lungnodules in thin-section CT images by use of selective enhance-ment filters and an automated rule-based classifierrdquo AcademicRadiology vol 15 no 2 pp 165ndash175 2008

[12] A El-Bazl M Nitzken E Vanbogaertl G Gimelrsquojarb RFalfi and M Abo El-Ghar ldquoA novel shaped-based diagnosticapproach for early diagnosis of lung nodulesrdquo in Proceedingsof the IEEE International Symposium in Biomedical Imaging(ISBI rsquo11) pp 137ndash140 Chicago Ill USA 2011

[13] M T Coimbra and J P S Cunha ldquoMPEG-7 visual descriptorsmdashcontributions for automated feature extraction in capsuleendoscopyrdquo IEEE Transactions on Circuits and Systems for VideoTechnology vol 16 no 5 pp 628ndash636 2006

[14] B Li and M Q H Meng ldquoTexture analysis for ulcer detectionin capsule endoscopy imagesrdquo Image andVision Computing vol27 no 9 pp 1336ndash1342 2009

[15] C K Chui An Introduction to Wavelets Academic Press SanDiego Calif USA 1992

[16] M Vetterli andCHerley ldquoWavelets and filter banks theory anddesignrdquo IEEE Transactions on Signal Processing vol 40 no 9pp 2207ndash2232 1992

12 Journal of Medical Engineering

[17] J G Daugman ldquoUncertainty relation for resolution inspace spatial frequency and orientation optimized by two-dimensional visual cortical filtersrdquo Journal of the Optical Societyof America A vol 2 no 7 pp 1160ndash1169 1985

[18] I W Selesnick R G Baraniuk and N G Kingsbury ldquoThedual-tree complex wavelet transformrdquo IEEE Signal ProcessingMagazine vol 22 no 6 pp 123ndash151 2005

[19] E Candes and D Donoho ldquoRidgelets a key to higher-dimensional intermittencyrdquo Philosophical Transactions of theLondon Royal Society vol 357 pp 2495ndash2509 1999

[20] E J Candes and D L Donoho ldquoContinuous curvelettransformmdashI Resolution of the wavefront setrdquo Applied andComputational Harmonic Analysis vol 19 no 2 pp 162ndash1972005

[21] M N Do and M Vetterli ldquoThe contourlet transform an effi-cient directional multiresolution image representationrdquo IEEETransactions on Image Processing vol 14 no 12 pp 2091ndash21062005

[22] R J Ferrari R M Rangayyan J E L Desautels and A FFrere ldquoAnalysis of asymmetry in mammograms via directionalfiltering with Gabor waveletsrdquo IEEE Transactions on MedicalImaging vol 20 no 9 pp 953ndash964 2001

[23] Z Cui and G Zhang ldquoA novel medical image dynamic fuzzyclassification model based on ridgelet transformrdquo Journal ofSoftware vol 5 no 5 pp 458ndash465 2010

[24] T Geback and P Koumoutsakos ldquoEdge detection inmicroscopyimages using curveletsrdquo BMC Bioinformatics vol 10 article 752009

[25] J Ma andG Plonka ldquoThe curvelet transform a review of recentapplicationsrdquo IEEE Signal ProcessingMagazine vol 27 no 2 pp118ndash133 2010

[26] N Kingsbury ldquoComplex wavelets and shift invariancerdquo inProceedings of the IEEE Seminar on Time-Scale and Time-Frequency Analysis and Applications pp 501ndash510 London UK2000

[27] Y L Qiao C Y Song and C H Zhao ldquoM-band ridgelettransform based texture classificationrdquo Pattern RecognitionLetters vol 31 no 3 pp 244ndash249 2010

[28] F Gomez and E Romero ldquoTexture characterization using acurvelet based descriptorrdquo Lecture Notes in Computer Sciencevol 5856 pp 113ndash120 2009

[29] H Shan and J Ma ldquoCurvelet-based geodesic snakes for imagesegmentation with multiple objectsrdquo Pattern Recognition Let-ters vol 31 no 5 pp 355ndash360 2010

[30] R Eslami and H Radha ldquoNew image transforms using hybridwavelets and directional filter banks analysis and designrdquo inProceedings of the IEEE International Conference on ImageProcessing (ICIP rsquo05) pp 733ndash736 Genova Italy September2005

[31] O O V Villegas and V G C Sanchez ldquoThe wavelet basedcontourlet transform and its application to feature preservingimage codingrdquo Lecture Notes in Computer Science vol 4827 pp590ndash600 2007

[32] S Lahmir and M Boukadoum ldquoClassification of brain MRIusing the LH and HL wavelet transform sub-bandsrdquo in Pro-ceedings of the IEEE International Symposium on Circuits andSystems (ISCAS rsquo11) pp 1025ndash1028 Rio de Janeiro Brazil May2009 2011

[33] S Lahmir and M Boukadoum ldquoBrain MRI classificationusing an ensemble system and LH and HL wavelet Sub-bandsFeaturesrdquo in Proceedings of the IEEE Symposium Series on

Computational Intelligence (SSCI rsquo11) pp 1ndash7 Paris FranceApril2011

[34] S Lahmir and M Boukadoum ldquoHybrid Cosine and RadonTransform-based processing for Digital Mammogram FeatureExtraction and Classification with SVMrdquo in Proceedings of the33rd IEEE Annual International Conference on Engineering inMedecine and Biology Society (EMBS rsquo11) pp 5104ndash5107 BostonMass USA 2011

[35] S Lahmir andM Boukadoum ldquoDWT and RT-Based Approachfor Feature Extraction and classification of Mammograms withSVMrdquo in Proceedings of the IEEE Biomedical Circuits andSystems Conference (BioCAS rsquo11) pp 412ndash415 San Diego CalifUSA November 2011

[36] VN VapnikTheNature of Statistical LearningTheory Springer1995

[37] S Lahmiri and M Boukadoum ldquoHybrid discret wavelet trans-form and Gabor filter banks processing for mammogramfeatures extractionrdquo in Proceedings of the IEEE New Circuits andSystems (NEWCAS rsquo11) pp 53ndash56 Bordeaux France June 2011

[38] L M Bruce and N Shanmugam ldquoUsing neural networkswith wavelet transforms for an automated mammographicmass classifierrdquo in Proceedings of the 22nd Annual InternationalConference of the IEEE Engineering in Medicine and BiologySociety pp 985ndash987 Chicago Ill USA July 2000

[39] S M H Jamarani G Rezai-rad and H Behnam ldquoA novelmethod for breast cancer prognosis using wavelet packet basedneural networkrdquo in Proceedings of the 27th Annual InternationalConference of the Engineering in Medicine and Biology Society(IEEE-EMBS rsquo05) pp 3414ndash3417 Shanghai China September2005

[40] httpmedical-dictionarythefreedictionarycomcircinate+retinopathy

[41] C I O Martins F N S Medeiros R M S Veras F N Bezerraand R M Cesar Jr ldquoEvaluation of retinal vessel segmentationmethods for microaneurysms detectionrdquo in Proceedings of theIEEE International Conference on Image Processing (ICIP rsquo09)pp 3365ndash3368 Cairo Egypt November 2009

[42] C Plant S J Teipel A Oswald et al ldquoAutomated detectionof brain atrophy patterns based on MRI for the prediction ofAlzheimerrsquos diseaserdquo NeuroImage vol 50 no 1 pp 162ndash1742010

[43] J Meier R Bock L G Nyul and G Michelson ldquoEye fun-dus image processing system for automated glaucoma clas-sificationrdquo in Proceedings of the 52nd Internationales Wis-senschaftliches Kolloquium Technische Universitat Ilmenau2007

[44] E Sakka A Prentza I E Lamprinos and D KoutsourisldquoMicrocalcification detection using multiresolution analysisbased on wavelet transformrdquo in Proceedings of the IEEE Inter-national Special Topic Conference on Information Technology inBiomedicine Ioannina Greece October 2006

[45] S G Mallat ldquoTheory for multiresolution signal decompositionthe wavelet representationrdquo IEEE Transactions on Pattern Anal-ysis and Machine Intelligence vol 11 no 7 pp 674ndash693 1989

[46] J Suckling J Parker and D R Dance ldquoThe mammographicimage analysis society digital mammogram databaserdquo in Pro-ceedings of the the 2nd International Workshop on DigitalMammography A G Gale S M Astley D D Dance and AY Cairns Eds pp 375ndash378 Elsevier York UK 1994

[47] H J Chiou C Y Chen T C Liu et al ldquoComputer-aideddiagnosis of peripheral soft tissue masses based on ultrasound

Journal of Medical Engineering 13

imagingrdquo Computerized Medical Imaging and Graphics vol 33no 5 pp 408ndash413 2009

[48] J K Kim J M Park K S Song and H W Park ldquoAdaptivemammographic image enhancement using first derivative andlocal statisticsrdquo IEEE Transactions on Medical Imaging vol 16no 5 pp 495ndash502 1997

[49] H S Sheshadri andAKandaswamy ldquoBreast tissue classificationusing statistical feature extraction of mammogramsrdquo MedicalImaging and Information Sciences vol 23 no 3 pp 105ndash1072006

[50] N Cristianini and J Shawe-Taylor Introduction to SupportVector Machines and Other Kernel-Based Learning MethodsCambridge University Press Cambridge UK 2000

[51] L Chen X Mao Y Xue and L L Cheng ldquoSpeech emotionrecognition features and classification modelsrdquo Digital SignalProcessing vol 22 pp 1154ndash1160 2012

[52] F Palmieri U Fiore A Castiglione and A De Santis ldquoOn thedetection of card-sharing traffic through wavelet analysis andSupport Vector Machinesrdquo Applied Soft Computing vol 13 no1 pp 615ndash627 2013

[53] A Azadeh M Saberi A Kazem V Ebrahimipour A Nour-mohammadzadeh and Z Saberi ldquoA flexible algorithm forfault diagnosis in a centrifugal pump with corrupted data andnoise based on ANN and support vector machine with hyper-parameters optimizationrdquoApplied Soft Computing vol 13 no 3pp 1478ndash1485 2013

[54] R J Martis U R Acharya K M Mandana A K Ray andC Chakraborty ldquoCardiac decision making using higher orderspectrardquo Biomedical Signal Processing and Control vol 8 pp193ndash203 2013

[55] M R Mohammad ldquoChi-square distance kernel of the gaitsfor the diagnosis of Parkinsonrsquos diseaserdquo Biomedical SignalProcessing and Control vol 8 pp 66ndash70 2013

[56] R Vandenberghe N Nelissen E Salmon et al ldquoBinary clas-sification of 18F-flutemetamol PET using machine learningcomparison with visual reads and structuralMRIrdquoNeuroImagevol 64 no 1 pp 517ndash525 2013

[57] httpmarathoncseeusfeduMammographyDatabasehtml[58] httpwwwcesclemsonedusimahooverstare[59] httpwwwmedharvardeduaanlib

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Journal of Medical Engineering 3

In a preliminary work [37] the previously mentionedDWT-Gabor hybrid systemwas successfully applied tomam-mograms to extract features that allow discriminating normaland cancer images More specifically the goal was to detectthe presence of malign microcalcifications (specs of calciumin the breast tissue that appear in the mammogram as smallbright spots that are scattered or grouped in clusters) whoseearly detection is important for cancer screening [38 39]The results showed the superiority of the approach oversimply using the DWT alone In the present work wewiden our study to retina digital images and brain magneticresonance images to investigate the effectiveness of theDWT-Gabor approach across application domains with similarimage features Indeed the images of some pathologiesrelated to brain retina and breast present similar contrastfeatures characterized by abrupt changes in image texturewith directional properties (see examples in Figures 1 2and 3) For instance breast cancer is characterized by denseconcentration of contrast cells in the biological tissue cancerin brain magnetic resonance images is often characterized bylarge cells with high contrast and many forms of retinopathyare characterized by the presence of spots on the retina orcovering the macula As a result the DWT-Gabor hybridsystem we have used in our previous work [37] to detectcancer in mammograms could potentially also be applied tobrain magnetic resonance images and retina digital imageswith similar properties Next is a brief description of thepathologies that were studied in this work

Circinate retinopathy is a retinal degeneration charac-terized by a circle of white spots encircling the maculathat causes complete foveal blindness [40] retinal microa-neurysms are due to a swelling of the capillaries caused bya weakening of the vessel wall [41] and are considered tobe the earliest sign of diabetic retinopathy among othersMagnetic resonance imaging (MRI) is a noninvasive imagingmodality largely used for brain imaging to detect diseasessuch as Alzheimerrsquos and multiple sclerosis [6 8] Alzheimerrsquosdisease is the most frequent cause of age-related dementiaand multiple sclerosis is a progressive neurological disorderthat can result in various dysfunctions [42] Additional brainpathologies that can be detected fromMR images and that areinvestigated in this work include glioma herpes encephalitisand metastatic bronchogenic carcinoma (Figure 1) All ofthese are characterized by large cells with high contrasthence the interest in being able to detect them with the samealgorithm

The contribution of our work can be summarized asfollows First we propose a relatively simple and fast approachto biomedical image characterization that relies on the direc-tional properties of high-frequency componentsTheDWT isapplied first to extract high-frequency components that char-acterize abrupt changes in the biological tissue and then theGabor filter is applied to the obtainedHH subimage to extractdirectional features Second the statistical features extractedfrom the hybrid DWT-Gabor transform are processed by anSVM for classification This statistical binary classifier hasproven its efficiency [4ndash6 32ndash35 37] and ease of tuning incomparison to alternatives such as artificial neural networksAnother desirable feature is its scalability and ability to avoid

local minima [36] Third contrary to alternatives that focuson ROIs or specific image details the proposed methodologyis ofmore general reach as three different types of images usedfor validation show

The paper is organized as follows Section 2 reviewsprevious works related to the automatic classification ofnormal versus abnormal images in the context of brain mag-netic (MR) resonance imaging mammograms and retinadigital images Section 3 describes our proposed approach fordirectional features extraction from biomedical images usingdiscrete wavelet transform followed by Gabor filter banksand support vector machines classifier Section 4 presentsexperimental results Finally Section 5 draws the conclusionsand gives future work to be done

2 Related Works

Mammograms retina and MR images are the subject ofmany research efforts on feature extraction and subsequentclassification Next is a summary of some recent worksrelated to DWTs andor Gabor filters In the problem ofautomatic classification of mammograms the authors in [3]used Gabor filter banks to process images and 119896 nearestneighbour (119896-NN) algorithm as classifier The obtained clas-sification rate was 80 In [4] the dual-tree complex wavelettransform (DT-CWT) and support vector machine (SVM)were employed to classify benign and malignant images Theexperimental result achieved 8864 classification accuracyThe authors in [5] employed the contourlet transform andsuccessive enhancement learning (SEL) weighted SVM toobtain 966 correct classification rate The previous studiesall used images of size 1024 times 1024 pixels

In the problem of retina digital image classification theauthors in [1] employed the Belkynsrsquos shift-invariant DWTto classify normal against abnormal retina images of size700 times 605 pixels The pathologies of the abnormal imagesincluded exudates large drusen fine drusen choroidal neo-vascularization central vein and artery occlusion histoplas-mosis arteriosclerotic retinopathy hemicentral retinal veinocclusion and more In order to capture texture directionalfeatures they employed normalized gray level cooccurrencematrices (GLCMs)The obtained classification accuracy withlinear discriminant analysis (LDA) was 822 The authorsin [2] employed the probabilistic boosting algorithm andmorphological scale space analysis and GLCM to extracttexture features The purpose was to classify normal imagesversus drusen images with various texture complexities Thedetection accuracy of normal images varied between 813and 922 and that of abnormal images varied between717 and 852 depending on texture complexity (gradeof pathology) The authors in [43] used four approaches toextract features from retina digital images of size 300 times 300pixels to automatically classify glaucoma images The first setof features is obtained by taking the pixel intensities as inputto principal component analysis The second features areobtained from Gabor texture filter responses The third set offeatures is computed from the coefficients of the fast Fouriertransform and the fourth set of features is obtained from

4 Journal of Medical Engineering

(a) Normal (b) Alzheimerrsquos disease (c) Glioma

(d) Herpes encephalitis (e) Metastatic bronchogenic carcinoma (f) Multiple sclerosis

Figure 1 Examples of brain MR images

(a) Normal (b) Microaneurysms (c) Circinate

Figure 2 Examples of retina images

the histogram of the intensity distribution of the imageFinally support vector machines were employed for the clas-sification task The performance of the classifications usingone feature set only was 73 with the histogram features76with the fast Fourier transform coefficients 80with theGabor textures and 83 with the pixel intensities

Finally in the problem of brain MRI classification theauthors in [6] used the wavelet coefficients as input to asupport vector machine to classify normal and abnormal

Alzheimerrsquos disease images of size 256 times 256 pixels Theclassification accuracy was 98 using SVM with a radialbasis kernel More recently the authors in [42] used voxelsto represent each brain MRI of size 512 times 512 pixels Usingcross-validated tests the obtained correct classification ratesof normal and Alzheimer images were 90 92 and78 respectively when using classification by SVM naıveBayes classifier and voting feature intervals (VFIs) Stillmore recently the authors in [8] employed the DWT to

Journal of Medical Engineering 5

(a) Normal (b) Abnormal

Figure 3 Examples of mammograms

extract features from brain magnetic resonance images ofsize 256 times 256 pixels and then principal component analysiswas used to reduce the dimensions of the features space Theabnormal images included gliomameningiomaAlzheimerrsquosAlzheimerrsquos plus visual agnosia Pickrsquos disease sarcomaand Huntingtonrsquos disease The classification accuracies usingbackpropagation neural network (BPNN) were 100 usinglearning and testing sets of 33 images each

In this work we are interested in how a DWT-Gabor-based approach for feature extraction may provide betterclassification results than those reported in the previousworks particularly those based on the DWT alone The nextsection provides the details of our methodology

3 Methodology

The overall methodology proceeds as follows First theDWT is applied to the biomedical image to obtain its high-frequency image component since it often contains mostof the desired information about the biological tissue [39]Indeed sudden changes in the texture of the image are typicalindicators of the presence of abnormal biological tissueSecond a bank of Gabor filters with different scales andorientations is applied to the high-frequency image to obtainGabor-filtered images along different spatial orientationsThird statistical features are extracted from the Gabor-filtered images Finally the SVM is used to classify theresulting feature vector for final diagnosisThe block diagramof the DWT-Gabor system is shown in Figure 4 Figure 5summarizes the DWT approach in comparison

31 Discrete Wavelet Transform The two-dimensional dis-crete wavelet transform (2D-DWT) [14ndash16] performs asubband coding of an image in terms of spectral spa-tialfrequency components using an iterative and recursiveprocess Figure 6 illustrates the case of two-level decompo-sition The image is first represented by LH HL and HHsubbands that encode the image details in three directionsand an LL subband which provides an approximation of it

Source image DWTExtract DWT

high frequency

Gabor filter

Extract Gabor filtered images

Compute statistics

SVM Classification

Figure 4 Schematic diagram of the DWT-Gabor approach

Mammogram DWT Extract DWT high frequency

Compute statistics

SVMClassification

Figure 5 Schematic diagram of the DWT approach

The obtained detail or approximation images can be decom-posed again to obtain second-level detail and approximationimages and the process can be repeated for finer analysis aseach iteration doubles the image scale

The computation of the 2D-DWT proceeds from that ofthe 1D-DWT the discrete version of the one-dimensionalcontinuous wavelet transform The one-dimensional contin-uous wavelet transform of a signal 119909(119905) is defined by [7 8]

119882120595 (119886 119887) = int

+infin

minusinfin

119909 (119905) 120595lowast

119886119887(119905) 119889119905 (1)

120595119886119887 (119905) =1

radic|119886|120595(

119905 minus 119887

119886) (2)

where 120595119886119887(sdot) stands for a given wavelet function and 119886 and 119887are the scale and translation parameters respectivelyThe 1D-DWT is obtained by sampling 119886 and 119887 so that (1) becomes that

6 Journal of Medical Engineering

Image

LL

LH

HL

HH

darr

darr

darr

darr

darr

darr

119892(119899)

119892(119899)

119892(119899)

ℎ(119899)

ℎ(119899)ℎ(119899)

Figure 6 2D-DWT decomposition of an image

of a sequence In dyadic sampling 119886 and 119887 are respectively apower of 2 and multiples thereof and the sequence elements(wavelet coefficients) are given by

119888119895119896 = 119882120595 (2minus119895 2minus119895119896) (3a)

where 119895 represents the discrete scale factor and 119896 the discretetranslation factor In other words 119886 and 119887 in (1) are replacedby 2119895 and 2119895119896 respectively

The one-dimensional wavelet decomposition is extendedto an image by applying it to the row variable first and thento the column variable of the obtained result [44] At eachstep two subimages are created with half the number ofpixels of the row or column that was processed In the endan 119872 times 119873 image is decomposed into 4 subimages eachwith 1198722 times 1198732 resolution and preserved scale However(1) has only theoretical merit due to the infinite ranges of119886 and 119887 For a practical implementation the fact that (1)is essentially a measure of correlation between a signal andvarious wavelets derived from a mother is exploited and theDWT decomposition is turned into a filtering operation witha sequence of high-pass and low-pass filters [45] Followingthe notation in [7 8] the discrete form of (1) can then bewritten as

119888119886119895119896 [119909 (119905)] = DS [sum119909 (119905) 119892lowast

119895(119905 minus 2

119895119896)]

119888119889119895119896 [119909 (119905)] = DS [sum119909 (119905) ℎlowast

119895(119905 minus 2

119895119896)]

(4)

where coefficients 119888119886119895119896 and 119888119889119895119896 specify approximationand details components provided by the 119892(119899) low-pass andℎ(119899) high-pass impulse responses respectively and the DSoperator performs downsampling by a factor of 2 The one-dimensional wavelet decomposition is extended to two-dimensional objects by using row and column decomposi-tions as shown in Figure 6 In our work the most frequentlyused wavelet (Daubechies-4) [25] is considered to extract theHH image component

32 Gabor Filter The two-dimensional (2D) Gabor filterdecomposes an image into components corresponding todifferent scales and orientations [22] thus capturing visualproperties such as spatial localization orientation selectivityand spatial frequency The 2D Gabor filter consists of acomplex exponential centered at a given frequency andmodulated by a Gaussian envelope Because of the complex

exponential the filter has both real and imaginary parts Thegeneral form of the real part is defined as follows

119866(119909 119910 120590119909 120590119910 119891 120579) = exp[minus12((

1199091015840

120590119909

)

2

+ (1199101015840

120590119910

)

2

)]

times cos (21205871198911199091015840)

(5)

where

1199091015840= 119909 cos (120579) + 119910 sin (120579)

1199101015840= 119910 cos (120579) minus 119909 sin (120579)

(6)

and where 120590119909 and 120590119910 are the standard deviations of theGaussian envelope along the 119909 and 119910 axes The parameters 119891and 120579 are respectively the central frequency and the rotationof the Gabor filter To obtain the Gabor-filtered image119891(119909 119910)of a given image 119894(119909 119910) the 2D convolution operation (lowast) isperformed

119891 (119909 119910) = 119866 (119909 119910 120590119909 120590119910 119891 120579) lowast 119894 (119909 119910) (7)

The selection of parameters 120590119909 120590119910 119891 and 120579 plays animportant role in the filterrsquos operation However no formaltechnique exists for choosing them and experience-guidedintuition trial and error or heuristic searchmust be used Forretina digital images and brain MR images 120590119909 and 120590119910 werearbitrarily set to unity In the case of mammograms 120590119909 and120590119910 were set to the values used in [46] which were determinedempirically Consequently we used 120590119909 = 120591235 where 120591 isthe full width at half-maximumof theGaussian and120590119910 = 8120590119909No values of120590119909 and120590119910 other than the previous oneswere triedsince optimalitywas not the primary concern of thiswork andwe obtained satisfactory results with these values

Four orientations 120579 = 0 1205874 1205872 and 31205874 were used asin [22 33] These values seemed reasonable as a first try sincethey covered both image axis directions and the forward andbackward diagonals Finally the central frequency 119891 was setto 2 25 and 3 Given that the Gabor filter is modulated bythe cosine of 119891 large values of 119891 lead to a compressed cosineand consequently the filter output is more likely to showfast or frequent changes in biological tissue texture This inturn would help verify our hypothesis that abnormal imagesare characterized by sudden and frequent variations in imagetexture In the end the application of the Gabor filter bank tothe HH image component obtained with the 2D-DWT leadsto twelve Gabor-filtered HH images components for eachchoice of 119891 and 120579

33 Feature Extraction Statistical measures are employed toextract features from both the DWTHH subband image andthe real Gabor-filtered HH image components More pre-cisely the entropy (119864) and uniformity (119880) of the coefficientsof each one are computed Entropy and uniformity wereselected as features because previousworks onmammogramshave shown that uniformity is correlated with suspiciousmalignancy [47] and that entropy can successfully charac-terize breast biological tissue [48] In this study the entropy

Journal of Medical Engineering 7

and uniformity statistics are hypothesized to also characterizeretina andbrainMR imageswith similar contrast information(ie abrupt andor frequent variations in texture) Entropy(119864) and uniformity (119880) are defined by [49]

119864 = minussum119901 (119911) times log (119901 (119911))

119880 = sum1199012(119911)

(8)

where 119911 is a random variable that represents a coefficientin the Gabor filtered image and 119901(119911) is its probability ofoccurrence as estimated by its relative frequency

To investigate the performance of the previous approach-es the image features were extracted from HH at bothlevel-one DWT decomposition (HH1) and level-two DWTdecomposition (HH2) with and without filtering by a Gaborfilter bank We also applied the Gabor filter directly to theoriginal image without the DWT for comparison purposeFor eachDWTHH subband image the feature vector is givenby

119881DWT119886 = [119864119886 119880119886] (9)

where 119886 is the level of wavelet analysis (decomposition)Similarly for each of the twelve outputs generated by theGabor filter bank (4 angles times 3 central frequencies) theentropy and uniformity are computed and a twenty-fourcomponent feature vector is formed to represent the initialimage We thus have

119881Gabor119886 = [1198641119886 1198642119886 11986412119886 1198801119886 1198802119886 11988012119886] (10)

Either feature vector is subsequently fed to the SVM toclassify normal versus pathological images

34 The Support Vector Machine Classifier Introduced byVapnik [36] the support vector machine (SVM) classifieris based on statistical learning theory It implements theprinciple of structural risk minimization and has excellentgeneralization ability as a result even when the data sample issmall Moreover SVM can tolerate high-dimensional andorincomplete data [50] It has been used with great success invarious applications including speech emotion recognition[51] card-sharing traffic detection [52] fault diagnosis [53]cardiac decision making [54] Parkinsonrsquos disease diagnosis[55] and Alzheimerrsquos disease detection [56]

The support vector machine performs classification tasksby constructing an optimal separating hyperplane that maxi-mizes themargin between the two nearest data points belong-ing to two separate classes Given a training set (119909119894 119910119894) 119894 =1 2 119898 where the input 119909119894 isin 119877

119889 and class labels 119910119894 isin+1 minus1 the separation hyperplane for a linearly separablebinary classification problem is given by

119891 (119909) = ⟨119908 sdot 119909⟩ + 119887 (11)

where 119908 is a weight vector and 119887 is a bias The optimalseparation hyperplane is found by solving the followingoptimization problem

minimize119908119887120585

1

2⟨119908 sdot 119908⟩ + 119862

119898

sum

119894=1

120585119894

Subject to 119910119894 (⟨119908 sdot 119909119894⟩ + 119887) + 120585119894 minus 1 ge 0 120585119894 ge 0

(12)

where 119862 is a penalty parameter that controls the tradeoffbetween the complexity of the decision function and thenumber of misclassified training examples and 120585 is a positiveslack variableTheprevious optimizationmodel can be solvedby introducing Lagrange multipliers and using the Karush-Kuhn-Tucker theorem of optimization to obtain the solutionas

119908 =

119898

sum

119894=1

120572119894119910119894119909119894 (13)

The 119909119894 values corresponding to positive Lagrange multipliers120572119894 are called support vectors which define the decisionboundary The 119909119894 values corresponding to zero 120572119894 are irrel-evant Once the optimal solution 120572

lowast

119894is found the optimal

hyperplane parameters 119908lowast and 119887lowast are determined Then thediscriminant function of the SVM for a linearly separablebinary classification problem is [32]

119892 (119909) = sign(119898

sum

119894=1

119910119894120572lowast

119894⟨119909119894 sdot 119909⟩ + 119887

lowast) (14)

In the nonlinearly separable case the SVM classifier nonlin-early maps the training points to a high-dimensional featurespace using a kernel function Φ where linear separation canbe possible The scalar product ⟨Φ(119909119894) sdot Φ(119909119895)⟩ is computedby Mercer kernel function 119870 as 119870(119909119894 119909119895) = ⟨Φ(119909119894) sdot Φ(119909119895)⟩Then the nonlinear SVM classifier has the following form

119892 (119909) = sign(119898

sum

119894=1

119910119894120572lowast

119894119870⟨119909 119909119894⟩ + 119887

lowast) (15)

In this study a polynomial kernel of degree 2 was used forthe SVM As a global kernel it allows data points that are faraway from each other to also have an influence on the kernelvalues The general polynomial kernel is given by

119870(119909 119909119894) = ((119909119894 sdot 119909) + 1)119889 (16)

where 119889 is the degree of the polynomial to be used

4 Experimental Results

Asmentioned previouslymammograms and retina and brainMR images corresponding to given pathologies are consid-ered in this work and the aim is to classify normal versusabnormal images for each image category To do so onehundred digitalmammograms (171times 364 pixels) consisting offifty normal images and fifty cancer images were taken fromThe Digital Database for Screening Mammography (DDSM)

8 Journal of Medical Engineering

Table 1 Average SVM classification accuracy as a function of feature extraction method and level of DWT decompositionlowast

DWT DWT-Gabor DWT DWT-GaborDecomposition level One One Two TwoMammograms 9598 (plusmn004) 9667 (plusmn005) 8913 (plusmn001) 9109 (plusmn005)Retina 7469 (plusmn005) 100 9098 (plusmn003) 100Brain MRI 8780 (plusmn000) 9736 (plusmn002) 8576 (plusmn000) 9118 (plusmn004)lowastTenfold cross-validation used for mammograms and retina images leave-one-out used for brain MRIs

Table 2 SVM classification specificity as a function of feature extraction method and level of DWT decompositionlowast

DWT DWT-Gabor DWT DWT-GaborDecomposition level One One Two TwoMammograms 9781 (plusmn002) 100 8885 (plusmn005) 9209 (plusmn006)Retina 681 (plusmn004) 100 232 (plusmn010) 100Brain MRI 2155 (plusmn001) 9958 (plusmn001) 0 9724 (plusmn001)lowastTenfold cross-validation used for mammograms and retina images leave-one-out used for brain MRIs

[57] For retina a set of 69 color images (150 times 130 pixels)from the STARE [58] database were employed including23 normal images 24 with microaneurysms and 22 withcircinate Finally a collection of 56 axial T2-weighted andMR brain images (256 times 256 pixels) were taken from theAANLIB database [59] of the Harvard Medical School Theyconsisted of 7 normal images 9 with Alzheimerrsquos disease 13with glioma 8 with Herpes encephalitis 8 with metastaticbronchogenic carcinoma and 14 with multiple sclerosis Itis unfortunate that the number of images was not constantacross pathologies but we had no control over this and usedwhat was available with tenfold cross-validation or leave-one-out cross-validation of the results depending on samplesize All experiments were based on a binary classificationapproach of normal versus abnormal images Many kinds ofbiomedical images could be considered for our experimentswe focused on mammograms retina and brain magneticresonance images mainly because of public availability Anexample of the processing of a normal retina and a retina withcircinate is illustrated in Figures 7 and 8 respectively

For each image type the average and standard deviationof the correct classification rate (CCR) sensitivity and speci-ficity were computed to evaluate the performance featureextraction techniquewhen used in conjunctionwith the SVMclassifier The three performance measures are defined by

CCR =Classified Samples

Total Number of Samples

Sensitivity =Correctly Classified Positive Samples

True Positive Samples

Specificity =Correctly Classified Negative Samples

True Negative Samples

(17)

where positive samples and negative samples are respectivelyabnormal and normal images

Finally all experiments were performed with tenfoldcross-validation except those for MR images which usedleave-one-out cross-validation due to the small sample sizeof each brain image category

Table 1 shows the obtained average results for the threetypes of images that were investigated The performance ofthe SVM classifier improved for all types of images and alllevels of HH decomposition by the DWT At level one theaverage correct classification rate increased by respectively069 2531 and 956 percentage points for mammogramsretina and brain magnetic resonance images when usingthe DWT-Gabor approach At level-two decomposition theimprovement was respectively 196 902 and 542 percent-age points

Tables 2 and 3 provide the average results for classifiersensitivity and specificity At level-one DWT decompositionthe DWT-Gabor approach improved classification specificityfor mammograms and retina images to make it reach 100while it improved it by 7803 percentage points for brain MRimages At level-two DWT decomposition the improvementwas 324 percentage points for mammograms 9768 percent-age points for retina images (100 specificity) and 9724percentage points for MR images Regarding sensitivity theresults were mixed

At level-one DWT decomposition the values were aboutthe same for the DWT-Gabor and DWT-only approachesfor mammograms and retina images with respectivelyminus081 and 0 percentage points differences but there was adegradation of minus1133 and minus4616 percentage points for brainMR images at level one and level two of decompositionrespectively

Following the same cross-validation protocol we alsoconducted classification experiments with features extractedfrom aGabor filtered image of the original biomedical imageThe purpose was to check whether Gabor-based featuresalone help characterize images better than DWT or DWT-Gabor-based features The results are given in Table 4

Journal of Medical Engineering 9

20 40 60 80 100 120 140

20

40

60

80

100

120020406080100120140160180200

(a) Original normal retina image

10 20 30 40 50 60 70

10

20

30

40

50

60 minus60

minus40

minus20

0

20

40

60

(b) Normal retina HH1 subimage

0

5

10

15

20

25

30

35

20406080100

minus100

minus80

minus60

minus40

minus20

5 10 15 20 25 30 35 40(c) Normal retina HH2 subimage

10 20 30 40 50 60 70

10

20

30

40

50

60 10

20

30

40

50

60

70

(d) Gabor filtered HH1 at 119891 = 2 and 120579 = 1205874

5

10

15

20

25

30

355 10 15 20 25 30 35 40

102030405060708090100

(e) Gabor filtered HH2 119891 = 2 and 120579 = 1205874

Figure 7 Analysis of a normal retina

20 40 60 80 100 120 140

20

40

60

80

100

120 20406080100120140160180200220

(a) Original retina image with circinate

10 20 30 40 50 60 70

10

20

30

40

50

60minus20

minus10

0

10

20

30

(b) HH1 subimage

5

10

15

20

25

30

355 10 15 20 25 30 35 40

minus80

minus60

minus40

minus20020406080100

(c) HH2 subimage

10 20 30 40 50 60 70

10

20

30

40

50

600

5

10

15

20

25

30

35

(d) Gabor filtered HH1 at 119891 = 2 and 120579 = 1205874

102030405060708090100

5

10

15

20

25

30

355 10 15 20 25 30 35 40

(e) Gabor filtered HH2 at 119891 = 2 and 120579 = 1205874

Figure 8 Analysis of a retina with circinate

10 Journal of Medical Engineering

Table 3 SVM classification sensitivity as a function of feature extraction method and level of DWT decompositionlowast

DWT DWT-Gabor DWT DWT-GaborDecomposition level One One Two TwoMammograms 9414 (plusmn006) 9333 (plusmn006) 9029 (plusmn0039) 8978 (plusmn004)Retina 100 100 100 100Brain MRI 9384 (plusmn000) 8251 (plusmn016) 100 5384 (plusmn023)lowastTenfold cross-validation used for mammograms and retina images leave-one-out used for brain MRIs

Table 4 SVM classification performance measures obtained withGabor-based features

Accuracy Specificity SensitivityMammograms 6803 (plusmn001) 100 0Retina 5000 (plusmn000) 100 0Brain MRI 8661 (plusmn003) 100 0

The obtained correct classification rate of mammogramsretina and brain magnetic resonance images is respectively6803 (plusmn001) 5000 (plusmn000) and 8661 (plusmn003) Theaverage results for classifier specificity and sensitivity for allimages are 100 and 0 This finding indicates that Gabor-based features are suitable to detect pathological imagesbut fails to detect normal images In sum the results showthat Gabor-based features do not perform better than DWTand DWT-Gabor-based features These findings confirm thesuperiority of combining theDWT andGabor filter banks forfeature extraction

Based on the previous results it appears that the DWT-Gabor approach for feature extraction is effective for detect-ing the abrupt changes in biological tissue that characterizethe pathological patterns that were investigated and it yieldsbetter classification accuracy and specificity than the DWT-only approach It also offers equal of better sensitivity exceptfor brain MRIs For brain MRIs the obtained specificityand sensitivity results with the DWT-Gabor approach showimproved true negative detection but lower true positiveperformance Finally the obtained results reveal also thatlevel-one DWT decomposition is preferable to level-twodecomposition

Finally Table 5 compares the results obtained with theDWT-Gabor approach to other work that we surveyed Inmany cases the DWT-Gabor method yields higher clas-sification rates particularly for mammograms and retinaFor the problem of brain MRI classification our obtainedperformance is better than the results of [38] but less thanwhat is reported in [6 8] However these comparisons shouldbe viewed with caution as not all the results stem froma common image database and the different authors usedifferent sample and image sizesMoreovermany authors useno cross-validation and simply performa single arbitrary splitof their data into training and test sets to obtain their accuracyresults Obviously one cannot generalize or draw definiteconclusion from such efforts and comparisons between

works cannot be made other than in general terms In thisrespect it can only be concluded from our results that theDWT-Gabor for feature extraction is effective for obtaininghigh image classification accuracy by an SVM and that itmay outperform other feature extraction and classificationtechniques reported in the literature at least those basedon DWT-only image decomposition Unfortunately a moredefinite conclusion is impossiblewithout gaining access to theimage databases used by the other authors

5 Computational Complexity

Finally the computational complexity of the DWT Gaborcontourlet and curvelet for an 119873 times 119873 image is respectively119874(119873) 119874(1198732 times 119872

2) with 119872 being the width of Gabor

(Gaussian) mask filter119874(1198732) and119874(1198732 log(119873)) As a resultthe computational complexity of the combination of theDWT and Gabor filter is 119874(119873) + 119874(1198732 times 1198722) In terms offeatures extraction processing time the average time requiredto process a brain a mammogram and a retina imagewith the DWT approach (DWT-Gabor) was respectively019 (031) 017 (032) and 015 seconds (035) using MatlabR2009a on a 15 GHz Core2 Duo processor

6 Conclusion

We proposed a supervised system for biomedical imagesclassification that uses statistical features obtained from thecombination of the discrete wavelet transform and Gaborfilter to classify normal images versus cancer images usingsupport vector machines as classifiers Our experimentalresults show that such a hybrid processing model achieveshigher accuracy in comparison to using DWT or Gaborfilter banks alone Therefore the proposed image processingand features extraction approach seem to be very promis-ing for the detection of certain pathologies in biomedicalimages

For future works it is recommended to consider a largerset of features and a selection process to identify the mostdiscriminant ones In addition the Gabor parameters willbe adjusted for each type of image separately to improvethe accuracy Furthermore the DWT-Gabor will be directlycompared to the dual-tree complex wavelet curvelet andcontourlet using the same databases and images in orderto draw general conclusions Also multilabels classificationswill be considered in future works to investigate the discrim-inative power of our approach for each type of pathology

Journal of Medical Engineering 11

Table 5 Comparison with the literature

Features Classifier Accuracylowast

Mammograms[3] Gabor k-NN 80[4] DT-CWT SVM 8864[5] Contourlet SVM 966Our approach DWT-Gabor SVM 9667 (plusmn005)

Retina[1] DWT + GLCM LDA 822

[2] Morphological + GLCM Probabilistic boosting algorithm 813ndash922717ndash852

[39] Gabor SVM 83Our approach DWT-Gabor SVM 100

Brain[6] DWT SVM 98[8] DWT + PCA BPNN 100

SVM 90[38] Voxels Bayes 92

VFI 78Our approach DWT-Gabor SVM 9736 (plusmn002)

lowastCorrect classification rate

Finally more experiments on the effect of kernel choice andits parameter on classification accuracy will be investigated

References

[1] A Khademi and S Krishnan ldquoShift-invariant discrete wavelettransform analysis for retinal image classificationrdquoMedical andBiological Engineering and Computing vol 45 no 12 pp 1211ndash1222 2007

[2] N Lee A F Laine andT R Smith ldquoLearning non-homogenoustextures and the unlearning problemwith application to drusendetection in retinal imagesrdquo in Proceedings of the 5th IEEEInternational Symposium on Biomedical Imaging From Nano toMacro (ISBI rsquo08) pp 1215ndash1218 Paris France May 2008

[3] A Dong and B Wang ldquoFeature selection and analysis onmammogram classificationrdquo in Proceedings of the IEEE PacificRim Conference on Communications Computers and SignalProcessing (PACRIM rsquo09) pp 731ndash735 Victoria BC CanadaAugust 2009

[4] A Tirtajaya and D D Santika ldquoClassification of microcalcifi-cation using dual-tree complex wavelet transform and supportvector machinerdquo in Proceedings of the 2nd International Confer-ence onAdvances in Computing Control and TelecommunicationTechnologies (ACT rsquo10) pp 164ndash166 Jakarta Indonesia Decem-ber 2010

[5] F Moayedi Z Azimifar R Boostani and S Katebi ldquoContour-let-based mammography mass classification using the SVMfamilyrdquo Computers in Biology and Medicine vol 40 no 4 pp373ndash383 2010

[6] S Chaplot LM Patnaik andN R Jagannathan ldquoClassificationof magnetic resonance brain images using wavelets as input tosupport vector machine and neural networkrdquo Biomedical SignalProcessing and Control vol 1 no 1 pp 86ndash92 2006

[7] Y Zhang S Wang and L Wu ldquoA novel method for magneticresonance brain image classification based on adaptive chaoticPSOrdquo Progress in Electromagnetics Research vol 109 pp 325ndash343 2010

[8] Y Zhang Z Dong L Wu and S Wang ldquoA hybrid method forMRI brain image classificationrdquo Expert Systems with Applica-tions vol 38 no 8 pp 10049ndash10053 2011

[9] M E Celebi H Iyatomi G Schaefer and W V StoeckerldquoLesion border detection in dermoscopy imagesrdquoComputerizedMedical Imaging and Graphics vol 33 no 2 pp 148ndash153 2009

[10] Q AbbasM E Celebi and I F Garcıa ldquoSkin tumor area extrac-tion using an improved dynamic programming approachrdquo SkinResearch and Technology vol 18 pp 133ndash142 2012

[11] Q Li F Li and K Doi ldquoComputerized detection of lungnodules in thin-section CT images by use of selective enhance-ment filters and an automated rule-based classifierrdquo AcademicRadiology vol 15 no 2 pp 165ndash175 2008

[12] A El-Bazl M Nitzken E Vanbogaertl G Gimelrsquojarb RFalfi and M Abo El-Ghar ldquoA novel shaped-based diagnosticapproach for early diagnosis of lung nodulesrdquo in Proceedingsof the IEEE International Symposium in Biomedical Imaging(ISBI rsquo11) pp 137ndash140 Chicago Ill USA 2011

[13] M T Coimbra and J P S Cunha ldquoMPEG-7 visual descriptorsmdashcontributions for automated feature extraction in capsuleendoscopyrdquo IEEE Transactions on Circuits and Systems for VideoTechnology vol 16 no 5 pp 628ndash636 2006

[14] B Li and M Q H Meng ldquoTexture analysis for ulcer detectionin capsule endoscopy imagesrdquo Image andVision Computing vol27 no 9 pp 1336ndash1342 2009

[15] C K Chui An Introduction to Wavelets Academic Press SanDiego Calif USA 1992

[16] M Vetterli andCHerley ldquoWavelets and filter banks theory anddesignrdquo IEEE Transactions on Signal Processing vol 40 no 9pp 2207ndash2232 1992

12 Journal of Medical Engineering

[17] J G Daugman ldquoUncertainty relation for resolution inspace spatial frequency and orientation optimized by two-dimensional visual cortical filtersrdquo Journal of the Optical Societyof America A vol 2 no 7 pp 1160ndash1169 1985

[18] I W Selesnick R G Baraniuk and N G Kingsbury ldquoThedual-tree complex wavelet transformrdquo IEEE Signal ProcessingMagazine vol 22 no 6 pp 123ndash151 2005

[19] E Candes and D Donoho ldquoRidgelets a key to higher-dimensional intermittencyrdquo Philosophical Transactions of theLondon Royal Society vol 357 pp 2495ndash2509 1999

[20] E J Candes and D L Donoho ldquoContinuous curvelettransformmdashI Resolution of the wavefront setrdquo Applied andComputational Harmonic Analysis vol 19 no 2 pp 162ndash1972005

[21] M N Do and M Vetterli ldquoThe contourlet transform an effi-cient directional multiresolution image representationrdquo IEEETransactions on Image Processing vol 14 no 12 pp 2091ndash21062005

[22] R J Ferrari R M Rangayyan J E L Desautels and A FFrere ldquoAnalysis of asymmetry in mammograms via directionalfiltering with Gabor waveletsrdquo IEEE Transactions on MedicalImaging vol 20 no 9 pp 953ndash964 2001

[23] Z Cui and G Zhang ldquoA novel medical image dynamic fuzzyclassification model based on ridgelet transformrdquo Journal ofSoftware vol 5 no 5 pp 458ndash465 2010

[24] T Geback and P Koumoutsakos ldquoEdge detection inmicroscopyimages using curveletsrdquo BMC Bioinformatics vol 10 article 752009

[25] J Ma andG Plonka ldquoThe curvelet transform a review of recentapplicationsrdquo IEEE Signal ProcessingMagazine vol 27 no 2 pp118ndash133 2010

[26] N Kingsbury ldquoComplex wavelets and shift invariancerdquo inProceedings of the IEEE Seminar on Time-Scale and Time-Frequency Analysis and Applications pp 501ndash510 London UK2000

[27] Y L Qiao C Y Song and C H Zhao ldquoM-band ridgelettransform based texture classificationrdquo Pattern RecognitionLetters vol 31 no 3 pp 244ndash249 2010

[28] F Gomez and E Romero ldquoTexture characterization using acurvelet based descriptorrdquo Lecture Notes in Computer Sciencevol 5856 pp 113ndash120 2009

[29] H Shan and J Ma ldquoCurvelet-based geodesic snakes for imagesegmentation with multiple objectsrdquo Pattern Recognition Let-ters vol 31 no 5 pp 355ndash360 2010

[30] R Eslami and H Radha ldquoNew image transforms using hybridwavelets and directional filter banks analysis and designrdquo inProceedings of the IEEE International Conference on ImageProcessing (ICIP rsquo05) pp 733ndash736 Genova Italy September2005

[31] O O V Villegas and V G C Sanchez ldquoThe wavelet basedcontourlet transform and its application to feature preservingimage codingrdquo Lecture Notes in Computer Science vol 4827 pp590ndash600 2007

[32] S Lahmir and M Boukadoum ldquoClassification of brain MRIusing the LH and HL wavelet transform sub-bandsrdquo in Pro-ceedings of the IEEE International Symposium on Circuits andSystems (ISCAS rsquo11) pp 1025ndash1028 Rio de Janeiro Brazil May2009 2011

[33] S Lahmir and M Boukadoum ldquoBrain MRI classificationusing an ensemble system and LH and HL wavelet Sub-bandsFeaturesrdquo in Proceedings of the IEEE Symposium Series on

Computational Intelligence (SSCI rsquo11) pp 1ndash7 Paris FranceApril2011

[34] S Lahmir and M Boukadoum ldquoHybrid Cosine and RadonTransform-based processing for Digital Mammogram FeatureExtraction and Classification with SVMrdquo in Proceedings of the33rd IEEE Annual International Conference on Engineering inMedecine and Biology Society (EMBS rsquo11) pp 5104ndash5107 BostonMass USA 2011

[35] S Lahmir andM Boukadoum ldquoDWT and RT-Based Approachfor Feature Extraction and classification of Mammograms withSVMrdquo in Proceedings of the IEEE Biomedical Circuits andSystems Conference (BioCAS rsquo11) pp 412ndash415 San Diego CalifUSA November 2011

[36] VN VapnikTheNature of Statistical LearningTheory Springer1995

[37] S Lahmiri and M Boukadoum ldquoHybrid discret wavelet trans-form and Gabor filter banks processing for mammogramfeatures extractionrdquo in Proceedings of the IEEE New Circuits andSystems (NEWCAS rsquo11) pp 53ndash56 Bordeaux France June 2011

[38] L M Bruce and N Shanmugam ldquoUsing neural networkswith wavelet transforms for an automated mammographicmass classifierrdquo in Proceedings of the 22nd Annual InternationalConference of the IEEE Engineering in Medicine and BiologySociety pp 985ndash987 Chicago Ill USA July 2000

[39] S M H Jamarani G Rezai-rad and H Behnam ldquoA novelmethod for breast cancer prognosis using wavelet packet basedneural networkrdquo in Proceedings of the 27th Annual InternationalConference of the Engineering in Medicine and Biology Society(IEEE-EMBS rsquo05) pp 3414ndash3417 Shanghai China September2005

[40] httpmedical-dictionarythefreedictionarycomcircinate+retinopathy

[41] C I O Martins F N S Medeiros R M S Veras F N Bezerraand R M Cesar Jr ldquoEvaluation of retinal vessel segmentationmethods for microaneurysms detectionrdquo in Proceedings of theIEEE International Conference on Image Processing (ICIP rsquo09)pp 3365ndash3368 Cairo Egypt November 2009

[42] C Plant S J Teipel A Oswald et al ldquoAutomated detectionof brain atrophy patterns based on MRI for the prediction ofAlzheimerrsquos diseaserdquo NeuroImage vol 50 no 1 pp 162ndash1742010

[43] J Meier R Bock L G Nyul and G Michelson ldquoEye fun-dus image processing system for automated glaucoma clas-sificationrdquo in Proceedings of the 52nd Internationales Wis-senschaftliches Kolloquium Technische Universitat Ilmenau2007

[44] E Sakka A Prentza I E Lamprinos and D KoutsourisldquoMicrocalcification detection using multiresolution analysisbased on wavelet transformrdquo in Proceedings of the IEEE Inter-national Special Topic Conference on Information Technology inBiomedicine Ioannina Greece October 2006

[45] S G Mallat ldquoTheory for multiresolution signal decompositionthe wavelet representationrdquo IEEE Transactions on Pattern Anal-ysis and Machine Intelligence vol 11 no 7 pp 674ndash693 1989

[46] J Suckling J Parker and D R Dance ldquoThe mammographicimage analysis society digital mammogram databaserdquo in Pro-ceedings of the the 2nd International Workshop on DigitalMammography A G Gale S M Astley D D Dance and AY Cairns Eds pp 375ndash378 Elsevier York UK 1994

[47] H J Chiou C Y Chen T C Liu et al ldquoComputer-aideddiagnosis of peripheral soft tissue masses based on ultrasound

Journal of Medical Engineering 13

imagingrdquo Computerized Medical Imaging and Graphics vol 33no 5 pp 408ndash413 2009

[48] J K Kim J M Park K S Song and H W Park ldquoAdaptivemammographic image enhancement using first derivative andlocal statisticsrdquo IEEE Transactions on Medical Imaging vol 16no 5 pp 495ndash502 1997

[49] H S Sheshadri andAKandaswamy ldquoBreast tissue classificationusing statistical feature extraction of mammogramsrdquo MedicalImaging and Information Sciences vol 23 no 3 pp 105ndash1072006

[50] N Cristianini and J Shawe-Taylor Introduction to SupportVector Machines and Other Kernel-Based Learning MethodsCambridge University Press Cambridge UK 2000

[51] L Chen X Mao Y Xue and L L Cheng ldquoSpeech emotionrecognition features and classification modelsrdquo Digital SignalProcessing vol 22 pp 1154ndash1160 2012

[52] F Palmieri U Fiore A Castiglione and A De Santis ldquoOn thedetection of card-sharing traffic through wavelet analysis andSupport Vector Machinesrdquo Applied Soft Computing vol 13 no1 pp 615ndash627 2013

[53] A Azadeh M Saberi A Kazem V Ebrahimipour A Nour-mohammadzadeh and Z Saberi ldquoA flexible algorithm forfault diagnosis in a centrifugal pump with corrupted data andnoise based on ANN and support vector machine with hyper-parameters optimizationrdquoApplied Soft Computing vol 13 no 3pp 1478ndash1485 2013

[54] R J Martis U R Acharya K M Mandana A K Ray andC Chakraborty ldquoCardiac decision making using higher orderspectrardquo Biomedical Signal Processing and Control vol 8 pp193ndash203 2013

[55] M R Mohammad ldquoChi-square distance kernel of the gaitsfor the diagnosis of Parkinsonrsquos diseaserdquo Biomedical SignalProcessing and Control vol 8 pp 66ndash70 2013

[56] R Vandenberghe N Nelissen E Salmon et al ldquoBinary clas-sification of 18F-flutemetamol PET using machine learningcomparison with visual reads and structuralMRIrdquoNeuroImagevol 64 no 1 pp 517ndash525 2013

[57] httpmarathoncseeusfeduMammographyDatabasehtml[58] httpwwwcesclemsonedusimahooverstare[59] httpwwwmedharvardeduaanlib

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

4 Journal of Medical Engineering

(a) Normal (b) Alzheimerrsquos disease (c) Glioma

(d) Herpes encephalitis (e) Metastatic bronchogenic carcinoma (f) Multiple sclerosis

Figure 1 Examples of brain MR images

(a) Normal (b) Microaneurysms (c) Circinate

Figure 2 Examples of retina images

the histogram of the intensity distribution of the imageFinally support vector machines were employed for the clas-sification task The performance of the classifications usingone feature set only was 73 with the histogram features76with the fast Fourier transform coefficients 80with theGabor textures and 83 with the pixel intensities

Finally in the problem of brain MRI classification theauthors in [6] used the wavelet coefficients as input to asupport vector machine to classify normal and abnormal

Alzheimerrsquos disease images of size 256 times 256 pixels Theclassification accuracy was 98 using SVM with a radialbasis kernel More recently the authors in [42] used voxelsto represent each brain MRI of size 512 times 512 pixels Usingcross-validated tests the obtained correct classification ratesof normal and Alzheimer images were 90 92 and78 respectively when using classification by SVM naıveBayes classifier and voting feature intervals (VFIs) Stillmore recently the authors in [8] employed the DWT to

Journal of Medical Engineering 5

(a) Normal (b) Abnormal

Figure 3 Examples of mammograms

extract features from brain magnetic resonance images ofsize 256 times 256 pixels and then principal component analysiswas used to reduce the dimensions of the features space Theabnormal images included gliomameningiomaAlzheimerrsquosAlzheimerrsquos plus visual agnosia Pickrsquos disease sarcomaand Huntingtonrsquos disease The classification accuracies usingbackpropagation neural network (BPNN) were 100 usinglearning and testing sets of 33 images each

In this work we are interested in how a DWT-Gabor-based approach for feature extraction may provide betterclassification results than those reported in the previousworks particularly those based on the DWT alone The nextsection provides the details of our methodology

3 Methodology

The overall methodology proceeds as follows First theDWT is applied to the biomedical image to obtain its high-frequency image component since it often contains mostof the desired information about the biological tissue [39]Indeed sudden changes in the texture of the image are typicalindicators of the presence of abnormal biological tissueSecond a bank of Gabor filters with different scales andorientations is applied to the high-frequency image to obtainGabor-filtered images along different spatial orientationsThird statistical features are extracted from the Gabor-filtered images Finally the SVM is used to classify theresulting feature vector for final diagnosisThe block diagramof the DWT-Gabor system is shown in Figure 4 Figure 5summarizes the DWT approach in comparison

31 Discrete Wavelet Transform The two-dimensional dis-crete wavelet transform (2D-DWT) [14ndash16] performs asubband coding of an image in terms of spectral spa-tialfrequency components using an iterative and recursiveprocess Figure 6 illustrates the case of two-level decompo-sition The image is first represented by LH HL and HHsubbands that encode the image details in three directionsand an LL subband which provides an approximation of it

Source image DWTExtract DWT

high frequency

Gabor filter

Extract Gabor filtered images

Compute statistics

SVM Classification

Figure 4 Schematic diagram of the DWT-Gabor approach

Mammogram DWT Extract DWT high frequency

Compute statistics

SVMClassification

Figure 5 Schematic diagram of the DWT approach

The obtained detail or approximation images can be decom-posed again to obtain second-level detail and approximationimages and the process can be repeated for finer analysis aseach iteration doubles the image scale

The computation of the 2D-DWT proceeds from that ofthe 1D-DWT the discrete version of the one-dimensionalcontinuous wavelet transform The one-dimensional contin-uous wavelet transform of a signal 119909(119905) is defined by [7 8]

119882120595 (119886 119887) = int

+infin

minusinfin

119909 (119905) 120595lowast

119886119887(119905) 119889119905 (1)

120595119886119887 (119905) =1

radic|119886|120595(

119905 minus 119887

119886) (2)

where 120595119886119887(sdot) stands for a given wavelet function and 119886 and 119887are the scale and translation parameters respectivelyThe 1D-DWT is obtained by sampling 119886 and 119887 so that (1) becomes that

6 Journal of Medical Engineering

Image

LL

LH

HL

HH

darr

darr

darr

darr

darr

darr

119892(119899)

119892(119899)

119892(119899)

ℎ(119899)

ℎ(119899)ℎ(119899)

Figure 6 2D-DWT decomposition of an image

of a sequence In dyadic sampling 119886 and 119887 are respectively apower of 2 and multiples thereof and the sequence elements(wavelet coefficients) are given by

119888119895119896 = 119882120595 (2minus119895 2minus119895119896) (3a)

where 119895 represents the discrete scale factor and 119896 the discretetranslation factor In other words 119886 and 119887 in (1) are replacedby 2119895 and 2119895119896 respectively

The one-dimensional wavelet decomposition is extendedto an image by applying it to the row variable first and thento the column variable of the obtained result [44] At eachstep two subimages are created with half the number ofpixels of the row or column that was processed In the endan 119872 times 119873 image is decomposed into 4 subimages eachwith 1198722 times 1198732 resolution and preserved scale However(1) has only theoretical merit due to the infinite ranges of119886 and 119887 For a practical implementation the fact that (1)is essentially a measure of correlation between a signal andvarious wavelets derived from a mother is exploited and theDWT decomposition is turned into a filtering operation witha sequence of high-pass and low-pass filters [45] Followingthe notation in [7 8] the discrete form of (1) can then bewritten as

119888119886119895119896 [119909 (119905)] = DS [sum119909 (119905) 119892lowast

119895(119905 minus 2

119895119896)]

119888119889119895119896 [119909 (119905)] = DS [sum119909 (119905) ℎlowast

119895(119905 minus 2

119895119896)]

(4)

where coefficients 119888119886119895119896 and 119888119889119895119896 specify approximationand details components provided by the 119892(119899) low-pass andℎ(119899) high-pass impulse responses respectively and the DSoperator performs downsampling by a factor of 2 The one-dimensional wavelet decomposition is extended to two-dimensional objects by using row and column decomposi-tions as shown in Figure 6 In our work the most frequentlyused wavelet (Daubechies-4) [25] is considered to extract theHH image component

32 Gabor Filter The two-dimensional (2D) Gabor filterdecomposes an image into components corresponding todifferent scales and orientations [22] thus capturing visualproperties such as spatial localization orientation selectivityand spatial frequency The 2D Gabor filter consists of acomplex exponential centered at a given frequency andmodulated by a Gaussian envelope Because of the complex

exponential the filter has both real and imaginary parts Thegeneral form of the real part is defined as follows

119866(119909 119910 120590119909 120590119910 119891 120579) = exp[minus12((

1199091015840

120590119909

)

2

+ (1199101015840

120590119910

)

2

)]

times cos (21205871198911199091015840)

(5)

where

1199091015840= 119909 cos (120579) + 119910 sin (120579)

1199101015840= 119910 cos (120579) minus 119909 sin (120579)

(6)

and where 120590119909 and 120590119910 are the standard deviations of theGaussian envelope along the 119909 and 119910 axes The parameters 119891and 120579 are respectively the central frequency and the rotationof the Gabor filter To obtain the Gabor-filtered image119891(119909 119910)of a given image 119894(119909 119910) the 2D convolution operation (lowast) isperformed

119891 (119909 119910) = 119866 (119909 119910 120590119909 120590119910 119891 120579) lowast 119894 (119909 119910) (7)

The selection of parameters 120590119909 120590119910 119891 and 120579 plays animportant role in the filterrsquos operation However no formaltechnique exists for choosing them and experience-guidedintuition trial and error or heuristic searchmust be used Forretina digital images and brain MR images 120590119909 and 120590119910 werearbitrarily set to unity In the case of mammograms 120590119909 and120590119910 were set to the values used in [46] which were determinedempirically Consequently we used 120590119909 = 120591235 where 120591 isthe full width at half-maximumof theGaussian and120590119910 = 8120590119909No values of120590119909 and120590119910 other than the previous oneswere triedsince optimalitywas not the primary concern of thiswork andwe obtained satisfactory results with these values

Four orientations 120579 = 0 1205874 1205872 and 31205874 were used asin [22 33] These values seemed reasonable as a first try sincethey covered both image axis directions and the forward andbackward diagonals Finally the central frequency 119891 was setto 2 25 and 3 Given that the Gabor filter is modulated bythe cosine of 119891 large values of 119891 lead to a compressed cosineand consequently the filter output is more likely to showfast or frequent changes in biological tissue texture This inturn would help verify our hypothesis that abnormal imagesare characterized by sudden and frequent variations in imagetexture In the end the application of the Gabor filter bank tothe HH image component obtained with the 2D-DWT leadsto twelve Gabor-filtered HH images components for eachchoice of 119891 and 120579

33 Feature Extraction Statistical measures are employed toextract features from both the DWTHH subband image andthe real Gabor-filtered HH image components More pre-cisely the entropy (119864) and uniformity (119880) of the coefficientsof each one are computed Entropy and uniformity wereselected as features because previousworks onmammogramshave shown that uniformity is correlated with suspiciousmalignancy [47] and that entropy can successfully charac-terize breast biological tissue [48] In this study the entropy

Journal of Medical Engineering 7

and uniformity statistics are hypothesized to also characterizeretina andbrainMR imageswith similar contrast information(ie abrupt andor frequent variations in texture) Entropy(119864) and uniformity (119880) are defined by [49]

119864 = minussum119901 (119911) times log (119901 (119911))

119880 = sum1199012(119911)

(8)

where 119911 is a random variable that represents a coefficientin the Gabor filtered image and 119901(119911) is its probability ofoccurrence as estimated by its relative frequency

To investigate the performance of the previous approach-es the image features were extracted from HH at bothlevel-one DWT decomposition (HH1) and level-two DWTdecomposition (HH2) with and without filtering by a Gaborfilter bank We also applied the Gabor filter directly to theoriginal image without the DWT for comparison purposeFor eachDWTHH subband image the feature vector is givenby

119881DWT119886 = [119864119886 119880119886] (9)

where 119886 is the level of wavelet analysis (decomposition)Similarly for each of the twelve outputs generated by theGabor filter bank (4 angles times 3 central frequencies) theentropy and uniformity are computed and a twenty-fourcomponent feature vector is formed to represent the initialimage We thus have

119881Gabor119886 = [1198641119886 1198642119886 11986412119886 1198801119886 1198802119886 11988012119886] (10)

Either feature vector is subsequently fed to the SVM toclassify normal versus pathological images

34 The Support Vector Machine Classifier Introduced byVapnik [36] the support vector machine (SVM) classifieris based on statistical learning theory It implements theprinciple of structural risk minimization and has excellentgeneralization ability as a result even when the data sample issmall Moreover SVM can tolerate high-dimensional andorincomplete data [50] It has been used with great success invarious applications including speech emotion recognition[51] card-sharing traffic detection [52] fault diagnosis [53]cardiac decision making [54] Parkinsonrsquos disease diagnosis[55] and Alzheimerrsquos disease detection [56]

The support vector machine performs classification tasksby constructing an optimal separating hyperplane that maxi-mizes themargin between the two nearest data points belong-ing to two separate classes Given a training set (119909119894 119910119894) 119894 =1 2 119898 where the input 119909119894 isin 119877

119889 and class labels 119910119894 isin+1 minus1 the separation hyperplane for a linearly separablebinary classification problem is given by

119891 (119909) = ⟨119908 sdot 119909⟩ + 119887 (11)

where 119908 is a weight vector and 119887 is a bias The optimalseparation hyperplane is found by solving the followingoptimization problem

minimize119908119887120585

1

2⟨119908 sdot 119908⟩ + 119862

119898

sum

119894=1

120585119894

Subject to 119910119894 (⟨119908 sdot 119909119894⟩ + 119887) + 120585119894 minus 1 ge 0 120585119894 ge 0

(12)

where 119862 is a penalty parameter that controls the tradeoffbetween the complexity of the decision function and thenumber of misclassified training examples and 120585 is a positiveslack variableTheprevious optimizationmodel can be solvedby introducing Lagrange multipliers and using the Karush-Kuhn-Tucker theorem of optimization to obtain the solutionas

119908 =

119898

sum

119894=1

120572119894119910119894119909119894 (13)

The 119909119894 values corresponding to positive Lagrange multipliers120572119894 are called support vectors which define the decisionboundary The 119909119894 values corresponding to zero 120572119894 are irrel-evant Once the optimal solution 120572

lowast

119894is found the optimal

hyperplane parameters 119908lowast and 119887lowast are determined Then thediscriminant function of the SVM for a linearly separablebinary classification problem is [32]

119892 (119909) = sign(119898

sum

119894=1

119910119894120572lowast

119894⟨119909119894 sdot 119909⟩ + 119887

lowast) (14)

In the nonlinearly separable case the SVM classifier nonlin-early maps the training points to a high-dimensional featurespace using a kernel function Φ where linear separation canbe possible The scalar product ⟨Φ(119909119894) sdot Φ(119909119895)⟩ is computedby Mercer kernel function 119870 as 119870(119909119894 119909119895) = ⟨Φ(119909119894) sdot Φ(119909119895)⟩Then the nonlinear SVM classifier has the following form

119892 (119909) = sign(119898

sum

119894=1

119910119894120572lowast

119894119870⟨119909 119909119894⟩ + 119887

lowast) (15)

In this study a polynomial kernel of degree 2 was used forthe SVM As a global kernel it allows data points that are faraway from each other to also have an influence on the kernelvalues The general polynomial kernel is given by

119870(119909 119909119894) = ((119909119894 sdot 119909) + 1)119889 (16)

where 119889 is the degree of the polynomial to be used

4 Experimental Results

Asmentioned previouslymammograms and retina and brainMR images corresponding to given pathologies are consid-ered in this work and the aim is to classify normal versusabnormal images for each image category To do so onehundred digitalmammograms (171times 364 pixels) consisting offifty normal images and fifty cancer images were taken fromThe Digital Database for Screening Mammography (DDSM)

8 Journal of Medical Engineering

Table 1 Average SVM classification accuracy as a function of feature extraction method and level of DWT decompositionlowast

DWT DWT-Gabor DWT DWT-GaborDecomposition level One One Two TwoMammograms 9598 (plusmn004) 9667 (plusmn005) 8913 (plusmn001) 9109 (plusmn005)Retina 7469 (plusmn005) 100 9098 (plusmn003) 100Brain MRI 8780 (plusmn000) 9736 (plusmn002) 8576 (plusmn000) 9118 (plusmn004)lowastTenfold cross-validation used for mammograms and retina images leave-one-out used for brain MRIs

Table 2 SVM classification specificity as a function of feature extraction method and level of DWT decompositionlowast

DWT DWT-Gabor DWT DWT-GaborDecomposition level One One Two TwoMammograms 9781 (plusmn002) 100 8885 (plusmn005) 9209 (plusmn006)Retina 681 (plusmn004) 100 232 (plusmn010) 100Brain MRI 2155 (plusmn001) 9958 (plusmn001) 0 9724 (plusmn001)lowastTenfold cross-validation used for mammograms and retina images leave-one-out used for brain MRIs

[57] For retina a set of 69 color images (150 times 130 pixels)from the STARE [58] database were employed including23 normal images 24 with microaneurysms and 22 withcircinate Finally a collection of 56 axial T2-weighted andMR brain images (256 times 256 pixels) were taken from theAANLIB database [59] of the Harvard Medical School Theyconsisted of 7 normal images 9 with Alzheimerrsquos disease 13with glioma 8 with Herpes encephalitis 8 with metastaticbronchogenic carcinoma and 14 with multiple sclerosis Itis unfortunate that the number of images was not constantacross pathologies but we had no control over this and usedwhat was available with tenfold cross-validation or leave-one-out cross-validation of the results depending on samplesize All experiments were based on a binary classificationapproach of normal versus abnormal images Many kinds ofbiomedical images could be considered for our experimentswe focused on mammograms retina and brain magneticresonance images mainly because of public availability Anexample of the processing of a normal retina and a retina withcircinate is illustrated in Figures 7 and 8 respectively

For each image type the average and standard deviationof the correct classification rate (CCR) sensitivity and speci-ficity were computed to evaluate the performance featureextraction techniquewhen used in conjunctionwith the SVMclassifier The three performance measures are defined by

CCR =Classified Samples

Total Number of Samples

Sensitivity =Correctly Classified Positive Samples

True Positive Samples

Specificity =Correctly Classified Negative Samples

True Negative Samples

(17)

where positive samples and negative samples are respectivelyabnormal and normal images

Finally all experiments were performed with tenfoldcross-validation except those for MR images which usedleave-one-out cross-validation due to the small sample sizeof each brain image category

Table 1 shows the obtained average results for the threetypes of images that were investigated The performance ofthe SVM classifier improved for all types of images and alllevels of HH decomposition by the DWT At level one theaverage correct classification rate increased by respectively069 2531 and 956 percentage points for mammogramsretina and brain magnetic resonance images when usingthe DWT-Gabor approach At level-two decomposition theimprovement was respectively 196 902 and 542 percent-age points

Tables 2 and 3 provide the average results for classifiersensitivity and specificity At level-one DWT decompositionthe DWT-Gabor approach improved classification specificityfor mammograms and retina images to make it reach 100while it improved it by 7803 percentage points for brain MRimages At level-two DWT decomposition the improvementwas 324 percentage points for mammograms 9768 percent-age points for retina images (100 specificity) and 9724percentage points for MR images Regarding sensitivity theresults were mixed

At level-one DWT decomposition the values were aboutthe same for the DWT-Gabor and DWT-only approachesfor mammograms and retina images with respectivelyminus081 and 0 percentage points differences but there was adegradation of minus1133 and minus4616 percentage points for brainMR images at level one and level two of decompositionrespectively

Following the same cross-validation protocol we alsoconducted classification experiments with features extractedfrom aGabor filtered image of the original biomedical imageThe purpose was to check whether Gabor-based featuresalone help characterize images better than DWT or DWT-Gabor-based features The results are given in Table 4

Journal of Medical Engineering 9

20 40 60 80 100 120 140

20

40

60

80

100

120020406080100120140160180200

(a) Original normal retina image

10 20 30 40 50 60 70

10

20

30

40

50

60 minus60

minus40

minus20

0

20

40

60

(b) Normal retina HH1 subimage

0

5

10

15

20

25

30

35

20406080100

minus100

minus80

minus60

minus40

minus20

5 10 15 20 25 30 35 40(c) Normal retina HH2 subimage

10 20 30 40 50 60 70

10

20

30

40

50

60 10

20

30

40

50

60

70

(d) Gabor filtered HH1 at 119891 = 2 and 120579 = 1205874

5

10

15

20

25

30

355 10 15 20 25 30 35 40

102030405060708090100

(e) Gabor filtered HH2 119891 = 2 and 120579 = 1205874

Figure 7 Analysis of a normal retina

20 40 60 80 100 120 140

20

40

60

80

100

120 20406080100120140160180200220

(a) Original retina image with circinate

10 20 30 40 50 60 70

10

20

30

40

50

60minus20

minus10

0

10

20

30

(b) HH1 subimage

5

10

15

20

25

30

355 10 15 20 25 30 35 40

minus80

minus60

minus40

minus20020406080100

(c) HH2 subimage

10 20 30 40 50 60 70

10

20

30

40

50

600

5

10

15

20

25

30

35

(d) Gabor filtered HH1 at 119891 = 2 and 120579 = 1205874

102030405060708090100

5

10

15

20

25

30

355 10 15 20 25 30 35 40

(e) Gabor filtered HH2 at 119891 = 2 and 120579 = 1205874

Figure 8 Analysis of a retina with circinate

10 Journal of Medical Engineering

Table 3 SVM classification sensitivity as a function of feature extraction method and level of DWT decompositionlowast

DWT DWT-Gabor DWT DWT-GaborDecomposition level One One Two TwoMammograms 9414 (plusmn006) 9333 (plusmn006) 9029 (plusmn0039) 8978 (plusmn004)Retina 100 100 100 100Brain MRI 9384 (plusmn000) 8251 (plusmn016) 100 5384 (plusmn023)lowastTenfold cross-validation used for mammograms and retina images leave-one-out used for brain MRIs

Table 4 SVM classification performance measures obtained withGabor-based features

Accuracy Specificity SensitivityMammograms 6803 (plusmn001) 100 0Retina 5000 (plusmn000) 100 0Brain MRI 8661 (plusmn003) 100 0

The obtained correct classification rate of mammogramsretina and brain magnetic resonance images is respectively6803 (plusmn001) 5000 (plusmn000) and 8661 (plusmn003) Theaverage results for classifier specificity and sensitivity for allimages are 100 and 0 This finding indicates that Gabor-based features are suitable to detect pathological imagesbut fails to detect normal images In sum the results showthat Gabor-based features do not perform better than DWTand DWT-Gabor-based features These findings confirm thesuperiority of combining theDWT andGabor filter banks forfeature extraction

Based on the previous results it appears that the DWT-Gabor approach for feature extraction is effective for detect-ing the abrupt changes in biological tissue that characterizethe pathological patterns that were investigated and it yieldsbetter classification accuracy and specificity than the DWT-only approach It also offers equal of better sensitivity exceptfor brain MRIs For brain MRIs the obtained specificityand sensitivity results with the DWT-Gabor approach showimproved true negative detection but lower true positiveperformance Finally the obtained results reveal also thatlevel-one DWT decomposition is preferable to level-twodecomposition

Finally Table 5 compares the results obtained with theDWT-Gabor approach to other work that we surveyed Inmany cases the DWT-Gabor method yields higher clas-sification rates particularly for mammograms and retinaFor the problem of brain MRI classification our obtainedperformance is better than the results of [38] but less thanwhat is reported in [6 8] However these comparisons shouldbe viewed with caution as not all the results stem froma common image database and the different authors usedifferent sample and image sizesMoreovermany authors useno cross-validation and simply performa single arbitrary splitof their data into training and test sets to obtain their accuracyresults Obviously one cannot generalize or draw definiteconclusion from such efforts and comparisons between

works cannot be made other than in general terms In thisrespect it can only be concluded from our results that theDWT-Gabor for feature extraction is effective for obtaininghigh image classification accuracy by an SVM and that itmay outperform other feature extraction and classificationtechniques reported in the literature at least those basedon DWT-only image decomposition Unfortunately a moredefinite conclusion is impossiblewithout gaining access to theimage databases used by the other authors

5 Computational Complexity

Finally the computational complexity of the DWT Gaborcontourlet and curvelet for an 119873 times 119873 image is respectively119874(119873) 119874(1198732 times 119872

2) with 119872 being the width of Gabor

(Gaussian) mask filter119874(1198732) and119874(1198732 log(119873)) As a resultthe computational complexity of the combination of theDWT and Gabor filter is 119874(119873) + 119874(1198732 times 1198722) In terms offeatures extraction processing time the average time requiredto process a brain a mammogram and a retina imagewith the DWT approach (DWT-Gabor) was respectively019 (031) 017 (032) and 015 seconds (035) using MatlabR2009a on a 15 GHz Core2 Duo processor

6 Conclusion

We proposed a supervised system for biomedical imagesclassification that uses statistical features obtained from thecombination of the discrete wavelet transform and Gaborfilter to classify normal images versus cancer images usingsupport vector machines as classifiers Our experimentalresults show that such a hybrid processing model achieveshigher accuracy in comparison to using DWT or Gaborfilter banks alone Therefore the proposed image processingand features extraction approach seem to be very promis-ing for the detection of certain pathologies in biomedicalimages

For future works it is recommended to consider a largerset of features and a selection process to identify the mostdiscriminant ones In addition the Gabor parameters willbe adjusted for each type of image separately to improvethe accuracy Furthermore the DWT-Gabor will be directlycompared to the dual-tree complex wavelet curvelet andcontourlet using the same databases and images in orderto draw general conclusions Also multilabels classificationswill be considered in future works to investigate the discrim-inative power of our approach for each type of pathology

Journal of Medical Engineering 11

Table 5 Comparison with the literature

Features Classifier Accuracylowast

Mammograms[3] Gabor k-NN 80[4] DT-CWT SVM 8864[5] Contourlet SVM 966Our approach DWT-Gabor SVM 9667 (plusmn005)

Retina[1] DWT + GLCM LDA 822

[2] Morphological + GLCM Probabilistic boosting algorithm 813ndash922717ndash852

[39] Gabor SVM 83Our approach DWT-Gabor SVM 100

Brain[6] DWT SVM 98[8] DWT + PCA BPNN 100

SVM 90[38] Voxels Bayes 92

VFI 78Our approach DWT-Gabor SVM 9736 (plusmn002)

lowastCorrect classification rate

Finally more experiments on the effect of kernel choice andits parameter on classification accuracy will be investigated

References

[1] A Khademi and S Krishnan ldquoShift-invariant discrete wavelettransform analysis for retinal image classificationrdquoMedical andBiological Engineering and Computing vol 45 no 12 pp 1211ndash1222 2007

[2] N Lee A F Laine andT R Smith ldquoLearning non-homogenoustextures and the unlearning problemwith application to drusendetection in retinal imagesrdquo in Proceedings of the 5th IEEEInternational Symposium on Biomedical Imaging From Nano toMacro (ISBI rsquo08) pp 1215ndash1218 Paris France May 2008

[3] A Dong and B Wang ldquoFeature selection and analysis onmammogram classificationrdquo in Proceedings of the IEEE PacificRim Conference on Communications Computers and SignalProcessing (PACRIM rsquo09) pp 731ndash735 Victoria BC CanadaAugust 2009

[4] A Tirtajaya and D D Santika ldquoClassification of microcalcifi-cation using dual-tree complex wavelet transform and supportvector machinerdquo in Proceedings of the 2nd International Confer-ence onAdvances in Computing Control and TelecommunicationTechnologies (ACT rsquo10) pp 164ndash166 Jakarta Indonesia Decem-ber 2010

[5] F Moayedi Z Azimifar R Boostani and S Katebi ldquoContour-let-based mammography mass classification using the SVMfamilyrdquo Computers in Biology and Medicine vol 40 no 4 pp373ndash383 2010

[6] S Chaplot LM Patnaik andN R Jagannathan ldquoClassificationof magnetic resonance brain images using wavelets as input tosupport vector machine and neural networkrdquo Biomedical SignalProcessing and Control vol 1 no 1 pp 86ndash92 2006

[7] Y Zhang S Wang and L Wu ldquoA novel method for magneticresonance brain image classification based on adaptive chaoticPSOrdquo Progress in Electromagnetics Research vol 109 pp 325ndash343 2010

[8] Y Zhang Z Dong L Wu and S Wang ldquoA hybrid method forMRI brain image classificationrdquo Expert Systems with Applica-tions vol 38 no 8 pp 10049ndash10053 2011

[9] M E Celebi H Iyatomi G Schaefer and W V StoeckerldquoLesion border detection in dermoscopy imagesrdquoComputerizedMedical Imaging and Graphics vol 33 no 2 pp 148ndash153 2009

[10] Q AbbasM E Celebi and I F Garcıa ldquoSkin tumor area extrac-tion using an improved dynamic programming approachrdquo SkinResearch and Technology vol 18 pp 133ndash142 2012

[11] Q Li F Li and K Doi ldquoComputerized detection of lungnodules in thin-section CT images by use of selective enhance-ment filters and an automated rule-based classifierrdquo AcademicRadiology vol 15 no 2 pp 165ndash175 2008

[12] A El-Bazl M Nitzken E Vanbogaertl G Gimelrsquojarb RFalfi and M Abo El-Ghar ldquoA novel shaped-based diagnosticapproach for early diagnosis of lung nodulesrdquo in Proceedingsof the IEEE International Symposium in Biomedical Imaging(ISBI rsquo11) pp 137ndash140 Chicago Ill USA 2011

[13] M T Coimbra and J P S Cunha ldquoMPEG-7 visual descriptorsmdashcontributions for automated feature extraction in capsuleendoscopyrdquo IEEE Transactions on Circuits and Systems for VideoTechnology vol 16 no 5 pp 628ndash636 2006

[14] B Li and M Q H Meng ldquoTexture analysis for ulcer detectionin capsule endoscopy imagesrdquo Image andVision Computing vol27 no 9 pp 1336ndash1342 2009

[15] C K Chui An Introduction to Wavelets Academic Press SanDiego Calif USA 1992

[16] M Vetterli andCHerley ldquoWavelets and filter banks theory anddesignrdquo IEEE Transactions on Signal Processing vol 40 no 9pp 2207ndash2232 1992

12 Journal of Medical Engineering

[17] J G Daugman ldquoUncertainty relation for resolution inspace spatial frequency and orientation optimized by two-dimensional visual cortical filtersrdquo Journal of the Optical Societyof America A vol 2 no 7 pp 1160ndash1169 1985

[18] I W Selesnick R G Baraniuk and N G Kingsbury ldquoThedual-tree complex wavelet transformrdquo IEEE Signal ProcessingMagazine vol 22 no 6 pp 123ndash151 2005

[19] E Candes and D Donoho ldquoRidgelets a key to higher-dimensional intermittencyrdquo Philosophical Transactions of theLondon Royal Society vol 357 pp 2495ndash2509 1999

[20] E J Candes and D L Donoho ldquoContinuous curvelettransformmdashI Resolution of the wavefront setrdquo Applied andComputational Harmonic Analysis vol 19 no 2 pp 162ndash1972005

[21] M N Do and M Vetterli ldquoThe contourlet transform an effi-cient directional multiresolution image representationrdquo IEEETransactions on Image Processing vol 14 no 12 pp 2091ndash21062005

[22] R J Ferrari R M Rangayyan J E L Desautels and A FFrere ldquoAnalysis of asymmetry in mammograms via directionalfiltering with Gabor waveletsrdquo IEEE Transactions on MedicalImaging vol 20 no 9 pp 953ndash964 2001

[23] Z Cui and G Zhang ldquoA novel medical image dynamic fuzzyclassification model based on ridgelet transformrdquo Journal ofSoftware vol 5 no 5 pp 458ndash465 2010

[24] T Geback and P Koumoutsakos ldquoEdge detection inmicroscopyimages using curveletsrdquo BMC Bioinformatics vol 10 article 752009

[25] J Ma andG Plonka ldquoThe curvelet transform a review of recentapplicationsrdquo IEEE Signal ProcessingMagazine vol 27 no 2 pp118ndash133 2010

[26] N Kingsbury ldquoComplex wavelets and shift invariancerdquo inProceedings of the IEEE Seminar on Time-Scale and Time-Frequency Analysis and Applications pp 501ndash510 London UK2000

[27] Y L Qiao C Y Song and C H Zhao ldquoM-band ridgelettransform based texture classificationrdquo Pattern RecognitionLetters vol 31 no 3 pp 244ndash249 2010

[28] F Gomez and E Romero ldquoTexture characterization using acurvelet based descriptorrdquo Lecture Notes in Computer Sciencevol 5856 pp 113ndash120 2009

[29] H Shan and J Ma ldquoCurvelet-based geodesic snakes for imagesegmentation with multiple objectsrdquo Pattern Recognition Let-ters vol 31 no 5 pp 355ndash360 2010

[30] R Eslami and H Radha ldquoNew image transforms using hybridwavelets and directional filter banks analysis and designrdquo inProceedings of the IEEE International Conference on ImageProcessing (ICIP rsquo05) pp 733ndash736 Genova Italy September2005

[31] O O V Villegas and V G C Sanchez ldquoThe wavelet basedcontourlet transform and its application to feature preservingimage codingrdquo Lecture Notes in Computer Science vol 4827 pp590ndash600 2007

[32] S Lahmir and M Boukadoum ldquoClassification of brain MRIusing the LH and HL wavelet transform sub-bandsrdquo in Pro-ceedings of the IEEE International Symposium on Circuits andSystems (ISCAS rsquo11) pp 1025ndash1028 Rio de Janeiro Brazil May2009 2011

[33] S Lahmir and M Boukadoum ldquoBrain MRI classificationusing an ensemble system and LH and HL wavelet Sub-bandsFeaturesrdquo in Proceedings of the IEEE Symposium Series on

Computational Intelligence (SSCI rsquo11) pp 1ndash7 Paris FranceApril2011

[34] S Lahmir and M Boukadoum ldquoHybrid Cosine and RadonTransform-based processing for Digital Mammogram FeatureExtraction and Classification with SVMrdquo in Proceedings of the33rd IEEE Annual International Conference on Engineering inMedecine and Biology Society (EMBS rsquo11) pp 5104ndash5107 BostonMass USA 2011

[35] S Lahmir andM Boukadoum ldquoDWT and RT-Based Approachfor Feature Extraction and classification of Mammograms withSVMrdquo in Proceedings of the IEEE Biomedical Circuits andSystems Conference (BioCAS rsquo11) pp 412ndash415 San Diego CalifUSA November 2011

[36] VN VapnikTheNature of Statistical LearningTheory Springer1995

[37] S Lahmiri and M Boukadoum ldquoHybrid discret wavelet trans-form and Gabor filter banks processing for mammogramfeatures extractionrdquo in Proceedings of the IEEE New Circuits andSystems (NEWCAS rsquo11) pp 53ndash56 Bordeaux France June 2011

[38] L M Bruce and N Shanmugam ldquoUsing neural networkswith wavelet transforms for an automated mammographicmass classifierrdquo in Proceedings of the 22nd Annual InternationalConference of the IEEE Engineering in Medicine and BiologySociety pp 985ndash987 Chicago Ill USA July 2000

[39] S M H Jamarani G Rezai-rad and H Behnam ldquoA novelmethod for breast cancer prognosis using wavelet packet basedneural networkrdquo in Proceedings of the 27th Annual InternationalConference of the Engineering in Medicine and Biology Society(IEEE-EMBS rsquo05) pp 3414ndash3417 Shanghai China September2005

[40] httpmedical-dictionarythefreedictionarycomcircinate+retinopathy

[41] C I O Martins F N S Medeiros R M S Veras F N Bezerraand R M Cesar Jr ldquoEvaluation of retinal vessel segmentationmethods for microaneurysms detectionrdquo in Proceedings of theIEEE International Conference on Image Processing (ICIP rsquo09)pp 3365ndash3368 Cairo Egypt November 2009

[42] C Plant S J Teipel A Oswald et al ldquoAutomated detectionof brain atrophy patterns based on MRI for the prediction ofAlzheimerrsquos diseaserdquo NeuroImage vol 50 no 1 pp 162ndash1742010

[43] J Meier R Bock L G Nyul and G Michelson ldquoEye fun-dus image processing system for automated glaucoma clas-sificationrdquo in Proceedings of the 52nd Internationales Wis-senschaftliches Kolloquium Technische Universitat Ilmenau2007

[44] E Sakka A Prentza I E Lamprinos and D KoutsourisldquoMicrocalcification detection using multiresolution analysisbased on wavelet transformrdquo in Proceedings of the IEEE Inter-national Special Topic Conference on Information Technology inBiomedicine Ioannina Greece October 2006

[45] S G Mallat ldquoTheory for multiresolution signal decompositionthe wavelet representationrdquo IEEE Transactions on Pattern Anal-ysis and Machine Intelligence vol 11 no 7 pp 674ndash693 1989

[46] J Suckling J Parker and D R Dance ldquoThe mammographicimage analysis society digital mammogram databaserdquo in Pro-ceedings of the the 2nd International Workshop on DigitalMammography A G Gale S M Astley D D Dance and AY Cairns Eds pp 375ndash378 Elsevier York UK 1994

[47] H J Chiou C Y Chen T C Liu et al ldquoComputer-aideddiagnosis of peripheral soft tissue masses based on ultrasound

Journal of Medical Engineering 13

imagingrdquo Computerized Medical Imaging and Graphics vol 33no 5 pp 408ndash413 2009

[48] J K Kim J M Park K S Song and H W Park ldquoAdaptivemammographic image enhancement using first derivative andlocal statisticsrdquo IEEE Transactions on Medical Imaging vol 16no 5 pp 495ndash502 1997

[49] H S Sheshadri andAKandaswamy ldquoBreast tissue classificationusing statistical feature extraction of mammogramsrdquo MedicalImaging and Information Sciences vol 23 no 3 pp 105ndash1072006

[50] N Cristianini and J Shawe-Taylor Introduction to SupportVector Machines and Other Kernel-Based Learning MethodsCambridge University Press Cambridge UK 2000

[51] L Chen X Mao Y Xue and L L Cheng ldquoSpeech emotionrecognition features and classification modelsrdquo Digital SignalProcessing vol 22 pp 1154ndash1160 2012

[52] F Palmieri U Fiore A Castiglione and A De Santis ldquoOn thedetection of card-sharing traffic through wavelet analysis andSupport Vector Machinesrdquo Applied Soft Computing vol 13 no1 pp 615ndash627 2013

[53] A Azadeh M Saberi A Kazem V Ebrahimipour A Nour-mohammadzadeh and Z Saberi ldquoA flexible algorithm forfault diagnosis in a centrifugal pump with corrupted data andnoise based on ANN and support vector machine with hyper-parameters optimizationrdquoApplied Soft Computing vol 13 no 3pp 1478ndash1485 2013

[54] R J Martis U R Acharya K M Mandana A K Ray andC Chakraborty ldquoCardiac decision making using higher orderspectrardquo Biomedical Signal Processing and Control vol 8 pp193ndash203 2013

[55] M R Mohammad ldquoChi-square distance kernel of the gaitsfor the diagnosis of Parkinsonrsquos diseaserdquo Biomedical SignalProcessing and Control vol 8 pp 66ndash70 2013

[56] R Vandenberghe N Nelissen E Salmon et al ldquoBinary clas-sification of 18F-flutemetamol PET using machine learningcomparison with visual reads and structuralMRIrdquoNeuroImagevol 64 no 1 pp 517ndash525 2013

[57] httpmarathoncseeusfeduMammographyDatabasehtml[58] httpwwwcesclemsonedusimahooverstare[59] httpwwwmedharvardeduaanlib

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Journal of Medical Engineering 5

(a) Normal (b) Abnormal

Figure 3 Examples of mammograms

extract features from brain magnetic resonance images ofsize 256 times 256 pixels and then principal component analysiswas used to reduce the dimensions of the features space Theabnormal images included gliomameningiomaAlzheimerrsquosAlzheimerrsquos plus visual agnosia Pickrsquos disease sarcomaand Huntingtonrsquos disease The classification accuracies usingbackpropagation neural network (BPNN) were 100 usinglearning and testing sets of 33 images each

In this work we are interested in how a DWT-Gabor-based approach for feature extraction may provide betterclassification results than those reported in the previousworks particularly those based on the DWT alone The nextsection provides the details of our methodology

3 Methodology

The overall methodology proceeds as follows First theDWT is applied to the biomedical image to obtain its high-frequency image component since it often contains mostof the desired information about the biological tissue [39]Indeed sudden changes in the texture of the image are typicalindicators of the presence of abnormal biological tissueSecond a bank of Gabor filters with different scales andorientations is applied to the high-frequency image to obtainGabor-filtered images along different spatial orientationsThird statistical features are extracted from the Gabor-filtered images Finally the SVM is used to classify theresulting feature vector for final diagnosisThe block diagramof the DWT-Gabor system is shown in Figure 4 Figure 5summarizes the DWT approach in comparison

31 Discrete Wavelet Transform The two-dimensional dis-crete wavelet transform (2D-DWT) [14ndash16] performs asubband coding of an image in terms of spectral spa-tialfrequency components using an iterative and recursiveprocess Figure 6 illustrates the case of two-level decompo-sition The image is first represented by LH HL and HHsubbands that encode the image details in three directionsand an LL subband which provides an approximation of it

Source image DWTExtract DWT

high frequency

Gabor filter

Extract Gabor filtered images

Compute statistics

SVM Classification

Figure 4 Schematic diagram of the DWT-Gabor approach

Mammogram DWT Extract DWT high frequency

Compute statistics

SVMClassification

Figure 5 Schematic diagram of the DWT approach

The obtained detail or approximation images can be decom-posed again to obtain second-level detail and approximationimages and the process can be repeated for finer analysis aseach iteration doubles the image scale

The computation of the 2D-DWT proceeds from that ofthe 1D-DWT the discrete version of the one-dimensionalcontinuous wavelet transform The one-dimensional contin-uous wavelet transform of a signal 119909(119905) is defined by [7 8]

119882120595 (119886 119887) = int

+infin

minusinfin

119909 (119905) 120595lowast

119886119887(119905) 119889119905 (1)

120595119886119887 (119905) =1

radic|119886|120595(

119905 minus 119887

119886) (2)

where 120595119886119887(sdot) stands for a given wavelet function and 119886 and 119887are the scale and translation parameters respectivelyThe 1D-DWT is obtained by sampling 119886 and 119887 so that (1) becomes that

6 Journal of Medical Engineering

Image

LL

LH

HL

HH

darr

darr

darr

darr

darr

darr

119892(119899)

119892(119899)

119892(119899)

ℎ(119899)

ℎ(119899)ℎ(119899)

Figure 6 2D-DWT decomposition of an image

of a sequence In dyadic sampling 119886 and 119887 are respectively apower of 2 and multiples thereof and the sequence elements(wavelet coefficients) are given by

119888119895119896 = 119882120595 (2minus119895 2minus119895119896) (3a)

where 119895 represents the discrete scale factor and 119896 the discretetranslation factor In other words 119886 and 119887 in (1) are replacedby 2119895 and 2119895119896 respectively

The one-dimensional wavelet decomposition is extendedto an image by applying it to the row variable first and thento the column variable of the obtained result [44] At eachstep two subimages are created with half the number ofpixels of the row or column that was processed In the endan 119872 times 119873 image is decomposed into 4 subimages eachwith 1198722 times 1198732 resolution and preserved scale However(1) has only theoretical merit due to the infinite ranges of119886 and 119887 For a practical implementation the fact that (1)is essentially a measure of correlation between a signal andvarious wavelets derived from a mother is exploited and theDWT decomposition is turned into a filtering operation witha sequence of high-pass and low-pass filters [45] Followingthe notation in [7 8] the discrete form of (1) can then bewritten as

119888119886119895119896 [119909 (119905)] = DS [sum119909 (119905) 119892lowast

119895(119905 minus 2

119895119896)]

119888119889119895119896 [119909 (119905)] = DS [sum119909 (119905) ℎlowast

119895(119905 minus 2

119895119896)]

(4)

where coefficients 119888119886119895119896 and 119888119889119895119896 specify approximationand details components provided by the 119892(119899) low-pass andℎ(119899) high-pass impulse responses respectively and the DSoperator performs downsampling by a factor of 2 The one-dimensional wavelet decomposition is extended to two-dimensional objects by using row and column decomposi-tions as shown in Figure 6 In our work the most frequentlyused wavelet (Daubechies-4) [25] is considered to extract theHH image component

32 Gabor Filter The two-dimensional (2D) Gabor filterdecomposes an image into components corresponding todifferent scales and orientations [22] thus capturing visualproperties such as spatial localization orientation selectivityand spatial frequency The 2D Gabor filter consists of acomplex exponential centered at a given frequency andmodulated by a Gaussian envelope Because of the complex

exponential the filter has both real and imaginary parts Thegeneral form of the real part is defined as follows

119866(119909 119910 120590119909 120590119910 119891 120579) = exp[minus12((

1199091015840

120590119909

)

2

+ (1199101015840

120590119910

)

2

)]

times cos (21205871198911199091015840)

(5)

where

1199091015840= 119909 cos (120579) + 119910 sin (120579)

1199101015840= 119910 cos (120579) minus 119909 sin (120579)

(6)

and where 120590119909 and 120590119910 are the standard deviations of theGaussian envelope along the 119909 and 119910 axes The parameters 119891and 120579 are respectively the central frequency and the rotationof the Gabor filter To obtain the Gabor-filtered image119891(119909 119910)of a given image 119894(119909 119910) the 2D convolution operation (lowast) isperformed

119891 (119909 119910) = 119866 (119909 119910 120590119909 120590119910 119891 120579) lowast 119894 (119909 119910) (7)

The selection of parameters 120590119909 120590119910 119891 and 120579 plays animportant role in the filterrsquos operation However no formaltechnique exists for choosing them and experience-guidedintuition trial and error or heuristic searchmust be used Forretina digital images and brain MR images 120590119909 and 120590119910 werearbitrarily set to unity In the case of mammograms 120590119909 and120590119910 were set to the values used in [46] which were determinedempirically Consequently we used 120590119909 = 120591235 where 120591 isthe full width at half-maximumof theGaussian and120590119910 = 8120590119909No values of120590119909 and120590119910 other than the previous oneswere triedsince optimalitywas not the primary concern of thiswork andwe obtained satisfactory results with these values

Four orientations 120579 = 0 1205874 1205872 and 31205874 were used asin [22 33] These values seemed reasonable as a first try sincethey covered both image axis directions and the forward andbackward diagonals Finally the central frequency 119891 was setto 2 25 and 3 Given that the Gabor filter is modulated bythe cosine of 119891 large values of 119891 lead to a compressed cosineand consequently the filter output is more likely to showfast or frequent changes in biological tissue texture This inturn would help verify our hypothesis that abnormal imagesare characterized by sudden and frequent variations in imagetexture In the end the application of the Gabor filter bank tothe HH image component obtained with the 2D-DWT leadsto twelve Gabor-filtered HH images components for eachchoice of 119891 and 120579

33 Feature Extraction Statistical measures are employed toextract features from both the DWTHH subband image andthe real Gabor-filtered HH image components More pre-cisely the entropy (119864) and uniformity (119880) of the coefficientsof each one are computed Entropy and uniformity wereselected as features because previousworks onmammogramshave shown that uniformity is correlated with suspiciousmalignancy [47] and that entropy can successfully charac-terize breast biological tissue [48] In this study the entropy

Journal of Medical Engineering 7

and uniformity statistics are hypothesized to also characterizeretina andbrainMR imageswith similar contrast information(ie abrupt andor frequent variations in texture) Entropy(119864) and uniformity (119880) are defined by [49]

119864 = minussum119901 (119911) times log (119901 (119911))

119880 = sum1199012(119911)

(8)

where 119911 is a random variable that represents a coefficientin the Gabor filtered image and 119901(119911) is its probability ofoccurrence as estimated by its relative frequency

To investigate the performance of the previous approach-es the image features were extracted from HH at bothlevel-one DWT decomposition (HH1) and level-two DWTdecomposition (HH2) with and without filtering by a Gaborfilter bank We also applied the Gabor filter directly to theoriginal image without the DWT for comparison purposeFor eachDWTHH subband image the feature vector is givenby

119881DWT119886 = [119864119886 119880119886] (9)

where 119886 is the level of wavelet analysis (decomposition)Similarly for each of the twelve outputs generated by theGabor filter bank (4 angles times 3 central frequencies) theentropy and uniformity are computed and a twenty-fourcomponent feature vector is formed to represent the initialimage We thus have

119881Gabor119886 = [1198641119886 1198642119886 11986412119886 1198801119886 1198802119886 11988012119886] (10)

Either feature vector is subsequently fed to the SVM toclassify normal versus pathological images

34 The Support Vector Machine Classifier Introduced byVapnik [36] the support vector machine (SVM) classifieris based on statistical learning theory It implements theprinciple of structural risk minimization and has excellentgeneralization ability as a result even when the data sample issmall Moreover SVM can tolerate high-dimensional andorincomplete data [50] It has been used with great success invarious applications including speech emotion recognition[51] card-sharing traffic detection [52] fault diagnosis [53]cardiac decision making [54] Parkinsonrsquos disease diagnosis[55] and Alzheimerrsquos disease detection [56]

The support vector machine performs classification tasksby constructing an optimal separating hyperplane that maxi-mizes themargin between the two nearest data points belong-ing to two separate classes Given a training set (119909119894 119910119894) 119894 =1 2 119898 where the input 119909119894 isin 119877

119889 and class labels 119910119894 isin+1 minus1 the separation hyperplane for a linearly separablebinary classification problem is given by

119891 (119909) = ⟨119908 sdot 119909⟩ + 119887 (11)

where 119908 is a weight vector and 119887 is a bias The optimalseparation hyperplane is found by solving the followingoptimization problem

minimize119908119887120585

1

2⟨119908 sdot 119908⟩ + 119862

119898

sum

119894=1

120585119894

Subject to 119910119894 (⟨119908 sdot 119909119894⟩ + 119887) + 120585119894 minus 1 ge 0 120585119894 ge 0

(12)

where 119862 is a penalty parameter that controls the tradeoffbetween the complexity of the decision function and thenumber of misclassified training examples and 120585 is a positiveslack variableTheprevious optimizationmodel can be solvedby introducing Lagrange multipliers and using the Karush-Kuhn-Tucker theorem of optimization to obtain the solutionas

119908 =

119898

sum

119894=1

120572119894119910119894119909119894 (13)

The 119909119894 values corresponding to positive Lagrange multipliers120572119894 are called support vectors which define the decisionboundary The 119909119894 values corresponding to zero 120572119894 are irrel-evant Once the optimal solution 120572

lowast

119894is found the optimal

hyperplane parameters 119908lowast and 119887lowast are determined Then thediscriminant function of the SVM for a linearly separablebinary classification problem is [32]

119892 (119909) = sign(119898

sum

119894=1

119910119894120572lowast

119894⟨119909119894 sdot 119909⟩ + 119887

lowast) (14)

In the nonlinearly separable case the SVM classifier nonlin-early maps the training points to a high-dimensional featurespace using a kernel function Φ where linear separation canbe possible The scalar product ⟨Φ(119909119894) sdot Φ(119909119895)⟩ is computedby Mercer kernel function 119870 as 119870(119909119894 119909119895) = ⟨Φ(119909119894) sdot Φ(119909119895)⟩Then the nonlinear SVM classifier has the following form

119892 (119909) = sign(119898

sum

119894=1

119910119894120572lowast

119894119870⟨119909 119909119894⟩ + 119887

lowast) (15)

In this study a polynomial kernel of degree 2 was used forthe SVM As a global kernel it allows data points that are faraway from each other to also have an influence on the kernelvalues The general polynomial kernel is given by

119870(119909 119909119894) = ((119909119894 sdot 119909) + 1)119889 (16)

where 119889 is the degree of the polynomial to be used

4 Experimental Results

Asmentioned previouslymammograms and retina and brainMR images corresponding to given pathologies are consid-ered in this work and the aim is to classify normal versusabnormal images for each image category To do so onehundred digitalmammograms (171times 364 pixels) consisting offifty normal images and fifty cancer images were taken fromThe Digital Database for Screening Mammography (DDSM)

8 Journal of Medical Engineering

Table 1 Average SVM classification accuracy as a function of feature extraction method and level of DWT decompositionlowast

DWT DWT-Gabor DWT DWT-GaborDecomposition level One One Two TwoMammograms 9598 (plusmn004) 9667 (plusmn005) 8913 (plusmn001) 9109 (plusmn005)Retina 7469 (plusmn005) 100 9098 (plusmn003) 100Brain MRI 8780 (plusmn000) 9736 (plusmn002) 8576 (plusmn000) 9118 (plusmn004)lowastTenfold cross-validation used for mammograms and retina images leave-one-out used for brain MRIs

Table 2 SVM classification specificity as a function of feature extraction method and level of DWT decompositionlowast

DWT DWT-Gabor DWT DWT-GaborDecomposition level One One Two TwoMammograms 9781 (plusmn002) 100 8885 (plusmn005) 9209 (plusmn006)Retina 681 (plusmn004) 100 232 (plusmn010) 100Brain MRI 2155 (plusmn001) 9958 (plusmn001) 0 9724 (plusmn001)lowastTenfold cross-validation used for mammograms and retina images leave-one-out used for brain MRIs

[57] For retina a set of 69 color images (150 times 130 pixels)from the STARE [58] database were employed including23 normal images 24 with microaneurysms and 22 withcircinate Finally a collection of 56 axial T2-weighted andMR brain images (256 times 256 pixels) were taken from theAANLIB database [59] of the Harvard Medical School Theyconsisted of 7 normal images 9 with Alzheimerrsquos disease 13with glioma 8 with Herpes encephalitis 8 with metastaticbronchogenic carcinoma and 14 with multiple sclerosis Itis unfortunate that the number of images was not constantacross pathologies but we had no control over this and usedwhat was available with tenfold cross-validation or leave-one-out cross-validation of the results depending on samplesize All experiments were based on a binary classificationapproach of normal versus abnormal images Many kinds ofbiomedical images could be considered for our experimentswe focused on mammograms retina and brain magneticresonance images mainly because of public availability Anexample of the processing of a normal retina and a retina withcircinate is illustrated in Figures 7 and 8 respectively

For each image type the average and standard deviationof the correct classification rate (CCR) sensitivity and speci-ficity were computed to evaluate the performance featureextraction techniquewhen used in conjunctionwith the SVMclassifier The three performance measures are defined by

CCR =Classified Samples

Total Number of Samples

Sensitivity =Correctly Classified Positive Samples

True Positive Samples

Specificity =Correctly Classified Negative Samples

True Negative Samples

(17)

where positive samples and negative samples are respectivelyabnormal and normal images

Finally all experiments were performed with tenfoldcross-validation except those for MR images which usedleave-one-out cross-validation due to the small sample sizeof each brain image category

Table 1 shows the obtained average results for the threetypes of images that were investigated The performance ofthe SVM classifier improved for all types of images and alllevels of HH decomposition by the DWT At level one theaverage correct classification rate increased by respectively069 2531 and 956 percentage points for mammogramsretina and brain magnetic resonance images when usingthe DWT-Gabor approach At level-two decomposition theimprovement was respectively 196 902 and 542 percent-age points

Tables 2 and 3 provide the average results for classifiersensitivity and specificity At level-one DWT decompositionthe DWT-Gabor approach improved classification specificityfor mammograms and retina images to make it reach 100while it improved it by 7803 percentage points for brain MRimages At level-two DWT decomposition the improvementwas 324 percentage points for mammograms 9768 percent-age points for retina images (100 specificity) and 9724percentage points for MR images Regarding sensitivity theresults were mixed

At level-one DWT decomposition the values were aboutthe same for the DWT-Gabor and DWT-only approachesfor mammograms and retina images with respectivelyminus081 and 0 percentage points differences but there was adegradation of minus1133 and minus4616 percentage points for brainMR images at level one and level two of decompositionrespectively

Following the same cross-validation protocol we alsoconducted classification experiments with features extractedfrom aGabor filtered image of the original biomedical imageThe purpose was to check whether Gabor-based featuresalone help characterize images better than DWT or DWT-Gabor-based features The results are given in Table 4

Journal of Medical Engineering 9

20 40 60 80 100 120 140

20

40

60

80

100

120020406080100120140160180200

(a) Original normal retina image

10 20 30 40 50 60 70

10

20

30

40

50

60 minus60

minus40

minus20

0

20

40

60

(b) Normal retina HH1 subimage

0

5

10

15

20

25

30

35

20406080100

minus100

minus80

minus60

minus40

minus20

5 10 15 20 25 30 35 40(c) Normal retina HH2 subimage

10 20 30 40 50 60 70

10

20

30

40

50

60 10

20

30

40

50

60

70

(d) Gabor filtered HH1 at 119891 = 2 and 120579 = 1205874

5

10

15

20

25

30

355 10 15 20 25 30 35 40

102030405060708090100

(e) Gabor filtered HH2 119891 = 2 and 120579 = 1205874

Figure 7 Analysis of a normal retina

20 40 60 80 100 120 140

20

40

60

80

100

120 20406080100120140160180200220

(a) Original retina image with circinate

10 20 30 40 50 60 70

10

20

30

40

50

60minus20

minus10

0

10

20

30

(b) HH1 subimage

5

10

15

20

25

30

355 10 15 20 25 30 35 40

minus80

minus60

minus40

minus20020406080100

(c) HH2 subimage

10 20 30 40 50 60 70

10

20

30

40

50

600

5

10

15

20

25

30

35

(d) Gabor filtered HH1 at 119891 = 2 and 120579 = 1205874

102030405060708090100

5

10

15

20

25

30

355 10 15 20 25 30 35 40

(e) Gabor filtered HH2 at 119891 = 2 and 120579 = 1205874

Figure 8 Analysis of a retina with circinate

10 Journal of Medical Engineering

Table 3 SVM classification sensitivity as a function of feature extraction method and level of DWT decompositionlowast

DWT DWT-Gabor DWT DWT-GaborDecomposition level One One Two TwoMammograms 9414 (plusmn006) 9333 (plusmn006) 9029 (plusmn0039) 8978 (plusmn004)Retina 100 100 100 100Brain MRI 9384 (plusmn000) 8251 (plusmn016) 100 5384 (plusmn023)lowastTenfold cross-validation used for mammograms and retina images leave-one-out used for brain MRIs

Table 4 SVM classification performance measures obtained withGabor-based features

Accuracy Specificity SensitivityMammograms 6803 (plusmn001) 100 0Retina 5000 (plusmn000) 100 0Brain MRI 8661 (plusmn003) 100 0

The obtained correct classification rate of mammogramsretina and brain magnetic resonance images is respectively6803 (plusmn001) 5000 (plusmn000) and 8661 (plusmn003) Theaverage results for classifier specificity and sensitivity for allimages are 100 and 0 This finding indicates that Gabor-based features are suitable to detect pathological imagesbut fails to detect normal images In sum the results showthat Gabor-based features do not perform better than DWTand DWT-Gabor-based features These findings confirm thesuperiority of combining theDWT andGabor filter banks forfeature extraction

Based on the previous results it appears that the DWT-Gabor approach for feature extraction is effective for detect-ing the abrupt changes in biological tissue that characterizethe pathological patterns that were investigated and it yieldsbetter classification accuracy and specificity than the DWT-only approach It also offers equal of better sensitivity exceptfor brain MRIs For brain MRIs the obtained specificityand sensitivity results with the DWT-Gabor approach showimproved true negative detection but lower true positiveperformance Finally the obtained results reveal also thatlevel-one DWT decomposition is preferable to level-twodecomposition

Finally Table 5 compares the results obtained with theDWT-Gabor approach to other work that we surveyed Inmany cases the DWT-Gabor method yields higher clas-sification rates particularly for mammograms and retinaFor the problem of brain MRI classification our obtainedperformance is better than the results of [38] but less thanwhat is reported in [6 8] However these comparisons shouldbe viewed with caution as not all the results stem froma common image database and the different authors usedifferent sample and image sizesMoreovermany authors useno cross-validation and simply performa single arbitrary splitof their data into training and test sets to obtain their accuracyresults Obviously one cannot generalize or draw definiteconclusion from such efforts and comparisons between

works cannot be made other than in general terms In thisrespect it can only be concluded from our results that theDWT-Gabor for feature extraction is effective for obtaininghigh image classification accuracy by an SVM and that itmay outperform other feature extraction and classificationtechniques reported in the literature at least those basedon DWT-only image decomposition Unfortunately a moredefinite conclusion is impossiblewithout gaining access to theimage databases used by the other authors

5 Computational Complexity

Finally the computational complexity of the DWT Gaborcontourlet and curvelet for an 119873 times 119873 image is respectively119874(119873) 119874(1198732 times 119872

2) with 119872 being the width of Gabor

(Gaussian) mask filter119874(1198732) and119874(1198732 log(119873)) As a resultthe computational complexity of the combination of theDWT and Gabor filter is 119874(119873) + 119874(1198732 times 1198722) In terms offeatures extraction processing time the average time requiredto process a brain a mammogram and a retina imagewith the DWT approach (DWT-Gabor) was respectively019 (031) 017 (032) and 015 seconds (035) using MatlabR2009a on a 15 GHz Core2 Duo processor

6 Conclusion

We proposed a supervised system for biomedical imagesclassification that uses statistical features obtained from thecombination of the discrete wavelet transform and Gaborfilter to classify normal images versus cancer images usingsupport vector machines as classifiers Our experimentalresults show that such a hybrid processing model achieveshigher accuracy in comparison to using DWT or Gaborfilter banks alone Therefore the proposed image processingand features extraction approach seem to be very promis-ing for the detection of certain pathologies in biomedicalimages

For future works it is recommended to consider a largerset of features and a selection process to identify the mostdiscriminant ones In addition the Gabor parameters willbe adjusted for each type of image separately to improvethe accuracy Furthermore the DWT-Gabor will be directlycompared to the dual-tree complex wavelet curvelet andcontourlet using the same databases and images in orderto draw general conclusions Also multilabels classificationswill be considered in future works to investigate the discrim-inative power of our approach for each type of pathology

Journal of Medical Engineering 11

Table 5 Comparison with the literature

Features Classifier Accuracylowast

Mammograms[3] Gabor k-NN 80[4] DT-CWT SVM 8864[5] Contourlet SVM 966Our approach DWT-Gabor SVM 9667 (plusmn005)

Retina[1] DWT + GLCM LDA 822

[2] Morphological + GLCM Probabilistic boosting algorithm 813ndash922717ndash852

[39] Gabor SVM 83Our approach DWT-Gabor SVM 100

Brain[6] DWT SVM 98[8] DWT + PCA BPNN 100

SVM 90[38] Voxels Bayes 92

VFI 78Our approach DWT-Gabor SVM 9736 (plusmn002)

lowastCorrect classification rate

Finally more experiments on the effect of kernel choice andits parameter on classification accuracy will be investigated

References

[1] A Khademi and S Krishnan ldquoShift-invariant discrete wavelettransform analysis for retinal image classificationrdquoMedical andBiological Engineering and Computing vol 45 no 12 pp 1211ndash1222 2007

[2] N Lee A F Laine andT R Smith ldquoLearning non-homogenoustextures and the unlearning problemwith application to drusendetection in retinal imagesrdquo in Proceedings of the 5th IEEEInternational Symposium on Biomedical Imaging From Nano toMacro (ISBI rsquo08) pp 1215ndash1218 Paris France May 2008

[3] A Dong and B Wang ldquoFeature selection and analysis onmammogram classificationrdquo in Proceedings of the IEEE PacificRim Conference on Communications Computers and SignalProcessing (PACRIM rsquo09) pp 731ndash735 Victoria BC CanadaAugust 2009

[4] A Tirtajaya and D D Santika ldquoClassification of microcalcifi-cation using dual-tree complex wavelet transform and supportvector machinerdquo in Proceedings of the 2nd International Confer-ence onAdvances in Computing Control and TelecommunicationTechnologies (ACT rsquo10) pp 164ndash166 Jakarta Indonesia Decem-ber 2010

[5] F Moayedi Z Azimifar R Boostani and S Katebi ldquoContour-let-based mammography mass classification using the SVMfamilyrdquo Computers in Biology and Medicine vol 40 no 4 pp373ndash383 2010

[6] S Chaplot LM Patnaik andN R Jagannathan ldquoClassificationof magnetic resonance brain images using wavelets as input tosupport vector machine and neural networkrdquo Biomedical SignalProcessing and Control vol 1 no 1 pp 86ndash92 2006

[7] Y Zhang S Wang and L Wu ldquoA novel method for magneticresonance brain image classification based on adaptive chaoticPSOrdquo Progress in Electromagnetics Research vol 109 pp 325ndash343 2010

[8] Y Zhang Z Dong L Wu and S Wang ldquoA hybrid method forMRI brain image classificationrdquo Expert Systems with Applica-tions vol 38 no 8 pp 10049ndash10053 2011

[9] M E Celebi H Iyatomi G Schaefer and W V StoeckerldquoLesion border detection in dermoscopy imagesrdquoComputerizedMedical Imaging and Graphics vol 33 no 2 pp 148ndash153 2009

[10] Q AbbasM E Celebi and I F Garcıa ldquoSkin tumor area extrac-tion using an improved dynamic programming approachrdquo SkinResearch and Technology vol 18 pp 133ndash142 2012

[11] Q Li F Li and K Doi ldquoComputerized detection of lungnodules in thin-section CT images by use of selective enhance-ment filters and an automated rule-based classifierrdquo AcademicRadiology vol 15 no 2 pp 165ndash175 2008

[12] A El-Bazl M Nitzken E Vanbogaertl G Gimelrsquojarb RFalfi and M Abo El-Ghar ldquoA novel shaped-based diagnosticapproach for early diagnosis of lung nodulesrdquo in Proceedingsof the IEEE International Symposium in Biomedical Imaging(ISBI rsquo11) pp 137ndash140 Chicago Ill USA 2011

[13] M T Coimbra and J P S Cunha ldquoMPEG-7 visual descriptorsmdashcontributions for automated feature extraction in capsuleendoscopyrdquo IEEE Transactions on Circuits and Systems for VideoTechnology vol 16 no 5 pp 628ndash636 2006

[14] B Li and M Q H Meng ldquoTexture analysis for ulcer detectionin capsule endoscopy imagesrdquo Image andVision Computing vol27 no 9 pp 1336ndash1342 2009

[15] C K Chui An Introduction to Wavelets Academic Press SanDiego Calif USA 1992

[16] M Vetterli andCHerley ldquoWavelets and filter banks theory anddesignrdquo IEEE Transactions on Signal Processing vol 40 no 9pp 2207ndash2232 1992

12 Journal of Medical Engineering

[17] J G Daugman ldquoUncertainty relation for resolution inspace spatial frequency and orientation optimized by two-dimensional visual cortical filtersrdquo Journal of the Optical Societyof America A vol 2 no 7 pp 1160ndash1169 1985

[18] I W Selesnick R G Baraniuk and N G Kingsbury ldquoThedual-tree complex wavelet transformrdquo IEEE Signal ProcessingMagazine vol 22 no 6 pp 123ndash151 2005

[19] E Candes and D Donoho ldquoRidgelets a key to higher-dimensional intermittencyrdquo Philosophical Transactions of theLondon Royal Society vol 357 pp 2495ndash2509 1999

[20] E J Candes and D L Donoho ldquoContinuous curvelettransformmdashI Resolution of the wavefront setrdquo Applied andComputational Harmonic Analysis vol 19 no 2 pp 162ndash1972005

[21] M N Do and M Vetterli ldquoThe contourlet transform an effi-cient directional multiresolution image representationrdquo IEEETransactions on Image Processing vol 14 no 12 pp 2091ndash21062005

[22] R J Ferrari R M Rangayyan J E L Desautels and A FFrere ldquoAnalysis of asymmetry in mammograms via directionalfiltering with Gabor waveletsrdquo IEEE Transactions on MedicalImaging vol 20 no 9 pp 953ndash964 2001

[23] Z Cui and G Zhang ldquoA novel medical image dynamic fuzzyclassification model based on ridgelet transformrdquo Journal ofSoftware vol 5 no 5 pp 458ndash465 2010

[24] T Geback and P Koumoutsakos ldquoEdge detection inmicroscopyimages using curveletsrdquo BMC Bioinformatics vol 10 article 752009

[25] J Ma andG Plonka ldquoThe curvelet transform a review of recentapplicationsrdquo IEEE Signal ProcessingMagazine vol 27 no 2 pp118ndash133 2010

[26] N Kingsbury ldquoComplex wavelets and shift invariancerdquo inProceedings of the IEEE Seminar on Time-Scale and Time-Frequency Analysis and Applications pp 501ndash510 London UK2000

[27] Y L Qiao C Y Song and C H Zhao ldquoM-band ridgelettransform based texture classificationrdquo Pattern RecognitionLetters vol 31 no 3 pp 244ndash249 2010

[28] F Gomez and E Romero ldquoTexture characterization using acurvelet based descriptorrdquo Lecture Notes in Computer Sciencevol 5856 pp 113ndash120 2009

[29] H Shan and J Ma ldquoCurvelet-based geodesic snakes for imagesegmentation with multiple objectsrdquo Pattern Recognition Let-ters vol 31 no 5 pp 355ndash360 2010

[30] R Eslami and H Radha ldquoNew image transforms using hybridwavelets and directional filter banks analysis and designrdquo inProceedings of the IEEE International Conference on ImageProcessing (ICIP rsquo05) pp 733ndash736 Genova Italy September2005

[31] O O V Villegas and V G C Sanchez ldquoThe wavelet basedcontourlet transform and its application to feature preservingimage codingrdquo Lecture Notes in Computer Science vol 4827 pp590ndash600 2007

[32] S Lahmir and M Boukadoum ldquoClassification of brain MRIusing the LH and HL wavelet transform sub-bandsrdquo in Pro-ceedings of the IEEE International Symposium on Circuits andSystems (ISCAS rsquo11) pp 1025ndash1028 Rio de Janeiro Brazil May2009 2011

[33] S Lahmir and M Boukadoum ldquoBrain MRI classificationusing an ensemble system and LH and HL wavelet Sub-bandsFeaturesrdquo in Proceedings of the IEEE Symposium Series on

Computational Intelligence (SSCI rsquo11) pp 1ndash7 Paris FranceApril2011

[34] S Lahmir and M Boukadoum ldquoHybrid Cosine and RadonTransform-based processing for Digital Mammogram FeatureExtraction and Classification with SVMrdquo in Proceedings of the33rd IEEE Annual International Conference on Engineering inMedecine and Biology Society (EMBS rsquo11) pp 5104ndash5107 BostonMass USA 2011

[35] S Lahmir andM Boukadoum ldquoDWT and RT-Based Approachfor Feature Extraction and classification of Mammograms withSVMrdquo in Proceedings of the IEEE Biomedical Circuits andSystems Conference (BioCAS rsquo11) pp 412ndash415 San Diego CalifUSA November 2011

[36] VN VapnikTheNature of Statistical LearningTheory Springer1995

[37] S Lahmiri and M Boukadoum ldquoHybrid discret wavelet trans-form and Gabor filter banks processing for mammogramfeatures extractionrdquo in Proceedings of the IEEE New Circuits andSystems (NEWCAS rsquo11) pp 53ndash56 Bordeaux France June 2011

[38] L M Bruce and N Shanmugam ldquoUsing neural networkswith wavelet transforms for an automated mammographicmass classifierrdquo in Proceedings of the 22nd Annual InternationalConference of the IEEE Engineering in Medicine and BiologySociety pp 985ndash987 Chicago Ill USA July 2000

[39] S M H Jamarani G Rezai-rad and H Behnam ldquoA novelmethod for breast cancer prognosis using wavelet packet basedneural networkrdquo in Proceedings of the 27th Annual InternationalConference of the Engineering in Medicine and Biology Society(IEEE-EMBS rsquo05) pp 3414ndash3417 Shanghai China September2005

[40] httpmedical-dictionarythefreedictionarycomcircinate+retinopathy

[41] C I O Martins F N S Medeiros R M S Veras F N Bezerraand R M Cesar Jr ldquoEvaluation of retinal vessel segmentationmethods for microaneurysms detectionrdquo in Proceedings of theIEEE International Conference on Image Processing (ICIP rsquo09)pp 3365ndash3368 Cairo Egypt November 2009

[42] C Plant S J Teipel A Oswald et al ldquoAutomated detectionof brain atrophy patterns based on MRI for the prediction ofAlzheimerrsquos diseaserdquo NeuroImage vol 50 no 1 pp 162ndash1742010

[43] J Meier R Bock L G Nyul and G Michelson ldquoEye fun-dus image processing system for automated glaucoma clas-sificationrdquo in Proceedings of the 52nd Internationales Wis-senschaftliches Kolloquium Technische Universitat Ilmenau2007

[44] E Sakka A Prentza I E Lamprinos and D KoutsourisldquoMicrocalcification detection using multiresolution analysisbased on wavelet transformrdquo in Proceedings of the IEEE Inter-national Special Topic Conference on Information Technology inBiomedicine Ioannina Greece October 2006

[45] S G Mallat ldquoTheory for multiresolution signal decompositionthe wavelet representationrdquo IEEE Transactions on Pattern Anal-ysis and Machine Intelligence vol 11 no 7 pp 674ndash693 1989

[46] J Suckling J Parker and D R Dance ldquoThe mammographicimage analysis society digital mammogram databaserdquo in Pro-ceedings of the the 2nd International Workshop on DigitalMammography A G Gale S M Astley D D Dance and AY Cairns Eds pp 375ndash378 Elsevier York UK 1994

[47] H J Chiou C Y Chen T C Liu et al ldquoComputer-aideddiagnosis of peripheral soft tissue masses based on ultrasound

Journal of Medical Engineering 13

imagingrdquo Computerized Medical Imaging and Graphics vol 33no 5 pp 408ndash413 2009

[48] J K Kim J M Park K S Song and H W Park ldquoAdaptivemammographic image enhancement using first derivative andlocal statisticsrdquo IEEE Transactions on Medical Imaging vol 16no 5 pp 495ndash502 1997

[49] H S Sheshadri andAKandaswamy ldquoBreast tissue classificationusing statistical feature extraction of mammogramsrdquo MedicalImaging and Information Sciences vol 23 no 3 pp 105ndash1072006

[50] N Cristianini and J Shawe-Taylor Introduction to SupportVector Machines and Other Kernel-Based Learning MethodsCambridge University Press Cambridge UK 2000

[51] L Chen X Mao Y Xue and L L Cheng ldquoSpeech emotionrecognition features and classification modelsrdquo Digital SignalProcessing vol 22 pp 1154ndash1160 2012

[52] F Palmieri U Fiore A Castiglione and A De Santis ldquoOn thedetection of card-sharing traffic through wavelet analysis andSupport Vector Machinesrdquo Applied Soft Computing vol 13 no1 pp 615ndash627 2013

[53] A Azadeh M Saberi A Kazem V Ebrahimipour A Nour-mohammadzadeh and Z Saberi ldquoA flexible algorithm forfault diagnosis in a centrifugal pump with corrupted data andnoise based on ANN and support vector machine with hyper-parameters optimizationrdquoApplied Soft Computing vol 13 no 3pp 1478ndash1485 2013

[54] R J Martis U R Acharya K M Mandana A K Ray andC Chakraborty ldquoCardiac decision making using higher orderspectrardquo Biomedical Signal Processing and Control vol 8 pp193ndash203 2013

[55] M R Mohammad ldquoChi-square distance kernel of the gaitsfor the diagnosis of Parkinsonrsquos diseaserdquo Biomedical SignalProcessing and Control vol 8 pp 66ndash70 2013

[56] R Vandenberghe N Nelissen E Salmon et al ldquoBinary clas-sification of 18F-flutemetamol PET using machine learningcomparison with visual reads and structuralMRIrdquoNeuroImagevol 64 no 1 pp 517ndash525 2013

[57] httpmarathoncseeusfeduMammographyDatabasehtml[58] httpwwwcesclemsonedusimahooverstare[59] httpwwwmedharvardeduaanlib

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

6 Journal of Medical Engineering

Image

LL

LH

HL

HH

darr

darr

darr

darr

darr

darr

119892(119899)

119892(119899)

119892(119899)

ℎ(119899)

ℎ(119899)ℎ(119899)

Figure 6 2D-DWT decomposition of an image

of a sequence In dyadic sampling 119886 and 119887 are respectively apower of 2 and multiples thereof and the sequence elements(wavelet coefficients) are given by

119888119895119896 = 119882120595 (2minus119895 2minus119895119896) (3a)

where 119895 represents the discrete scale factor and 119896 the discretetranslation factor In other words 119886 and 119887 in (1) are replacedby 2119895 and 2119895119896 respectively

The one-dimensional wavelet decomposition is extendedto an image by applying it to the row variable first and thento the column variable of the obtained result [44] At eachstep two subimages are created with half the number ofpixels of the row or column that was processed In the endan 119872 times 119873 image is decomposed into 4 subimages eachwith 1198722 times 1198732 resolution and preserved scale However(1) has only theoretical merit due to the infinite ranges of119886 and 119887 For a practical implementation the fact that (1)is essentially a measure of correlation between a signal andvarious wavelets derived from a mother is exploited and theDWT decomposition is turned into a filtering operation witha sequence of high-pass and low-pass filters [45] Followingthe notation in [7 8] the discrete form of (1) can then bewritten as

119888119886119895119896 [119909 (119905)] = DS [sum119909 (119905) 119892lowast

119895(119905 minus 2

119895119896)]

119888119889119895119896 [119909 (119905)] = DS [sum119909 (119905) ℎlowast

119895(119905 minus 2

119895119896)]

(4)

where coefficients 119888119886119895119896 and 119888119889119895119896 specify approximationand details components provided by the 119892(119899) low-pass andℎ(119899) high-pass impulse responses respectively and the DSoperator performs downsampling by a factor of 2 The one-dimensional wavelet decomposition is extended to two-dimensional objects by using row and column decomposi-tions as shown in Figure 6 In our work the most frequentlyused wavelet (Daubechies-4) [25] is considered to extract theHH image component

32 Gabor Filter The two-dimensional (2D) Gabor filterdecomposes an image into components corresponding todifferent scales and orientations [22] thus capturing visualproperties such as spatial localization orientation selectivityand spatial frequency The 2D Gabor filter consists of acomplex exponential centered at a given frequency andmodulated by a Gaussian envelope Because of the complex

exponential the filter has both real and imaginary parts Thegeneral form of the real part is defined as follows

119866(119909 119910 120590119909 120590119910 119891 120579) = exp[minus12((

1199091015840

120590119909

)

2

+ (1199101015840

120590119910

)

2

)]

times cos (21205871198911199091015840)

(5)

where

1199091015840= 119909 cos (120579) + 119910 sin (120579)

1199101015840= 119910 cos (120579) minus 119909 sin (120579)

(6)

and where 120590119909 and 120590119910 are the standard deviations of theGaussian envelope along the 119909 and 119910 axes The parameters 119891and 120579 are respectively the central frequency and the rotationof the Gabor filter To obtain the Gabor-filtered image119891(119909 119910)of a given image 119894(119909 119910) the 2D convolution operation (lowast) isperformed

119891 (119909 119910) = 119866 (119909 119910 120590119909 120590119910 119891 120579) lowast 119894 (119909 119910) (7)

The selection of parameters 120590119909 120590119910 119891 and 120579 plays animportant role in the filterrsquos operation However no formaltechnique exists for choosing them and experience-guidedintuition trial and error or heuristic searchmust be used Forretina digital images and brain MR images 120590119909 and 120590119910 werearbitrarily set to unity In the case of mammograms 120590119909 and120590119910 were set to the values used in [46] which were determinedempirically Consequently we used 120590119909 = 120591235 where 120591 isthe full width at half-maximumof theGaussian and120590119910 = 8120590119909No values of120590119909 and120590119910 other than the previous oneswere triedsince optimalitywas not the primary concern of thiswork andwe obtained satisfactory results with these values

Four orientations 120579 = 0 1205874 1205872 and 31205874 were used asin [22 33] These values seemed reasonable as a first try sincethey covered both image axis directions and the forward andbackward diagonals Finally the central frequency 119891 was setto 2 25 and 3 Given that the Gabor filter is modulated bythe cosine of 119891 large values of 119891 lead to a compressed cosineand consequently the filter output is more likely to showfast or frequent changes in biological tissue texture This inturn would help verify our hypothesis that abnormal imagesare characterized by sudden and frequent variations in imagetexture In the end the application of the Gabor filter bank tothe HH image component obtained with the 2D-DWT leadsto twelve Gabor-filtered HH images components for eachchoice of 119891 and 120579

33 Feature Extraction Statistical measures are employed toextract features from both the DWTHH subband image andthe real Gabor-filtered HH image components More pre-cisely the entropy (119864) and uniformity (119880) of the coefficientsof each one are computed Entropy and uniformity wereselected as features because previousworks onmammogramshave shown that uniformity is correlated with suspiciousmalignancy [47] and that entropy can successfully charac-terize breast biological tissue [48] In this study the entropy

Journal of Medical Engineering 7

and uniformity statistics are hypothesized to also characterizeretina andbrainMR imageswith similar contrast information(ie abrupt andor frequent variations in texture) Entropy(119864) and uniformity (119880) are defined by [49]

119864 = minussum119901 (119911) times log (119901 (119911))

119880 = sum1199012(119911)

(8)

where 119911 is a random variable that represents a coefficientin the Gabor filtered image and 119901(119911) is its probability ofoccurrence as estimated by its relative frequency

To investigate the performance of the previous approach-es the image features were extracted from HH at bothlevel-one DWT decomposition (HH1) and level-two DWTdecomposition (HH2) with and without filtering by a Gaborfilter bank We also applied the Gabor filter directly to theoriginal image without the DWT for comparison purposeFor eachDWTHH subband image the feature vector is givenby

119881DWT119886 = [119864119886 119880119886] (9)

where 119886 is the level of wavelet analysis (decomposition)Similarly for each of the twelve outputs generated by theGabor filter bank (4 angles times 3 central frequencies) theentropy and uniformity are computed and a twenty-fourcomponent feature vector is formed to represent the initialimage We thus have

119881Gabor119886 = [1198641119886 1198642119886 11986412119886 1198801119886 1198802119886 11988012119886] (10)

Either feature vector is subsequently fed to the SVM toclassify normal versus pathological images

34 The Support Vector Machine Classifier Introduced byVapnik [36] the support vector machine (SVM) classifieris based on statistical learning theory It implements theprinciple of structural risk minimization and has excellentgeneralization ability as a result even when the data sample issmall Moreover SVM can tolerate high-dimensional andorincomplete data [50] It has been used with great success invarious applications including speech emotion recognition[51] card-sharing traffic detection [52] fault diagnosis [53]cardiac decision making [54] Parkinsonrsquos disease diagnosis[55] and Alzheimerrsquos disease detection [56]

The support vector machine performs classification tasksby constructing an optimal separating hyperplane that maxi-mizes themargin between the two nearest data points belong-ing to two separate classes Given a training set (119909119894 119910119894) 119894 =1 2 119898 where the input 119909119894 isin 119877

119889 and class labels 119910119894 isin+1 minus1 the separation hyperplane for a linearly separablebinary classification problem is given by

119891 (119909) = ⟨119908 sdot 119909⟩ + 119887 (11)

where 119908 is a weight vector and 119887 is a bias The optimalseparation hyperplane is found by solving the followingoptimization problem

minimize119908119887120585

1

2⟨119908 sdot 119908⟩ + 119862

119898

sum

119894=1

120585119894

Subject to 119910119894 (⟨119908 sdot 119909119894⟩ + 119887) + 120585119894 minus 1 ge 0 120585119894 ge 0

(12)

where 119862 is a penalty parameter that controls the tradeoffbetween the complexity of the decision function and thenumber of misclassified training examples and 120585 is a positiveslack variableTheprevious optimizationmodel can be solvedby introducing Lagrange multipliers and using the Karush-Kuhn-Tucker theorem of optimization to obtain the solutionas

119908 =

119898

sum

119894=1

120572119894119910119894119909119894 (13)

The 119909119894 values corresponding to positive Lagrange multipliers120572119894 are called support vectors which define the decisionboundary The 119909119894 values corresponding to zero 120572119894 are irrel-evant Once the optimal solution 120572

lowast

119894is found the optimal

hyperplane parameters 119908lowast and 119887lowast are determined Then thediscriminant function of the SVM for a linearly separablebinary classification problem is [32]

119892 (119909) = sign(119898

sum

119894=1

119910119894120572lowast

119894⟨119909119894 sdot 119909⟩ + 119887

lowast) (14)

In the nonlinearly separable case the SVM classifier nonlin-early maps the training points to a high-dimensional featurespace using a kernel function Φ where linear separation canbe possible The scalar product ⟨Φ(119909119894) sdot Φ(119909119895)⟩ is computedby Mercer kernel function 119870 as 119870(119909119894 119909119895) = ⟨Φ(119909119894) sdot Φ(119909119895)⟩Then the nonlinear SVM classifier has the following form

119892 (119909) = sign(119898

sum

119894=1

119910119894120572lowast

119894119870⟨119909 119909119894⟩ + 119887

lowast) (15)

In this study a polynomial kernel of degree 2 was used forthe SVM As a global kernel it allows data points that are faraway from each other to also have an influence on the kernelvalues The general polynomial kernel is given by

119870(119909 119909119894) = ((119909119894 sdot 119909) + 1)119889 (16)

where 119889 is the degree of the polynomial to be used

4 Experimental Results

Asmentioned previouslymammograms and retina and brainMR images corresponding to given pathologies are consid-ered in this work and the aim is to classify normal versusabnormal images for each image category To do so onehundred digitalmammograms (171times 364 pixels) consisting offifty normal images and fifty cancer images were taken fromThe Digital Database for Screening Mammography (DDSM)

8 Journal of Medical Engineering

Table 1 Average SVM classification accuracy as a function of feature extraction method and level of DWT decompositionlowast

DWT DWT-Gabor DWT DWT-GaborDecomposition level One One Two TwoMammograms 9598 (plusmn004) 9667 (plusmn005) 8913 (plusmn001) 9109 (plusmn005)Retina 7469 (plusmn005) 100 9098 (plusmn003) 100Brain MRI 8780 (plusmn000) 9736 (plusmn002) 8576 (plusmn000) 9118 (plusmn004)lowastTenfold cross-validation used for mammograms and retina images leave-one-out used for brain MRIs

Table 2 SVM classification specificity as a function of feature extraction method and level of DWT decompositionlowast

DWT DWT-Gabor DWT DWT-GaborDecomposition level One One Two TwoMammograms 9781 (plusmn002) 100 8885 (plusmn005) 9209 (plusmn006)Retina 681 (plusmn004) 100 232 (plusmn010) 100Brain MRI 2155 (plusmn001) 9958 (plusmn001) 0 9724 (plusmn001)lowastTenfold cross-validation used for mammograms and retina images leave-one-out used for brain MRIs

[57] For retina a set of 69 color images (150 times 130 pixels)from the STARE [58] database were employed including23 normal images 24 with microaneurysms and 22 withcircinate Finally a collection of 56 axial T2-weighted andMR brain images (256 times 256 pixels) were taken from theAANLIB database [59] of the Harvard Medical School Theyconsisted of 7 normal images 9 with Alzheimerrsquos disease 13with glioma 8 with Herpes encephalitis 8 with metastaticbronchogenic carcinoma and 14 with multiple sclerosis Itis unfortunate that the number of images was not constantacross pathologies but we had no control over this and usedwhat was available with tenfold cross-validation or leave-one-out cross-validation of the results depending on samplesize All experiments were based on a binary classificationapproach of normal versus abnormal images Many kinds ofbiomedical images could be considered for our experimentswe focused on mammograms retina and brain magneticresonance images mainly because of public availability Anexample of the processing of a normal retina and a retina withcircinate is illustrated in Figures 7 and 8 respectively

For each image type the average and standard deviationof the correct classification rate (CCR) sensitivity and speci-ficity were computed to evaluate the performance featureextraction techniquewhen used in conjunctionwith the SVMclassifier The three performance measures are defined by

CCR =Classified Samples

Total Number of Samples

Sensitivity =Correctly Classified Positive Samples

True Positive Samples

Specificity =Correctly Classified Negative Samples

True Negative Samples

(17)

where positive samples and negative samples are respectivelyabnormal and normal images

Finally all experiments were performed with tenfoldcross-validation except those for MR images which usedleave-one-out cross-validation due to the small sample sizeof each brain image category

Table 1 shows the obtained average results for the threetypes of images that were investigated The performance ofthe SVM classifier improved for all types of images and alllevels of HH decomposition by the DWT At level one theaverage correct classification rate increased by respectively069 2531 and 956 percentage points for mammogramsretina and brain magnetic resonance images when usingthe DWT-Gabor approach At level-two decomposition theimprovement was respectively 196 902 and 542 percent-age points

Tables 2 and 3 provide the average results for classifiersensitivity and specificity At level-one DWT decompositionthe DWT-Gabor approach improved classification specificityfor mammograms and retina images to make it reach 100while it improved it by 7803 percentage points for brain MRimages At level-two DWT decomposition the improvementwas 324 percentage points for mammograms 9768 percent-age points for retina images (100 specificity) and 9724percentage points for MR images Regarding sensitivity theresults were mixed

At level-one DWT decomposition the values were aboutthe same for the DWT-Gabor and DWT-only approachesfor mammograms and retina images with respectivelyminus081 and 0 percentage points differences but there was adegradation of minus1133 and minus4616 percentage points for brainMR images at level one and level two of decompositionrespectively

Following the same cross-validation protocol we alsoconducted classification experiments with features extractedfrom aGabor filtered image of the original biomedical imageThe purpose was to check whether Gabor-based featuresalone help characterize images better than DWT or DWT-Gabor-based features The results are given in Table 4

Journal of Medical Engineering 9

20 40 60 80 100 120 140

20

40

60

80

100

120020406080100120140160180200

(a) Original normal retina image

10 20 30 40 50 60 70

10

20

30

40

50

60 minus60

minus40

minus20

0

20

40

60

(b) Normal retina HH1 subimage

0

5

10

15

20

25

30

35

20406080100

minus100

minus80

minus60

minus40

minus20

5 10 15 20 25 30 35 40(c) Normal retina HH2 subimage

10 20 30 40 50 60 70

10

20

30

40

50

60 10

20

30

40

50

60

70

(d) Gabor filtered HH1 at 119891 = 2 and 120579 = 1205874

5

10

15

20

25

30

355 10 15 20 25 30 35 40

102030405060708090100

(e) Gabor filtered HH2 119891 = 2 and 120579 = 1205874

Figure 7 Analysis of a normal retina

20 40 60 80 100 120 140

20

40

60

80

100

120 20406080100120140160180200220

(a) Original retina image with circinate

10 20 30 40 50 60 70

10

20

30

40

50

60minus20

minus10

0

10

20

30

(b) HH1 subimage

5

10

15

20

25

30

355 10 15 20 25 30 35 40

minus80

minus60

minus40

minus20020406080100

(c) HH2 subimage

10 20 30 40 50 60 70

10

20

30

40

50

600

5

10

15

20

25

30

35

(d) Gabor filtered HH1 at 119891 = 2 and 120579 = 1205874

102030405060708090100

5

10

15

20

25

30

355 10 15 20 25 30 35 40

(e) Gabor filtered HH2 at 119891 = 2 and 120579 = 1205874

Figure 8 Analysis of a retina with circinate

10 Journal of Medical Engineering

Table 3 SVM classification sensitivity as a function of feature extraction method and level of DWT decompositionlowast

DWT DWT-Gabor DWT DWT-GaborDecomposition level One One Two TwoMammograms 9414 (plusmn006) 9333 (plusmn006) 9029 (plusmn0039) 8978 (plusmn004)Retina 100 100 100 100Brain MRI 9384 (plusmn000) 8251 (plusmn016) 100 5384 (plusmn023)lowastTenfold cross-validation used for mammograms and retina images leave-one-out used for brain MRIs

Table 4 SVM classification performance measures obtained withGabor-based features

Accuracy Specificity SensitivityMammograms 6803 (plusmn001) 100 0Retina 5000 (plusmn000) 100 0Brain MRI 8661 (plusmn003) 100 0

The obtained correct classification rate of mammogramsretina and brain magnetic resonance images is respectively6803 (plusmn001) 5000 (plusmn000) and 8661 (plusmn003) Theaverage results for classifier specificity and sensitivity for allimages are 100 and 0 This finding indicates that Gabor-based features are suitable to detect pathological imagesbut fails to detect normal images In sum the results showthat Gabor-based features do not perform better than DWTand DWT-Gabor-based features These findings confirm thesuperiority of combining theDWT andGabor filter banks forfeature extraction

Based on the previous results it appears that the DWT-Gabor approach for feature extraction is effective for detect-ing the abrupt changes in biological tissue that characterizethe pathological patterns that were investigated and it yieldsbetter classification accuracy and specificity than the DWT-only approach It also offers equal of better sensitivity exceptfor brain MRIs For brain MRIs the obtained specificityand sensitivity results with the DWT-Gabor approach showimproved true negative detection but lower true positiveperformance Finally the obtained results reveal also thatlevel-one DWT decomposition is preferable to level-twodecomposition

Finally Table 5 compares the results obtained with theDWT-Gabor approach to other work that we surveyed Inmany cases the DWT-Gabor method yields higher clas-sification rates particularly for mammograms and retinaFor the problem of brain MRI classification our obtainedperformance is better than the results of [38] but less thanwhat is reported in [6 8] However these comparisons shouldbe viewed with caution as not all the results stem froma common image database and the different authors usedifferent sample and image sizesMoreovermany authors useno cross-validation and simply performa single arbitrary splitof their data into training and test sets to obtain their accuracyresults Obviously one cannot generalize or draw definiteconclusion from such efforts and comparisons between

works cannot be made other than in general terms In thisrespect it can only be concluded from our results that theDWT-Gabor for feature extraction is effective for obtaininghigh image classification accuracy by an SVM and that itmay outperform other feature extraction and classificationtechniques reported in the literature at least those basedon DWT-only image decomposition Unfortunately a moredefinite conclusion is impossiblewithout gaining access to theimage databases used by the other authors

5 Computational Complexity

Finally the computational complexity of the DWT Gaborcontourlet and curvelet for an 119873 times 119873 image is respectively119874(119873) 119874(1198732 times 119872

2) with 119872 being the width of Gabor

(Gaussian) mask filter119874(1198732) and119874(1198732 log(119873)) As a resultthe computational complexity of the combination of theDWT and Gabor filter is 119874(119873) + 119874(1198732 times 1198722) In terms offeatures extraction processing time the average time requiredto process a brain a mammogram and a retina imagewith the DWT approach (DWT-Gabor) was respectively019 (031) 017 (032) and 015 seconds (035) using MatlabR2009a on a 15 GHz Core2 Duo processor

6 Conclusion

We proposed a supervised system for biomedical imagesclassification that uses statistical features obtained from thecombination of the discrete wavelet transform and Gaborfilter to classify normal images versus cancer images usingsupport vector machines as classifiers Our experimentalresults show that such a hybrid processing model achieveshigher accuracy in comparison to using DWT or Gaborfilter banks alone Therefore the proposed image processingand features extraction approach seem to be very promis-ing for the detection of certain pathologies in biomedicalimages

For future works it is recommended to consider a largerset of features and a selection process to identify the mostdiscriminant ones In addition the Gabor parameters willbe adjusted for each type of image separately to improvethe accuracy Furthermore the DWT-Gabor will be directlycompared to the dual-tree complex wavelet curvelet andcontourlet using the same databases and images in orderto draw general conclusions Also multilabels classificationswill be considered in future works to investigate the discrim-inative power of our approach for each type of pathology

Journal of Medical Engineering 11

Table 5 Comparison with the literature

Features Classifier Accuracylowast

Mammograms[3] Gabor k-NN 80[4] DT-CWT SVM 8864[5] Contourlet SVM 966Our approach DWT-Gabor SVM 9667 (plusmn005)

Retina[1] DWT + GLCM LDA 822

[2] Morphological + GLCM Probabilistic boosting algorithm 813ndash922717ndash852

[39] Gabor SVM 83Our approach DWT-Gabor SVM 100

Brain[6] DWT SVM 98[8] DWT + PCA BPNN 100

SVM 90[38] Voxels Bayes 92

VFI 78Our approach DWT-Gabor SVM 9736 (plusmn002)

lowastCorrect classification rate

Finally more experiments on the effect of kernel choice andits parameter on classification accuracy will be investigated

References

[1] A Khademi and S Krishnan ldquoShift-invariant discrete wavelettransform analysis for retinal image classificationrdquoMedical andBiological Engineering and Computing vol 45 no 12 pp 1211ndash1222 2007

[2] N Lee A F Laine andT R Smith ldquoLearning non-homogenoustextures and the unlearning problemwith application to drusendetection in retinal imagesrdquo in Proceedings of the 5th IEEEInternational Symposium on Biomedical Imaging From Nano toMacro (ISBI rsquo08) pp 1215ndash1218 Paris France May 2008

[3] A Dong and B Wang ldquoFeature selection and analysis onmammogram classificationrdquo in Proceedings of the IEEE PacificRim Conference on Communications Computers and SignalProcessing (PACRIM rsquo09) pp 731ndash735 Victoria BC CanadaAugust 2009

[4] A Tirtajaya and D D Santika ldquoClassification of microcalcifi-cation using dual-tree complex wavelet transform and supportvector machinerdquo in Proceedings of the 2nd International Confer-ence onAdvances in Computing Control and TelecommunicationTechnologies (ACT rsquo10) pp 164ndash166 Jakarta Indonesia Decem-ber 2010

[5] F Moayedi Z Azimifar R Boostani and S Katebi ldquoContour-let-based mammography mass classification using the SVMfamilyrdquo Computers in Biology and Medicine vol 40 no 4 pp373ndash383 2010

[6] S Chaplot LM Patnaik andN R Jagannathan ldquoClassificationof magnetic resonance brain images using wavelets as input tosupport vector machine and neural networkrdquo Biomedical SignalProcessing and Control vol 1 no 1 pp 86ndash92 2006

[7] Y Zhang S Wang and L Wu ldquoA novel method for magneticresonance brain image classification based on adaptive chaoticPSOrdquo Progress in Electromagnetics Research vol 109 pp 325ndash343 2010

[8] Y Zhang Z Dong L Wu and S Wang ldquoA hybrid method forMRI brain image classificationrdquo Expert Systems with Applica-tions vol 38 no 8 pp 10049ndash10053 2011

[9] M E Celebi H Iyatomi G Schaefer and W V StoeckerldquoLesion border detection in dermoscopy imagesrdquoComputerizedMedical Imaging and Graphics vol 33 no 2 pp 148ndash153 2009

[10] Q AbbasM E Celebi and I F Garcıa ldquoSkin tumor area extrac-tion using an improved dynamic programming approachrdquo SkinResearch and Technology vol 18 pp 133ndash142 2012

[11] Q Li F Li and K Doi ldquoComputerized detection of lungnodules in thin-section CT images by use of selective enhance-ment filters and an automated rule-based classifierrdquo AcademicRadiology vol 15 no 2 pp 165ndash175 2008

[12] A El-Bazl M Nitzken E Vanbogaertl G Gimelrsquojarb RFalfi and M Abo El-Ghar ldquoA novel shaped-based diagnosticapproach for early diagnosis of lung nodulesrdquo in Proceedingsof the IEEE International Symposium in Biomedical Imaging(ISBI rsquo11) pp 137ndash140 Chicago Ill USA 2011

[13] M T Coimbra and J P S Cunha ldquoMPEG-7 visual descriptorsmdashcontributions for automated feature extraction in capsuleendoscopyrdquo IEEE Transactions on Circuits and Systems for VideoTechnology vol 16 no 5 pp 628ndash636 2006

[14] B Li and M Q H Meng ldquoTexture analysis for ulcer detectionin capsule endoscopy imagesrdquo Image andVision Computing vol27 no 9 pp 1336ndash1342 2009

[15] C K Chui An Introduction to Wavelets Academic Press SanDiego Calif USA 1992

[16] M Vetterli andCHerley ldquoWavelets and filter banks theory anddesignrdquo IEEE Transactions on Signal Processing vol 40 no 9pp 2207ndash2232 1992

12 Journal of Medical Engineering

[17] J G Daugman ldquoUncertainty relation for resolution inspace spatial frequency and orientation optimized by two-dimensional visual cortical filtersrdquo Journal of the Optical Societyof America A vol 2 no 7 pp 1160ndash1169 1985

[18] I W Selesnick R G Baraniuk and N G Kingsbury ldquoThedual-tree complex wavelet transformrdquo IEEE Signal ProcessingMagazine vol 22 no 6 pp 123ndash151 2005

[19] E Candes and D Donoho ldquoRidgelets a key to higher-dimensional intermittencyrdquo Philosophical Transactions of theLondon Royal Society vol 357 pp 2495ndash2509 1999

[20] E J Candes and D L Donoho ldquoContinuous curvelettransformmdashI Resolution of the wavefront setrdquo Applied andComputational Harmonic Analysis vol 19 no 2 pp 162ndash1972005

[21] M N Do and M Vetterli ldquoThe contourlet transform an effi-cient directional multiresolution image representationrdquo IEEETransactions on Image Processing vol 14 no 12 pp 2091ndash21062005

[22] R J Ferrari R M Rangayyan J E L Desautels and A FFrere ldquoAnalysis of asymmetry in mammograms via directionalfiltering with Gabor waveletsrdquo IEEE Transactions on MedicalImaging vol 20 no 9 pp 953ndash964 2001

[23] Z Cui and G Zhang ldquoA novel medical image dynamic fuzzyclassification model based on ridgelet transformrdquo Journal ofSoftware vol 5 no 5 pp 458ndash465 2010

[24] T Geback and P Koumoutsakos ldquoEdge detection inmicroscopyimages using curveletsrdquo BMC Bioinformatics vol 10 article 752009

[25] J Ma andG Plonka ldquoThe curvelet transform a review of recentapplicationsrdquo IEEE Signal ProcessingMagazine vol 27 no 2 pp118ndash133 2010

[26] N Kingsbury ldquoComplex wavelets and shift invariancerdquo inProceedings of the IEEE Seminar on Time-Scale and Time-Frequency Analysis and Applications pp 501ndash510 London UK2000

[27] Y L Qiao C Y Song and C H Zhao ldquoM-band ridgelettransform based texture classificationrdquo Pattern RecognitionLetters vol 31 no 3 pp 244ndash249 2010

[28] F Gomez and E Romero ldquoTexture characterization using acurvelet based descriptorrdquo Lecture Notes in Computer Sciencevol 5856 pp 113ndash120 2009

[29] H Shan and J Ma ldquoCurvelet-based geodesic snakes for imagesegmentation with multiple objectsrdquo Pattern Recognition Let-ters vol 31 no 5 pp 355ndash360 2010

[30] R Eslami and H Radha ldquoNew image transforms using hybridwavelets and directional filter banks analysis and designrdquo inProceedings of the IEEE International Conference on ImageProcessing (ICIP rsquo05) pp 733ndash736 Genova Italy September2005

[31] O O V Villegas and V G C Sanchez ldquoThe wavelet basedcontourlet transform and its application to feature preservingimage codingrdquo Lecture Notes in Computer Science vol 4827 pp590ndash600 2007

[32] S Lahmir and M Boukadoum ldquoClassification of brain MRIusing the LH and HL wavelet transform sub-bandsrdquo in Pro-ceedings of the IEEE International Symposium on Circuits andSystems (ISCAS rsquo11) pp 1025ndash1028 Rio de Janeiro Brazil May2009 2011

[33] S Lahmir and M Boukadoum ldquoBrain MRI classificationusing an ensemble system and LH and HL wavelet Sub-bandsFeaturesrdquo in Proceedings of the IEEE Symposium Series on

Computational Intelligence (SSCI rsquo11) pp 1ndash7 Paris FranceApril2011

[34] S Lahmir and M Boukadoum ldquoHybrid Cosine and RadonTransform-based processing for Digital Mammogram FeatureExtraction and Classification with SVMrdquo in Proceedings of the33rd IEEE Annual International Conference on Engineering inMedecine and Biology Society (EMBS rsquo11) pp 5104ndash5107 BostonMass USA 2011

[35] S Lahmir andM Boukadoum ldquoDWT and RT-Based Approachfor Feature Extraction and classification of Mammograms withSVMrdquo in Proceedings of the IEEE Biomedical Circuits andSystems Conference (BioCAS rsquo11) pp 412ndash415 San Diego CalifUSA November 2011

[36] VN VapnikTheNature of Statistical LearningTheory Springer1995

[37] S Lahmiri and M Boukadoum ldquoHybrid discret wavelet trans-form and Gabor filter banks processing for mammogramfeatures extractionrdquo in Proceedings of the IEEE New Circuits andSystems (NEWCAS rsquo11) pp 53ndash56 Bordeaux France June 2011

[38] L M Bruce and N Shanmugam ldquoUsing neural networkswith wavelet transforms for an automated mammographicmass classifierrdquo in Proceedings of the 22nd Annual InternationalConference of the IEEE Engineering in Medicine and BiologySociety pp 985ndash987 Chicago Ill USA July 2000

[39] S M H Jamarani G Rezai-rad and H Behnam ldquoA novelmethod for breast cancer prognosis using wavelet packet basedneural networkrdquo in Proceedings of the 27th Annual InternationalConference of the Engineering in Medicine and Biology Society(IEEE-EMBS rsquo05) pp 3414ndash3417 Shanghai China September2005

[40] httpmedical-dictionarythefreedictionarycomcircinate+retinopathy

[41] C I O Martins F N S Medeiros R M S Veras F N Bezerraand R M Cesar Jr ldquoEvaluation of retinal vessel segmentationmethods for microaneurysms detectionrdquo in Proceedings of theIEEE International Conference on Image Processing (ICIP rsquo09)pp 3365ndash3368 Cairo Egypt November 2009

[42] C Plant S J Teipel A Oswald et al ldquoAutomated detectionof brain atrophy patterns based on MRI for the prediction ofAlzheimerrsquos diseaserdquo NeuroImage vol 50 no 1 pp 162ndash1742010

[43] J Meier R Bock L G Nyul and G Michelson ldquoEye fun-dus image processing system for automated glaucoma clas-sificationrdquo in Proceedings of the 52nd Internationales Wis-senschaftliches Kolloquium Technische Universitat Ilmenau2007

[44] E Sakka A Prentza I E Lamprinos and D KoutsourisldquoMicrocalcification detection using multiresolution analysisbased on wavelet transformrdquo in Proceedings of the IEEE Inter-national Special Topic Conference on Information Technology inBiomedicine Ioannina Greece October 2006

[45] S G Mallat ldquoTheory for multiresolution signal decompositionthe wavelet representationrdquo IEEE Transactions on Pattern Anal-ysis and Machine Intelligence vol 11 no 7 pp 674ndash693 1989

[46] J Suckling J Parker and D R Dance ldquoThe mammographicimage analysis society digital mammogram databaserdquo in Pro-ceedings of the the 2nd International Workshop on DigitalMammography A G Gale S M Astley D D Dance and AY Cairns Eds pp 375ndash378 Elsevier York UK 1994

[47] H J Chiou C Y Chen T C Liu et al ldquoComputer-aideddiagnosis of peripheral soft tissue masses based on ultrasound

Journal of Medical Engineering 13

imagingrdquo Computerized Medical Imaging and Graphics vol 33no 5 pp 408ndash413 2009

[48] J K Kim J M Park K S Song and H W Park ldquoAdaptivemammographic image enhancement using first derivative andlocal statisticsrdquo IEEE Transactions on Medical Imaging vol 16no 5 pp 495ndash502 1997

[49] H S Sheshadri andAKandaswamy ldquoBreast tissue classificationusing statistical feature extraction of mammogramsrdquo MedicalImaging and Information Sciences vol 23 no 3 pp 105ndash1072006

[50] N Cristianini and J Shawe-Taylor Introduction to SupportVector Machines and Other Kernel-Based Learning MethodsCambridge University Press Cambridge UK 2000

[51] L Chen X Mao Y Xue and L L Cheng ldquoSpeech emotionrecognition features and classification modelsrdquo Digital SignalProcessing vol 22 pp 1154ndash1160 2012

[52] F Palmieri U Fiore A Castiglione and A De Santis ldquoOn thedetection of card-sharing traffic through wavelet analysis andSupport Vector Machinesrdquo Applied Soft Computing vol 13 no1 pp 615ndash627 2013

[53] A Azadeh M Saberi A Kazem V Ebrahimipour A Nour-mohammadzadeh and Z Saberi ldquoA flexible algorithm forfault diagnosis in a centrifugal pump with corrupted data andnoise based on ANN and support vector machine with hyper-parameters optimizationrdquoApplied Soft Computing vol 13 no 3pp 1478ndash1485 2013

[54] R J Martis U R Acharya K M Mandana A K Ray andC Chakraborty ldquoCardiac decision making using higher orderspectrardquo Biomedical Signal Processing and Control vol 8 pp193ndash203 2013

[55] M R Mohammad ldquoChi-square distance kernel of the gaitsfor the diagnosis of Parkinsonrsquos diseaserdquo Biomedical SignalProcessing and Control vol 8 pp 66ndash70 2013

[56] R Vandenberghe N Nelissen E Salmon et al ldquoBinary clas-sification of 18F-flutemetamol PET using machine learningcomparison with visual reads and structuralMRIrdquoNeuroImagevol 64 no 1 pp 517ndash525 2013

[57] httpmarathoncseeusfeduMammographyDatabasehtml[58] httpwwwcesclemsonedusimahooverstare[59] httpwwwmedharvardeduaanlib

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Journal of Medical Engineering 7

and uniformity statistics are hypothesized to also characterizeretina andbrainMR imageswith similar contrast information(ie abrupt andor frequent variations in texture) Entropy(119864) and uniformity (119880) are defined by [49]

119864 = minussum119901 (119911) times log (119901 (119911))

119880 = sum1199012(119911)

(8)

where 119911 is a random variable that represents a coefficientin the Gabor filtered image and 119901(119911) is its probability ofoccurrence as estimated by its relative frequency

To investigate the performance of the previous approach-es the image features were extracted from HH at bothlevel-one DWT decomposition (HH1) and level-two DWTdecomposition (HH2) with and without filtering by a Gaborfilter bank We also applied the Gabor filter directly to theoriginal image without the DWT for comparison purposeFor eachDWTHH subband image the feature vector is givenby

119881DWT119886 = [119864119886 119880119886] (9)

where 119886 is the level of wavelet analysis (decomposition)Similarly for each of the twelve outputs generated by theGabor filter bank (4 angles times 3 central frequencies) theentropy and uniformity are computed and a twenty-fourcomponent feature vector is formed to represent the initialimage We thus have

119881Gabor119886 = [1198641119886 1198642119886 11986412119886 1198801119886 1198802119886 11988012119886] (10)

Either feature vector is subsequently fed to the SVM toclassify normal versus pathological images

34 The Support Vector Machine Classifier Introduced byVapnik [36] the support vector machine (SVM) classifieris based on statistical learning theory It implements theprinciple of structural risk minimization and has excellentgeneralization ability as a result even when the data sample issmall Moreover SVM can tolerate high-dimensional andorincomplete data [50] It has been used with great success invarious applications including speech emotion recognition[51] card-sharing traffic detection [52] fault diagnosis [53]cardiac decision making [54] Parkinsonrsquos disease diagnosis[55] and Alzheimerrsquos disease detection [56]

The support vector machine performs classification tasksby constructing an optimal separating hyperplane that maxi-mizes themargin between the two nearest data points belong-ing to two separate classes Given a training set (119909119894 119910119894) 119894 =1 2 119898 where the input 119909119894 isin 119877

119889 and class labels 119910119894 isin+1 minus1 the separation hyperplane for a linearly separablebinary classification problem is given by

119891 (119909) = ⟨119908 sdot 119909⟩ + 119887 (11)

where 119908 is a weight vector and 119887 is a bias The optimalseparation hyperplane is found by solving the followingoptimization problem

minimize119908119887120585

1

2⟨119908 sdot 119908⟩ + 119862

119898

sum

119894=1

120585119894

Subject to 119910119894 (⟨119908 sdot 119909119894⟩ + 119887) + 120585119894 minus 1 ge 0 120585119894 ge 0

(12)

where 119862 is a penalty parameter that controls the tradeoffbetween the complexity of the decision function and thenumber of misclassified training examples and 120585 is a positiveslack variableTheprevious optimizationmodel can be solvedby introducing Lagrange multipliers and using the Karush-Kuhn-Tucker theorem of optimization to obtain the solutionas

119908 =

119898

sum

119894=1

120572119894119910119894119909119894 (13)

The 119909119894 values corresponding to positive Lagrange multipliers120572119894 are called support vectors which define the decisionboundary The 119909119894 values corresponding to zero 120572119894 are irrel-evant Once the optimal solution 120572

lowast

119894is found the optimal

hyperplane parameters 119908lowast and 119887lowast are determined Then thediscriminant function of the SVM for a linearly separablebinary classification problem is [32]

119892 (119909) = sign(119898

sum

119894=1

119910119894120572lowast

119894⟨119909119894 sdot 119909⟩ + 119887

lowast) (14)

In the nonlinearly separable case the SVM classifier nonlin-early maps the training points to a high-dimensional featurespace using a kernel function Φ where linear separation canbe possible The scalar product ⟨Φ(119909119894) sdot Φ(119909119895)⟩ is computedby Mercer kernel function 119870 as 119870(119909119894 119909119895) = ⟨Φ(119909119894) sdot Φ(119909119895)⟩Then the nonlinear SVM classifier has the following form

119892 (119909) = sign(119898

sum

119894=1

119910119894120572lowast

119894119870⟨119909 119909119894⟩ + 119887

lowast) (15)

In this study a polynomial kernel of degree 2 was used forthe SVM As a global kernel it allows data points that are faraway from each other to also have an influence on the kernelvalues The general polynomial kernel is given by

119870(119909 119909119894) = ((119909119894 sdot 119909) + 1)119889 (16)

where 119889 is the degree of the polynomial to be used

4 Experimental Results

Asmentioned previouslymammograms and retina and brainMR images corresponding to given pathologies are consid-ered in this work and the aim is to classify normal versusabnormal images for each image category To do so onehundred digitalmammograms (171times 364 pixels) consisting offifty normal images and fifty cancer images were taken fromThe Digital Database for Screening Mammography (DDSM)

8 Journal of Medical Engineering

Table 1 Average SVM classification accuracy as a function of feature extraction method and level of DWT decompositionlowast

DWT DWT-Gabor DWT DWT-GaborDecomposition level One One Two TwoMammograms 9598 (plusmn004) 9667 (plusmn005) 8913 (plusmn001) 9109 (plusmn005)Retina 7469 (plusmn005) 100 9098 (plusmn003) 100Brain MRI 8780 (plusmn000) 9736 (plusmn002) 8576 (plusmn000) 9118 (plusmn004)lowastTenfold cross-validation used for mammograms and retina images leave-one-out used for brain MRIs

Table 2 SVM classification specificity as a function of feature extraction method and level of DWT decompositionlowast

DWT DWT-Gabor DWT DWT-GaborDecomposition level One One Two TwoMammograms 9781 (plusmn002) 100 8885 (plusmn005) 9209 (plusmn006)Retina 681 (plusmn004) 100 232 (plusmn010) 100Brain MRI 2155 (plusmn001) 9958 (plusmn001) 0 9724 (plusmn001)lowastTenfold cross-validation used for mammograms and retina images leave-one-out used for brain MRIs

[57] For retina a set of 69 color images (150 times 130 pixels)from the STARE [58] database were employed including23 normal images 24 with microaneurysms and 22 withcircinate Finally a collection of 56 axial T2-weighted andMR brain images (256 times 256 pixels) were taken from theAANLIB database [59] of the Harvard Medical School Theyconsisted of 7 normal images 9 with Alzheimerrsquos disease 13with glioma 8 with Herpes encephalitis 8 with metastaticbronchogenic carcinoma and 14 with multiple sclerosis Itis unfortunate that the number of images was not constantacross pathologies but we had no control over this and usedwhat was available with tenfold cross-validation or leave-one-out cross-validation of the results depending on samplesize All experiments were based on a binary classificationapproach of normal versus abnormal images Many kinds ofbiomedical images could be considered for our experimentswe focused on mammograms retina and brain magneticresonance images mainly because of public availability Anexample of the processing of a normal retina and a retina withcircinate is illustrated in Figures 7 and 8 respectively

For each image type the average and standard deviationof the correct classification rate (CCR) sensitivity and speci-ficity were computed to evaluate the performance featureextraction techniquewhen used in conjunctionwith the SVMclassifier The three performance measures are defined by

CCR =Classified Samples

Total Number of Samples

Sensitivity =Correctly Classified Positive Samples

True Positive Samples

Specificity =Correctly Classified Negative Samples

True Negative Samples

(17)

where positive samples and negative samples are respectivelyabnormal and normal images

Finally all experiments were performed with tenfoldcross-validation except those for MR images which usedleave-one-out cross-validation due to the small sample sizeof each brain image category

Table 1 shows the obtained average results for the threetypes of images that were investigated The performance ofthe SVM classifier improved for all types of images and alllevels of HH decomposition by the DWT At level one theaverage correct classification rate increased by respectively069 2531 and 956 percentage points for mammogramsretina and brain magnetic resonance images when usingthe DWT-Gabor approach At level-two decomposition theimprovement was respectively 196 902 and 542 percent-age points

Tables 2 and 3 provide the average results for classifiersensitivity and specificity At level-one DWT decompositionthe DWT-Gabor approach improved classification specificityfor mammograms and retina images to make it reach 100while it improved it by 7803 percentage points for brain MRimages At level-two DWT decomposition the improvementwas 324 percentage points for mammograms 9768 percent-age points for retina images (100 specificity) and 9724percentage points for MR images Regarding sensitivity theresults were mixed

At level-one DWT decomposition the values were aboutthe same for the DWT-Gabor and DWT-only approachesfor mammograms and retina images with respectivelyminus081 and 0 percentage points differences but there was adegradation of minus1133 and minus4616 percentage points for brainMR images at level one and level two of decompositionrespectively

Following the same cross-validation protocol we alsoconducted classification experiments with features extractedfrom aGabor filtered image of the original biomedical imageThe purpose was to check whether Gabor-based featuresalone help characterize images better than DWT or DWT-Gabor-based features The results are given in Table 4

Journal of Medical Engineering 9

20 40 60 80 100 120 140

20

40

60

80

100

120020406080100120140160180200

(a) Original normal retina image

10 20 30 40 50 60 70

10

20

30

40

50

60 minus60

minus40

minus20

0

20

40

60

(b) Normal retina HH1 subimage

0

5

10

15

20

25

30

35

20406080100

minus100

minus80

minus60

minus40

minus20

5 10 15 20 25 30 35 40(c) Normal retina HH2 subimage

10 20 30 40 50 60 70

10

20

30

40

50

60 10

20

30

40

50

60

70

(d) Gabor filtered HH1 at 119891 = 2 and 120579 = 1205874

5

10

15

20

25

30

355 10 15 20 25 30 35 40

102030405060708090100

(e) Gabor filtered HH2 119891 = 2 and 120579 = 1205874

Figure 7 Analysis of a normal retina

20 40 60 80 100 120 140

20

40

60

80

100

120 20406080100120140160180200220

(a) Original retina image with circinate

10 20 30 40 50 60 70

10

20

30

40

50

60minus20

minus10

0

10

20

30

(b) HH1 subimage

5

10

15

20

25

30

355 10 15 20 25 30 35 40

minus80

minus60

minus40

minus20020406080100

(c) HH2 subimage

10 20 30 40 50 60 70

10

20

30

40

50

600

5

10

15

20

25

30

35

(d) Gabor filtered HH1 at 119891 = 2 and 120579 = 1205874

102030405060708090100

5

10

15

20

25

30

355 10 15 20 25 30 35 40

(e) Gabor filtered HH2 at 119891 = 2 and 120579 = 1205874

Figure 8 Analysis of a retina with circinate

10 Journal of Medical Engineering

Table 3 SVM classification sensitivity as a function of feature extraction method and level of DWT decompositionlowast

DWT DWT-Gabor DWT DWT-GaborDecomposition level One One Two TwoMammograms 9414 (plusmn006) 9333 (plusmn006) 9029 (plusmn0039) 8978 (plusmn004)Retina 100 100 100 100Brain MRI 9384 (plusmn000) 8251 (plusmn016) 100 5384 (plusmn023)lowastTenfold cross-validation used for mammograms and retina images leave-one-out used for brain MRIs

Table 4 SVM classification performance measures obtained withGabor-based features

Accuracy Specificity SensitivityMammograms 6803 (plusmn001) 100 0Retina 5000 (plusmn000) 100 0Brain MRI 8661 (plusmn003) 100 0

The obtained correct classification rate of mammogramsretina and brain magnetic resonance images is respectively6803 (plusmn001) 5000 (plusmn000) and 8661 (plusmn003) Theaverage results for classifier specificity and sensitivity for allimages are 100 and 0 This finding indicates that Gabor-based features are suitable to detect pathological imagesbut fails to detect normal images In sum the results showthat Gabor-based features do not perform better than DWTand DWT-Gabor-based features These findings confirm thesuperiority of combining theDWT andGabor filter banks forfeature extraction

Based on the previous results it appears that the DWT-Gabor approach for feature extraction is effective for detect-ing the abrupt changes in biological tissue that characterizethe pathological patterns that were investigated and it yieldsbetter classification accuracy and specificity than the DWT-only approach It also offers equal of better sensitivity exceptfor brain MRIs For brain MRIs the obtained specificityand sensitivity results with the DWT-Gabor approach showimproved true negative detection but lower true positiveperformance Finally the obtained results reveal also thatlevel-one DWT decomposition is preferable to level-twodecomposition

Finally Table 5 compares the results obtained with theDWT-Gabor approach to other work that we surveyed Inmany cases the DWT-Gabor method yields higher clas-sification rates particularly for mammograms and retinaFor the problem of brain MRI classification our obtainedperformance is better than the results of [38] but less thanwhat is reported in [6 8] However these comparisons shouldbe viewed with caution as not all the results stem froma common image database and the different authors usedifferent sample and image sizesMoreovermany authors useno cross-validation and simply performa single arbitrary splitof their data into training and test sets to obtain their accuracyresults Obviously one cannot generalize or draw definiteconclusion from such efforts and comparisons between

works cannot be made other than in general terms In thisrespect it can only be concluded from our results that theDWT-Gabor for feature extraction is effective for obtaininghigh image classification accuracy by an SVM and that itmay outperform other feature extraction and classificationtechniques reported in the literature at least those basedon DWT-only image decomposition Unfortunately a moredefinite conclusion is impossiblewithout gaining access to theimage databases used by the other authors

5 Computational Complexity

Finally the computational complexity of the DWT Gaborcontourlet and curvelet for an 119873 times 119873 image is respectively119874(119873) 119874(1198732 times 119872

2) with 119872 being the width of Gabor

(Gaussian) mask filter119874(1198732) and119874(1198732 log(119873)) As a resultthe computational complexity of the combination of theDWT and Gabor filter is 119874(119873) + 119874(1198732 times 1198722) In terms offeatures extraction processing time the average time requiredto process a brain a mammogram and a retina imagewith the DWT approach (DWT-Gabor) was respectively019 (031) 017 (032) and 015 seconds (035) using MatlabR2009a on a 15 GHz Core2 Duo processor

6 Conclusion

We proposed a supervised system for biomedical imagesclassification that uses statistical features obtained from thecombination of the discrete wavelet transform and Gaborfilter to classify normal images versus cancer images usingsupport vector machines as classifiers Our experimentalresults show that such a hybrid processing model achieveshigher accuracy in comparison to using DWT or Gaborfilter banks alone Therefore the proposed image processingand features extraction approach seem to be very promis-ing for the detection of certain pathologies in biomedicalimages

For future works it is recommended to consider a largerset of features and a selection process to identify the mostdiscriminant ones In addition the Gabor parameters willbe adjusted for each type of image separately to improvethe accuracy Furthermore the DWT-Gabor will be directlycompared to the dual-tree complex wavelet curvelet andcontourlet using the same databases and images in orderto draw general conclusions Also multilabels classificationswill be considered in future works to investigate the discrim-inative power of our approach for each type of pathology

Journal of Medical Engineering 11

Table 5 Comparison with the literature

Features Classifier Accuracylowast

Mammograms[3] Gabor k-NN 80[4] DT-CWT SVM 8864[5] Contourlet SVM 966Our approach DWT-Gabor SVM 9667 (plusmn005)

Retina[1] DWT + GLCM LDA 822

[2] Morphological + GLCM Probabilistic boosting algorithm 813ndash922717ndash852

[39] Gabor SVM 83Our approach DWT-Gabor SVM 100

Brain[6] DWT SVM 98[8] DWT + PCA BPNN 100

SVM 90[38] Voxels Bayes 92

VFI 78Our approach DWT-Gabor SVM 9736 (plusmn002)

lowastCorrect classification rate

Finally more experiments on the effect of kernel choice andits parameter on classification accuracy will be investigated

References

[1] A Khademi and S Krishnan ldquoShift-invariant discrete wavelettransform analysis for retinal image classificationrdquoMedical andBiological Engineering and Computing vol 45 no 12 pp 1211ndash1222 2007

[2] N Lee A F Laine andT R Smith ldquoLearning non-homogenoustextures and the unlearning problemwith application to drusendetection in retinal imagesrdquo in Proceedings of the 5th IEEEInternational Symposium on Biomedical Imaging From Nano toMacro (ISBI rsquo08) pp 1215ndash1218 Paris France May 2008

[3] A Dong and B Wang ldquoFeature selection and analysis onmammogram classificationrdquo in Proceedings of the IEEE PacificRim Conference on Communications Computers and SignalProcessing (PACRIM rsquo09) pp 731ndash735 Victoria BC CanadaAugust 2009

[4] A Tirtajaya and D D Santika ldquoClassification of microcalcifi-cation using dual-tree complex wavelet transform and supportvector machinerdquo in Proceedings of the 2nd International Confer-ence onAdvances in Computing Control and TelecommunicationTechnologies (ACT rsquo10) pp 164ndash166 Jakarta Indonesia Decem-ber 2010

[5] F Moayedi Z Azimifar R Boostani and S Katebi ldquoContour-let-based mammography mass classification using the SVMfamilyrdquo Computers in Biology and Medicine vol 40 no 4 pp373ndash383 2010

[6] S Chaplot LM Patnaik andN R Jagannathan ldquoClassificationof magnetic resonance brain images using wavelets as input tosupport vector machine and neural networkrdquo Biomedical SignalProcessing and Control vol 1 no 1 pp 86ndash92 2006

[7] Y Zhang S Wang and L Wu ldquoA novel method for magneticresonance brain image classification based on adaptive chaoticPSOrdquo Progress in Electromagnetics Research vol 109 pp 325ndash343 2010

[8] Y Zhang Z Dong L Wu and S Wang ldquoA hybrid method forMRI brain image classificationrdquo Expert Systems with Applica-tions vol 38 no 8 pp 10049ndash10053 2011

[9] M E Celebi H Iyatomi G Schaefer and W V StoeckerldquoLesion border detection in dermoscopy imagesrdquoComputerizedMedical Imaging and Graphics vol 33 no 2 pp 148ndash153 2009

[10] Q AbbasM E Celebi and I F Garcıa ldquoSkin tumor area extrac-tion using an improved dynamic programming approachrdquo SkinResearch and Technology vol 18 pp 133ndash142 2012

[11] Q Li F Li and K Doi ldquoComputerized detection of lungnodules in thin-section CT images by use of selective enhance-ment filters and an automated rule-based classifierrdquo AcademicRadiology vol 15 no 2 pp 165ndash175 2008

[12] A El-Bazl M Nitzken E Vanbogaertl G Gimelrsquojarb RFalfi and M Abo El-Ghar ldquoA novel shaped-based diagnosticapproach for early diagnosis of lung nodulesrdquo in Proceedingsof the IEEE International Symposium in Biomedical Imaging(ISBI rsquo11) pp 137ndash140 Chicago Ill USA 2011

[13] M T Coimbra and J P S Cunha ldquoMPEG-7 visual descriptorsmdashcontributions for automated feature extraction in capsuleendoscopyrdquo IEEE Transactions on Circuits and Systems for VideoTechnology vol 16 no 5 pp 628ndash636 2006

[14] B Li and M Q H Meng ldquoTexture analysis for ulcer detectionin capsule endoscopy imagesrdquo Image andVision Computing vol27 no 9 pp 1336ndash1342 2009

[15] C K Chui An Introduction to Wavelets Academic Press SanDiego Calif USA 1992

[16] M Vetterli andCHerley ldquoWavelets and filter banks theory anddesignrdquo IEEE Transactions on Signal Processing vol 40 no 9pp 2207ndash2232 1992

12 Journal of Medical Engineering

[17] J G Daugman ldquoUncertainty relation for resolution inspace spatial frequency and orientation optimized by two-dimensional visual cortical filtersrdquo Journal of the Optical Societyof America A vol 2 no 7 pp 1160ndash1169 1985

[18] I W Selesnick R G Baraniuk and N G Kingsbury ldquoThedual-tree complex wavelet transformrdquo IEEE Signal ProcessingMagazine vol 22 no 6 pp 123ndash151 2005

[19] E Candes and D Donoho ldquoRidgelets a key to higher-dimensional intermittencyrdquo Philosophical Transactions of theLondon Royal Society vol 357 pp 2495ndash2509 1999

[20] E J Candes and D L Donoho ldquoContinuous curvelettransformmdashI Resolution of the wavefront setrdquo Applied andComputational Harmonic Analysis vol 19 no 2 pp 162ndash1972005

[21] M N Do and M Vetterli ldquoThe contourlet transform an effi-cient directional multiresolution image representationrdquo IEEETransactions on Image Processing vol 14 no 12 pp 2091ndash21062005

[22] R J Ferrari R M Rangayyan J E L Desautels and A FFrere ldquoAnalysis of asymmetry in mammograms via directionalfiltering with Gabor waveletsrdquo IEEE Transactions on MedicalImaging vol 20 no 9 pp 953ndash964 2001

[23] Z Cui and G Zhang ldquoA novel medical image dynamic fuzzyclassification model based on ridgelet transformrdquo Journal ofSoftware vol 5 no 5 pp 458ndash465 2010

[24] T Geback and P Koumoutsakos ldquoEdge detection inmicroscopyimages using curveletsrdquo BMC Bioinformatics vol 10 article 752009

[25] J Ma andG Plonka ldquoThe curvelet transform a review of recentapplicationsrdquo IEEE Signal ProcessingMagazine vol 27 no 2 pp118ndash133 2010

[26] N Kingsbury ldquoComplex wavelets and shift invariancerdquo inProceedings of the IEEE Seminar on Time-Scale and Time-Frequency Analysis and Applications pp 501ndash510 London UK2000

[27] Y L Qiao C Y Song and C H Zhao ldquoM-band ridgelettransform based texture classificationrdquo Pattern RecognitionLetters vol 31 no 3 pp 244ndash249 2010

[28] F Gomez and E Romero ldquoTexture characterization using acurvelet based descriptorrdquo Lecture Notes in Computer Sciencevol 5856 pp 113ndash120 2009

[29] H Shan and J Ma ldquoCurvelet-based geodesic snakes for imagesegmentation with multiple objectsrdquo Pattern Recognition Let-ters vol 31 no 5 pp 355ndash360 2010

[30] R Eslami and H Radha ldquoNew image transforms using hybridwavelets and directional filter banks analysis and designrdquo inProceedings of the IEEE International Conference on ImageProcessing (ICIP rsquo05) pp 733ndash736 Genova Italy September2005

[31] O O V Villegas and V G C Sanchez ldquoThe wavelet basedcontourlet transform and its application to feature preservingimage codingrdquo Lecture Notes in Computer Science vol 4827 pp590ndash600 2007

[32] S Lahmir and M Boukadoum ldquoClassification of brain MRIusing the LH and HL wavelet transform sub-bandsrdquo in Pro-ceedings of the IEEE International Symposium on Circuits andSystems (ISCAS rsquo11) pp 1025ndash1028 Rio de Janeiro Brazil May2009 2011

[33] S Lahmir and M Boukadoum ldquoBrain MRI classificationusing an ensemble system and LH and HL wavelet Sub-bandsFeaturesrdquo in Proceedings of the IEEE Symposium Series on

Computational Intelligence (SSCI rsquo11) pp 1ndash7 Paris FranceApril2011

[34] S Lahmir and M Boukadoum ldquoHybrid Cosine and RadonTransform-based processing for Digital Mammogram FeatureExtraction and Classification with SVMrdquo in Proceedings of the33rd IEEE Annual International Conference on Engineering inMedecine and Biology Society (EMBS rsquo11) pp 5104ndash5107 BostonMass USA 2011

[35] S Lahmir andM Boukadoum ldquoDWT and RT-Based Approachfor Feature Extraction and classification of Mammograms withSVMrdquo in Proceedings of the IEEE Biomedical Circuits andSystems Conference (BioCAS rsquo11) pp 412ndash415 San Diego CalifUSA November 2011

[36] VN VapnikTheNature of Statistical LearningTheory Springer1995

[37] S Lahmiri and M Boukadoum ldquoHybrid discret wavelet trans-form and Gabor filter banks processing for mammogramfeatures extractionrdquo in Proceedings of the IEEE New Circuits andSystems (NEWCAS rsquo11) pp 53ndash56 Bordeaux France June 2011

[38] L M Bruce and N Shanmugam ldquoUsing neural networkswith wavelet transforms for an automated mammographicmass classifierrdquo in Proceedings of the 22nd Annual InternationalConference of the IEEE Engineering in Medicine and BiologySociety pp 985ndash987 Chicago Ill USA July 2000

[39] S M H Jamarani G Rezai-rad and H Behnam ldquoA novelmethod for breast cancer prognosis using wavelet packet basedneural networkrdquo in Proceedings of the 27th Annual InternationalConference of the Engineering in Medicine and Biology Society(IEEE-EMBS rsquo05) pp 3414ndash3417 Shanghai China September2005

[40] httpmedical-dictionarythefreedictionarycomcircinate+retinopathy

[41] C I O Martins F N S Medeiros R M S Veras F N Bezerraand R M Cesar Jr ldquoEvaluation of retinal vessel segmentationmethods for microaneurysms detectionrdquo in Proceedings of theIEEE International Conference on Image Processing (ICIP rsquo09)pp 3365ndash3368 Cairo Egypt November 2009

[42] C Plant S J Teipel A Oswald et al ldquoAutomated detectionof brain atrophy patterns based on MRI for the prediction ofAlzheimerrsquos diseaserdquo NeuroImage vol 50 no 1 pp 162ndash1742010

[43] J Meier R Bock L G Nyul and G Michelson ldquoEye fun-dus image processing system for automated glaucoma clas-sificationrdquo in Proceedings of the 52nd Internationales Wis-senschaftliches Kolloquium Technische Universitat Ilmenau2007

[44] E Sakka A Prentza I E Lamprinos and D KoutsourisldquoMicrocalcification detection using multiresolution analysisbased on wavelet transformrdquo in Proceedings of the IEEE Inter-national Special Topic Conference on Information Technology inBiomedicine Ioannina Greece October 2006

[45] S G Mallat ldquoTheory for multiresolution signal decompositionthe wavelet representationrdquo IEEE Transactions on Pattern Anal-ysis and Machine Intelligence vol 11 no 7 pp 674ndash693 1989

[46] J Suckling J Parker and D R Dance ldquoThe mammographicimage analysis society digital mammogram databaserdquo in Pro-ceedings of the the 2nd International Workshop on DigitalMammography A G Gale S M Astley D D Dance and AY Cairns Eds pp 375ndash378 Elsevier York UK 1994

[47] H J Chiou C Y Chen T C Liu et al ldquoComputer-aideddiagnosis of peripheral soft tissue masses based on ultrasound

Journal of Medical Engineering 13

imagingrdquo Computerized Medical Imaging and Graphics vol 33no 5 pp 408ndash413 2009

[48] J K Kim J M Park K S Song and H W Park ldquoAdaptivemammographic image enhancement using first derivative andlocal statisticsrdquo IEEE Transactions on Medical Imaging vol 16no 5 pp 495ndash502 1997

[49] H S Sheshadri andAKandaswamy ldquoBreast tissue classificationusing statistical feature extraction of mammogramsrdquo MedicalImaging and Information Sciences vol 23 no 3 pp 105ndash1072006

[50] N Cristianini and J Shawe-Taylor Introduction to SupportVector Machines and Other Kernel-Based Learning MethodsCambridge University Press Cambridge UK 2000

[51] L Chen X Mao Y Xue and L L Cheng ldquoSpeech emotionrecognition features and classification modelsrdquo Digital SignalProcessing vol 22 pp 1154ndash1160 2012

[52] F Palmieri U Fiore A Castiglione and A De Santis ldquoOn thedetection of card-sharing traffic through wavelet analysis andSupport Vector Machinesrdquo Applied Soft Computing vol 13 no1 pp 615ndash627 2013

[53] A Azadeh M Saberi A Kazem V Ebrahimipour A Nour-mohammadzadeh and Z Saberi ldquoA flexible algorithm forfault diagnosis in a centrifugal pump with corrupted data andnoise based on ANN and support vector machine with hyper-parameters optimizationrdquoApplied Soft Computing vol 13 no 3pp 1478ndash1485 2013

[54] R J Martis U R Acharya K M Mandana A K Ray andC Chakraborty ldquoCardiac decision making using higher orderspectrardquo Biomedical Signal Processing and Control vol 8 pp193ndash203 2013

[55] M R Mohammad ldquoChi-square distance kernel of the gaitsfor the diagnosis of Parkinsonrsquos diseaserdquo Biomedical SignalProcessing and Control vol 8 pp 66ndash70 2013

[56] R Vandenberghe N Nelissen E Salmon et al ldquoBinary clas-sification of 18F-flutemetamol PET using machine learningcomparison with visual reads and structuralMRIrdquoNeuroImagevol 64 no 1 pp 517ndash525 2013

[57] httpmarathoncseeusfeduMammographyDatabasehtml[58] httpwwwcesclemsonedusimahooverstare[59] httpwwwmedharvardeduaanlib

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

8 Journal of Medical Engineering

Table 1 Average SVM classification accuracy as a function of feature extraction method and level of DWT decompositionlowast

DWT DWT-Gabor DWT DWT-GaborDecomposition level One One Two TwoMammograms 9598 (plusmn004) 9667 (plusmn005) 8913 (plusmn001) 9109 (plusmn005)Retina 7469 (plusmn005) 100 9098 (plusmn003) 100Brain MRI 8780 (plusmn000) 9736 (plusmn002) 8576 (plusmn000) 9118 (plusmn004)lowastTenfold cross-validation used for mammograms and retina images leave-one-out used for brain MRIs

Table 2 SVM classification specificity as a function of feature extraction method and level of DWT decompositionlowast

DWT DWT-Gabor DWT DWT-GaborDecomposition level One One Two TwoMammograms 9781 (plusmn002) 100 8885 (plusmn005) 9209 (plusmn006)Retina 681 (plusmn004) 100 232 (plusmn010) 100Brain MRI 2155 (plusmn001) 9958 (plusmn001) 0 9724 (plusmn001)lowastTenfold cross-validation used for mammograms and retina images leave-one-out used for brain MRIs

[57] For retina a set of 69 color images (150 times 130 pixels)from the STARE [58] database were employed including23 normal images 24 with microaneurysms and 22 withcircinate Finally a collection of 56 axial T2-weighted andMR brain images (256 times 256 pixels) were taken from theAANLIB database [59] of the Harvard Medical School Theyconsisted of 7 normal images 9 with Alzheimerrsquos disease 13with glioma 8 with Herpes encephalitis 8 with metastaticbronchogenic carcinoma and 14 with multiple sclerosis Itis unfortunate that the number of images was not constantacross pathologies but we had no control over this and usedwhat was available with tenfold cross-validation or leave-one-out cross-validation of the results depending on samplesize All experiments were based on a binary classificationapproach of normal versus abnormal images Many kinds ofbiomedical images could be considered for our experimentswe focused on mammograms retina and brain magneticresonance images mainly because of public availability Anexample of the processing of a normal retina and a retina withcircinate is illustrated in Figures 7 and 8 respectively

For each image type the average and standard deviationof the correct classification rate (CCR) sensitivity and speci-ficity were computed to evaluate the performance featureextraction techniquewhen used in conjunctionwith the SVMclassifier The three performance measures are defined by

CCR =Classified Samples

Total Number of Samples

Sensitivity =Correctly Classified Positive Samples

True Positive Samples

Specificity =Correctly Classified Negative Samples

True Negative Samples

(17)

where positive samples and negative samples are respectivelyabnormal and normal images

Finally all experiments were performed with tenfoldcross-validation except those for MR images which usedleave-one-out cross-validation due to the small sample sizeof each brain image category

Table 1 shows the obtained average results for the threetypes of images that were investigated The performance ofthe SVM classifier improved for all types of images and alllevels of HH decomposition by the DWT At level one theaverage correct classification rate increased by respectively069 2531 and 956 percentage points for mammogramsretina and brain magnetic resonance images when usingthe DWT-Gabor approach At level-two decomposition theimprovement was respectively 196 902 and 542 percent-age points

Tables 2 and 3 provide the average results for classifiersensitivity and specificity At level-one DWT decompositionthe DWT-Gabor approach improved classification specificityfor mammograms and retina images to make it reach 100while it improved it by 7803 percentage points for brain MRimages At level-two DWT decomposition the improvementwas 324 percentage points for mammograms 9768 percent-age points for retina images (100 specificity) and 9724percentage points for MR images Regarding sensitivity theresults were mixed

At level-one DWT decomposition the values were aboutthe same for the DWT-Gabor and DWT-only approachesfor mammograms and retina images with respectivelyminus081 and 0 percentage points differences but there was adegradation of minus1133 and minus4616 percentage points for brainMR images at level one and level two of decompositionrespectively

Following the same cross-validation protocol we alsoconducted classification experiments with features extractedfrom aGabor filtered image of the original biomedical imageThe purpose was to check whether Gabor-based featuresalone help characterize images better than DWT or DWT-Gabor-based features The results are given in Table 4

Journal of Medical Engineering 9

20 40 60 80 100 120 140

20

40

60

80

100

120020406080100120140160180200

(a) Original normal retina image

10 20 30 40 50 60 70

10

20

30

40

50

60 minus60

minus40

minus20

0

20

40

60

(b) Normal retina HH1 subimage

0

5

10

15

20

25

30

35

20406080100

minus100

minus80

minus60

minus40

minus20

5 10 15 20 25 30 35 40(c) Normal retina HH2 subimage

10 20 30 40 50 60 70

10

20

30

40

50

60 10

20

30

40

50

60

70

(d) Gabor filtered HH1 at 119891 = 2 and 120579 = 1205874

5

10

15

20

25

30

355 10 15 20 25 30 35 40

102030405060708090100

(e) Gabor filtered HH2 119891 = 2 and 120579 = 1205874

Figure 7 Analysis of a normal retina

20 40 60 80 100 120 140

20

40

60

80

100

120 20406080100120140160180200220

(a) Original retina image with circinate

10 20 30 40 50 60 70

10

20

30

40

50

60minus20

minus10

0

10

20

30

(b) HH1 subimage

5

10

15

20

25

30

355 10 15 20 25 30 35 40

minus80

minus60

minus40

minus20020406080100

(c) HH2 subimage

10 20 30 40 50 60 70

10

20

30

40

50

600

5

10

15

20

25

30

35

(d) Gabor filtered HH1 at 119891 = 2 and 120579 = 1205874

102030405060708090100

5

10

15

20

25

30

355 10 15 20 25 30 35 40

(e) Gabor filtered HH2 at 119891 = 2 and 120579 = 1205874

Figure 8 Analysis of a retina with circinate

10 Journal of Medical Engineering

Table 3 SVM classification sensitivity as a function of feature extraction method and level of DWT decompositionlowast

DWT DWT-Gabor DWT DWT-GaborDecomposition level One One Two TwoMammograms 9414 (plusmn006) 9333 (plusmn006) 9029 (plusmn0039) 8978 (plusmn004)Retina 100 100 100 100Brain MRI 9384 (plusmn000) 8251 (plusmn016) 100 5384 (plusmn023)lowastTenfold cross-validation used for mammograms and retina images leave-one-out used for brain MRIs

Table 4 SVM classification performance measures obtained withGabor-based features

Accuracy Specificity SensitivityMammograms 6803 (plusmn001) 100 0Retina 5000 (plusmn000) 100 0Brain MRI 8661 (plusmn003) 100 0

The obtained correct classification rate of mammogramsretina and brain magnetic resonance images is respectively6803 (plusmn001) 5000 (plusmn000) and 8661 (plusmn003) Theaverage results for classifier specificity and sensitivity for allimages are 100 and 0 This finding indicates that Gabor-based features are suitable to detect pathological imagesbut fails to detect normal images In sum the results showthat Gabor-based features do not perform better than DWTand DWT-Gabor-based features These findings confirm thesuperiority of combining theDWT andGabor filter banks forfeature extraction

Based on the previous results it appears that the DWT-Gabor approach for feature extraction is effective for detect-ing the abrupt changes in biological tissue that characterizethe pathological patterns that were investigated and it yieldsbetter classification accuracy and specificity than the DWT-only approach It also offers equal of better sensitivity exceptfor brain MRIs For brain MRIs the obtained specificityand sensitivity results with the DWT-Gabor approach showimproved true negative detection but lower true positiveperformance Finally the obtained results reveal also thatlevel-one DWT decomposition is preferable to level-twodecomposition

Finally Table 5 compares the results obtained with theDWT-Gabor approach to other work that we surveyed Inmany cases the DWT-Gabor method yields higher clas-sification rates particularly for mammograms and retinaFor the problem of brain MRI classification our obtainedperformance is better than the results of [38] but less thanwhat is reported in [6 8] However these comparisons shouldbe viewed with caution as not all the results stem froma common image database and the different authors usedifferent sample and image sizesMoreovermany authors useno cross-validation and simply performa single arbitrary splitof their data into training and test sets to obtain their accuracyresults Obviously one cannot generalize or draw definiteconclusion from such efforts and comparisons between

works cannot be made other than in general terms In thisrespect it can only be concluded from our results that theDWT-Gabor for feature extraction is effective for obtaininghigh image classification accuracy by an SVM and that itmay outperform other feature extraction and classificationtechniques reported in the literature at least those basedon DWT-only image decomposition Unfortunately a moredefinite conclusion is impossiblewithout gaining access to theimage databases used by the other authors

5 Computational Complexity

Finally the computational complexity of the DWT Gaborcontourlet and curvelet for an 119873 times 119873 image is respectively119874(119873) 119874(1198732 times 119872

2) with 119872 being the width of Gabor

(Gaussian) mask filter119874(1198732) and119874(1198732 log(119873)) As a resultthe computational complexity of the combination of theDWT and Gabor filter is 119874(119873) + 119874(1198732 times 1198722) In terms offeatures extraction processing time the average time requiredto process a brain a mammogram and a retina imagewith the DWT approach (DWT-Gabor) was respectively019 (031) 017 (032) and 015 seconds (035) using MatlabR2009a on a 15 GHz Core2 Duo processor

6 Conclusion

We proposed a supervised system for biomedical imagesclassification that uses statistical features obtained from thecombination of the discrete wavelet transform and Gaborfilter to classify normal images versus cancer images usingsupport vector machines as classifiers Our experimentalresults show that such a hybrid processing model achieveshigher accuracy in comparison to using DWT or Gaborfilter banks alone Therefore the proposed image processingand features extraction approach seem to be very promis-ing for the detection of certain pathologies in biomedicalimages

For future works it is recommended to consider a largerset of features and a selection process to identify the mostdiscriminant ones In addition the Gabor parameters willbe adjusted for each type of image separately to improvethe accuracy Furthermore the DWT-Gabor will be directlycompared to the dual-tree complex wavelet curvelet andcontourlet using the same databases and images in orderto draw general conclusions Also multilabels classificationswill be considered in future works to investigate the discrim-inative power of our approach for each type of pathology

Journal of Medical Engineering 11

Table 5 Comparison with the literature

Features Classifier Accuracylowast

Mammograms[3] Gabor k-NN 80[4] DT-CWT SVM 8864[5] Contourlet SVM 966Our approach DWT-Gabor SVM 9667 (plusmn005)

Retina[1] DWT + GLCM LDA 822

[2] Morphological + GLCM Probabilistic boosting algorithm 813ndash922717ndash852

[39] Gabor SVM 83Our approach DWT-Gabor SVM 100

Brain[6] DWT SVM 98[8] DWT + PCA BPNN 100

SVM 90[38] Voxels Bayes 92

VFI 78Our approach DWT-Gabor SVM 9736 (plusmn002)

lowastCorrect classification rate

Finally more experiments on the effect of kernel choice andits parameter on classification accuracy will be investigated

References

[1] A Khademi and S Krishnan ldquoShift-invariant discrete wavelettransform analysis for retinal image classificationrdquoMedical andBiological Engineering and Computing vol 45 no 12 pp 1211ndash1222 2007

[2] N Lee A F Laine andT R Smith ldquoLearning non-homogenoustextures and the unlearning problemwith application to drusendetection in retinal imagesrdquo in Proceedings of the 5th IEEEInternational Symposium on Biomedical Imaging From Nano toMacro (ISBI rsquo08) pp 1215ndash1218 Paris France May 2008

[3] A Dong and B Wang ldquoFeature selection and analysis onmammogram classificationrdquo in Proceedings of the IEEE PacificRim Conference on Communications Computers and SignalProcessing (PACRIM rsquo09) pp 731ndash735 Victoria BC CanadaAugust 2009

[4] A Tirtajaya and D D Santika ldquoClassification of microcalcifi-cation using dual-tree complex wavelet transform and supportvector machinerdquo in Proceedings of the 2nd International Confer-ence onAdvances in Computing Control and TelecommunicationTechnologies (ACT rsquo10) pp 164ndash166 Jakarta Indonesia Decem-ber 2010

[5] F Moayedi Z Azimifar R Boostani and S Katebi ldquoContour-let-based mammography mass classification using the SVMfamilyrdquo Computers in Biology and Medicine vol 40 no 4 pp373ndash383 2010

[6] S Chaplot LM Patnaik andN R Jagannathan ldquoClassificationof magnetic resonance brain images using wavelets as input tosupport vector machine and neural networkrdquo Biomedical SignalProcessing and Control vol 1 no 1 pp 86ndash92 2006

[7] Y Zhang S Wang and L Wu ldquoA novel method for magneticresonance brain image classification based on adaptive chaoticPSOrdquo Progress in Electromagnetics Research vol 109 pp 325ndash343 2010

[8] Y Zhang Z Dong L Wu and S Wang ldquoA hybrid method forMRI brain image classificationrdquo Expert Systems with Applica-tions vol 38 no 8 pp 10049ndash10053 2011

[9] M E Celebi H Iyatomi G Schaefer and W V StoeckerldquoLesion border detection in dermoscopy imagesrdquoComputerizedMedical Imaging and Graphics vol 33 no 2 pp 148ndash153 2009

[10] Q AbbasM E Celebi and I F Garcıa ldquoSkin tumor area extrac-tion using an improved dynamic programming approachrdquo SkinResearch and Technology vol 18 pp 133ndash142 2012

[11] Q Li F Li and K Doi ldquoComputerized detection of lungnodules in thin-section CT images by use of selective enhance-ment filters and an automated rule-based classifierrdquo AcademicRadiology vol 15 no 2 pp 165ndash175 2008

[12] A El-Bazl M Nitzken E Vanbogaertl G Gimelrsquojarb RFalfi and M Abo El-Ghar ldquoA novel shaped-based diagnosticapproach for early diagnosis of lung nodulesrdquo in Proceedingsof the IEEE International Symposium in Biomedical Imaging(ISBI rsquo11) pp 137ndash140 Chicago Ill USA 2011

[13] M T Coimbra and J P S Cunha ldquoMPEG-7 visual descriptorsmdashcontributions for automated feature extraction in capsuleendoscopyrdquo IEEE Transactions on Circuits and Systems for VideoTechnology vol 16 no 5 pp 628ndash636 2006

[14] B Li and M Q H Meng ldquoTexture analysis for ulcer detectionin capsule endoscopy imagesrdquo Image andVision Computing vol27 no 9 pp 1336ndash1342 2009

[15] C K Chui An Introduction to Wavelets Academic Press SanDiego Calif USA 1992

[16] M Vetterli andCHerley ldquoWavelets and filter banks theory anddesignrdquo IEEE Transactions on Signal Processing vol 40 no 9pp 2207ndash2232 1992

12 Journal of Medical Engineering

[17] J G Daugman ldquoUncertainty relation for resolution inspace spatial frequency and orientation optimized by two-dimensional visual cortical filtersrdquo Journal of the Optical Societyof America A vol 2 no 7 pp 1160ndash1169 1985

[18] I W Selesnick R G Baraniuk and N G Kingsbury ldquoThedual-tree complex wavelet transformrdquo IEEE Signal ProcessingMagazine vol 22 no 6 pp 123ndash151 2005

[19] E Candes and D Donoho ldquoRidgelets a key to higher-dimensional intermittencyrdquo Philosophical Transactions of theLondon Royal Society vol 357 pp 2495ndash2509 1999

[20] E J Candes and D L Donoho ldquoContinuous curvelettransformmdashI Resolution of the wavefront setrdquo Applied andComputational Harmonic Analysis vol 19 no 2 pp 162ndash1972005

[21] M N Do and M Vetterli ldquoThe contourlet transform an effi-cient directional multiresolution image representationrdquo IEEETransactions on Image Processing vol 14 no 12 pp 2091ndash21062005

[22] R J Ferrari R M Rangayyan J E L Desautels and A FFrere ldquoAnalysis of asymmetry in mammograms via directionalfiltering with Gabor waveletsrdquo IEEE Transactions on MedicalImaging vol 20 no 9 pp 953ndash964 2001

[23] Z Cui and G Zhang ldquoA novel medical image dynamic fuzzyclassification model based on ridgelet transformrdquo Journal ofSoftware vol 5 no 5 pp 458ndash465 2010

[24] T Geback and P Koumoutsakos ldquoEdge detection inmicroscopyimages using curveletsrdquo BMC Bioinformatics vol 10 article 752009

[25] J Ma andG Plonka ldquoThe curvelet transform a review of recentapplicationsrdquo IEEE Signal ProcessingMagazine vol 27 no 2 pp118ndash133 2010

[26] N Kingsbury ldquoComplex wavelets and shift invariancerdquo inProceedings of the IEEE Seminar on Time-Scale and Time-Frequency Analysis and Applications pp 501ndash510 London UK2000

[27] Y L Qiao C Y Song and C H Zhao ldquoM-band ridgelettransform based texture classificationrdquo Pattern RecognitionLetters vol 31 no 3 pp 244ndash249 2010

[28] F Gomez and E Romero ldquoTexture characterization using acurvelet based descriptorrdquo Lecture Notes in Computer Sciencevol 5856 pp 113ndash120 2009

[29] H Shan and J Ma ldquoCurvelet-based geodesic snakes for imagesegmentation with multiple objectsrdquo Pattern Recognition Let-ters vol 31 no 5 pp 355ndash360 2010

[30] R Eslami and H Radha ldquoNew image transforms using hybridwavelets and directional filter banks analysis and designrdquo inProceedings of the IEEE International Conference on ImageProcessing (ICIP rsquo05) pp 733ndash736 Genova Italy September2005

[31] O O V Villegas and V G C Sanchez ldquoThe wavelet basedcontourlet transform and its application to feature preservingimage codingrdquo Lecture Notes in Computer Science vol 4827 pp590ndash600 2007

[32] S Lahmir and M Boukadoum ldquoClassification of brain MRIusing the LH and HL wavelet transform sub-bandsrdquo in Pro-ceedings of the IEEE International Symposium on Circuits andSystems (ISCAS rsquo11) pp 1025ndash1028 Rio de Janeiro Brazil May2009 2011

[33] S Lahmir and M Boukadoum ldquoBrain MRI classificationusing an ensemble system and LH and HL wavelet Sub-bandsFeaturesrdquo in Proceedings of the IEEE Symposium Series on

Computational Intelligence (SSCI rsquo11) pp 1ndash7 Paris FranceApril2011

[34] S Lahmir and M Boukadoum ldquoHybrid Cosine and RadonTransform-based processing for Digital Mammogram FeatureExtraction and Classification with SVMrdquo in Proceedings of the33rd IEEE Annual International Conference on Engineering inMedecine and Biology Society (EMBS rsquo11) pp 5104ndash5107 BostonMass USA 2011

[35] S Lahmir andM Boukadoum ldquoDWT and RT-Based Approachfor Feature Extraction and classification of Mammograms withSVMrdquo in Proceedings of the IEEE Biomedical Circuits andSystems Conference (BioCAS rsquo11) pp 412ndash415 San Diego CalifUSA November 2011

[36] VN VapnikTheNature of Statistical LearningTheory Springer1995

[37] S Lahmiri and M Boukadoum ldquoHybrid discret wavelet trans-form and Gabor filter banks processing for mammogramfeatures extractionrdquo in Proceedings of the IEEE New Circuits andSystems (NEWCAS rsquo11) pp 53ndash56 Bordeaux France June 2011

[38] L M Bruce and N Shanmugam ldquoUsing neural networkswith wavelet transforms for an automated mammographicmass classifierrdquo in Proceedings of the 22nd Annual InternationalConference of the IEEE Engineering in Medicine and BiologySociety pp 985ndash987 Chicago Ill USA July 2000

[39] S M H Jamarani G Rezai-rad and H Behnam ldquoA novelmethod for breast cancer prognosis using wavelet packet basedneural networkrdquo in Proceedings of the 27th Annual InternationalConference of the Engineering in Medicine and Biology Society(IEEE-EMBS rsquo05) pp 3414ndash3417 Shanghai China September2005

[40] httpmedical-dictionarythefreedictionarycomcircinate+retinopathy

[41] C I O Martins F N S Medeiros R M S Veras F N Bezerraand R M Cesar Jr ldquoEvaluation of retinal vessel segmentationmethods for microaneurysms detectionrdquo in Proceedings of theIEEE International Conference on Image Processing (ICIP rsquo09)pp 3365ndash3368 Cairo Egypt November 2009

[42] C Plant S J Teipel A Oswald et al ldquoAutomated detectionof brain atrophy patterns based on MRI for the prediction ofAlzheimerrsquos diseaserdquo NeuroImage vol 50 no 1 pp 162ndash1742010

[43] J Meier R Bock L G Nyul and G Michelson ldquoEye fun-dus image processing system for automated glaucoma clas-sificationrdquo in Proceedings of the 52nd Internationales Wis-senschaftliches Kolloquium Technische Universitat Ilmenau2007

[44] E Sakka A Prentza I E Lamprinos and D KoutsourisldquoMicrocalcification detection using multiresolution analysisbased on wavelet transformrdquo in Proceedings of the IEEE Inter-national Special Topic Conference on Information Technology inBiomedicine Ioannina Greece October 2006

[45] S G Mallat ldquoTheory for multiresolution signal decompositionthe wavelet representationrdquo IEEE Transactions on Pattern Anal-ysis and Machine Intelligence vol 11 no 7 pp 674ndash693 1989

[46] J Suckling J Parker and D R Dance ldquoThe mammographicimage analysis society digital mammogram databaserdquo in Pro-ceedings of the the 2nd International Workshop on DigitalMammography A G Gale S M Astley D D Dance and AY Cairns Eds pp 375ndash378 Elsevier York UK 1994

[47] H J Chiou C Y Chen T C Liu et al ldquoComputer-aideddiagnosis of peripheral soft tissue masses based on ultrasound

Journal of Medical Engineering 13

imagingrdquo Computerized Medical Imaging and Graphics vol 33no 5 pp 408ndash413 2009

[48] J K Kim J M Park K S Song and H W Park ldquoAdaptivemammographic image enhancement using first derivative andlocal statisticsrdquo IEEE Transactions on Medical Imaging vol 16no 5 pp 495ndash502 1997

[49] H S Sheshadri andAKandaswamy ldquoBreast tissue classificationusing statistical feature extraction of mammogramsrdquo MedicalImaging and Information Sciences vol 23 no 3 pp 105ndash1072006

[50] N Cristianini and J Shawe-Taylor Introduction to SupportVector Machines and Other Kernel-Based Learning MethodsCambridge University Press Cambridge UK 2000

[51] L Chen X Mao Y Xue and L L Cheng ldquoSpeech emotionrecognition features and classification modelsrdquo Digital SignalProcessing vol 22 pp 1154ndash1160 2012

[52] F Palmieri U Fiore A Castiglione and A De Santis ldquoOn thedetection of card-sharing traffic through wavelet analysis andSupport Vector Machinesrdquo Applied Soft Computing vol 13 no1 pp 615ndash627 2013

[53] A Azadeh M Saberi A Kazem V Ebrahimipour A Nour-mohammadzadeh and Z Saberi ldquoA flexible algorithm forfault diagnosis in a centrifugal pump with corrupted data andnoise based on ANN and support vector machine with hyper-parameters optimizationrdquoApplied Soft Computing vol 13 no 3pp 1478ndash1485 2013

[54] R J Martis U R Acharya K M Mandana A K Ray andC Chakraborty ldquoCardiac decision making using higher orderspectrardquo Biomedical Signal Processing and Control vol 8 pp193ndash203 2013

[55] M R Mohammad ldquoChi-square distance kernel of the gaitsfor the diagnosis of Parkinsonrsquos diseaserdquo Biomedical SignalProcessing and Control vol 8 pp 66ndash70 2013

[56] R Vandenberghe N Nelissen E Salmon et al ldquoBinary clas-sification of 18F-flutemetamol PET using machine learningcomparison with visual reads and structuralMRIrdquoNeuroImagevol 64 no 1 pp 517ndash525 2013

[57] httpmarathoncseeusfeduMammographyDatabasehtml[58] httpwwwcesclemsonedusimahooverstare[59] httpwwwmedharvardeduaanlib

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Journal of Medical Engineering 9

20 40 60 80 100 120 140

20

40

60

80

100

120020406080100120140160180200

(a) Original normal retina image

10 20 30 40 50 60 70

10

20

30

40

50

60 minus60

minus40

minus20

0

20

40

60

(b) Normal retina HH1 subimage

0

5

10

15

20

25

30

35

20406080100

minus100

minus80

minus60

minus40

minus20

5 10 15 20 25 30 35 40(c) Normal retina HH2 subimage

10 20 30 40 50 60 70

10

20

30

40

50

60 10

20

30

40

50

60

70

(d) Gabor filtered HH1 at 119891 = 2 and 120579 = 1205874

5

10

15

20

25

30

355 10 15 20 25 30 35 40

102030405060708090100

(e) Gabor filtered HH2 119891 = 2 and 120579 = 1205874

Figure 7 Analysis of a normal retina

20 40 60 80 100 120 140

20

40

60

80

100

120 20406080100120140160180200220

(a) Original retina image with circinate

10 20 30 40 50 60 70

10

20

30

40

50

60minus20

minus10

0

10

20

30

(b) HH1 subimage

5

10

15

20

25

30

355 10 15 20 25 30 35 40

minus80

minus60

minus40

minus20020406080100

(c) HH2 subimage

10 20 30 40 50 60 70

10

20

30

40

50

600

5

10

15

20

25

30

35

(d) Gabor filtered HH1 at 119891 = 2 and 120579 = 1205874

102030405060708090100

5

10

15

20

25

30

355 10 15 20 25 30 35 40

(e) Gabor filtered HH2 at 119891 = 2 and 120579 = 1205874

Figure 8 Analysis of a retina with circinate

10 Journal of Medical Engineering

Table 3 SVM classification sensitivity as a function of feature extraction method and level of DWT decompositionlowast

DWT DWT-Gabor DWT DWT-GaborDecomposition level One One Two TwoMammograms 9414 (plusmn006) 9333 (plusmn006) 9029 (plusmn0039) 8978 (plusmn004)Retina 100 100 100 100Brain MRI 9384 (plusmn000) 8251 (plusmn016) 100 5384 (plusmn023)lowastTenfold cross-validation used for mammograms and retina images leave-one-out used for brain MRIs

Table 4 SVM classification performance measures obtained withGabor-based features

Accuracy Specificity SensitivityMammograms 6803 (plusmn001) 100 0Retina 5000 (plusmn000) 100 0Brain MRI 8661 (plusmn003) 100 0

The obtained correct classification rate of mammogramsretina and brain magnetic resonance images is respectively6803 (plusmn001) 5000 (plusmn000) and 8661 (plusmn003) Theaverage results for classifier specificity and sensitivity for allimages are 100 and 0 This finding indicates that Gabor-based features are suitable to detect pathological imagesbut fails to detect normal images In sum the results showthat Gabor-based features do not perform better than DWTand DWT-Gabor-based features These findings confirm thesuperiority of combining theDWT andGabor filter banks forfeature extraction

Based on the previous results it appears that the DWT-Gabor approach for feature extraction is effective for detect-ing the abrupt changes in biological tissue that characterizethe pathological patterns that were investigated and it yieldsbetter classification accuracy and specificity than the DWT-only approach It also offers equal of better sensitivity exceptfor brain MRIs For brain MRIs the obtained specificityand sensitivity results with the DWT-Gabor approach showimproved true negative detection but lower true positiveperformance Finally the obtained results reveal also thatlevel-one DWT decomposition is preferable to level-twodecomposition

Finally Table 5 compares the results obtained with theDWT-Gabor approach to other work that we surveyed Inmany cases the DWT-Gabor method yields higher clas-sification rates particularly for mammograms and retinaFor the problem of brain MRI classification our obtainedperformance is better than the results of [38] but less thanwhat is reported in [6 8] However these comparisons shouldbe viewed with caution as not all the results stem froma common image database and the different authors usedifferent sample and image sizesMoreovermany authors useno cross-validation and simply performa single arbitrary splitof their data into training and test sets to obtain their accuracyresults Obviously one cannot generalize or draw definiteconclusion from such efforts and comparisons between

works cannot be made other than in general terms In thisrespect it can only be concluded from our results that theDWT-Gabor for feature extraction is effective for obtaininghigh image classification accuracy by an SVM and that itmay outperform other feature extraction and classificationtechniques reported in the literature at least those basedon DWT-only image decomposition Unfortunately a moredefinite conclusion is impossiblewithout gaining access to theimage databases used by the other authors

5 Computational Complexity

Finally the computational complexity of the DWT Gaborcontourlet and curvelet for an 119873 times 119873 image is respectively119874(119873) 119874(1198732 times 119872

2) with 119872 being the width of Gabor

(Gaussian) mask filter119874(1198732) and119874(1198732 log(119873)) As a resultthe computational complexity of the combination of theDWT and Gabor filter is 119874(119873) + 119874(1198732 times 1198722) In terms offeatures extraction processing time the average time requiredto process a brain a mammogram and a retina imagewith the DWT approach (DWT-Gabor) was respectively019 (031) 017 (032) and 015 seconds (035) using MatlabR2009a on a 15 GHz Core2 Duo processor

6 Conclusion

We proposed a supervised system for biomedical imagesclassification that uses statistical features obtained from thecombination of the discrete wavelet transform and Gaborfilter to classify normal images versus cancer images usingsupport vector machines as classifiers Our experimentalresults show that such a hybrid processing model achieveshigher accuracy in comparison to using DWT or Gaborfilter banks alone Therefore the proposed image processingand features extraction approach seem to be very promis-ing for the detection of certain pathologies in biomedicalimages

For future works it is recommended to consider a largerset of features and a selection process to identify the mostdiscriminant ones In addition the Gabor parameters willbe adjusted for each type of image separately to improvethe accuracy Furthermore the DWT-Gabor will be directlycompared to the dual-tree complex wavelet curvelet andcontourlet using the same databases and images in orderto draw general conclusions Also multilabels classificationswill be considered in future works to investigate the discrim-inative power of our approach for each type of pathology

Journal of Medical Engineering 11

Table 5 Comparison with the literature

Features Classifier Accuracylowast

Mammograms[3] Gabor k-NN 80[4] DT-CWT SVM 8864[5] Contourlet SVM 966Our approach DWT-Gabor SVM 9667 (plusmn005)

Retina[1] DWT + GLCM LDA 822

[2] Morphological + GLCM Probabilistic boosting algorithm 813ndash922717ndash852

[39] Gabor SVM 83Our approach DWT-Gabor SVM 100

Brain[6] DWT SVM 98[8] DWT + PCA BPNN 100

SVM 90[38] Voxels Bayes 92

VFI 78Our approach DWT-Gabor SVM 9736 (plusmn002)

lowastCorrect classification rate

Finally more experiments on the effect of kernel choice andits parameter on classification accuracy will be investigated

References

[1] A Khademi and S Krishnan ldquoShift-invariant discrete wavelettransform analysis for retinal image classificationrdquoMedical andBiological Engineering and Computing vol 45 no 12 pp 1211ndash1222 2007

[2] N Lee A F Laine andT R Smith ldquoLearning non-homogenoustextures and the unlearning problemwith application to drusendetection in retinal imagesrdquo in Proceedings of the 5th IEEEInternational Symposium on Biomedical Imaging From Nano toMacro (ISBI rsquo08) pp 1215ndash1218 Paris France May 2008

[3] A Dong and B Wang ldquoFeature selection and analysis onmammogram classificationrdquo in Proceedings of the IEEE PacificRim Conference on Communications Computers and SignalProcessing (PACRIM rsquo09) pp 731ndash735 Victoria BC CanadaAugust 2009

[4] A Tirtajaya and D D Santika ldquoClassification of microcalcifi-cation using dual-tree complex wavelet transform and supportvector machinerdquo in Proceedings of the 2nd International Confer-ence onAdvances in Computing Control and TelecommunicationTechnologies (ACT rsquo10) pp 164ndash166 Jakarta Indonesia Decem-ber 2010

[5] F Moayedi Z Azimifar R Boostani and S Katebi ldquoContour-let-based mammography mass classification using the SVMfamilyrdquo Computers in Biology and Medicine vol 40 no 4 pp373ndash383 2010

[6] S Chaplot LM Patnaik andN R Jagannathan ldquoClassificationof magnetic resonance brain images using wavelets as input tosupport vector machine and neural networkrdquo Biomedical SignalProcessing and Control vol 1 no 1 pp 86ndash92 2006

[7] Y Zhang S Wang and L Wu ldquoA novel method for magneticresonance brain image classification based on adaptive chaoticPSOrdquo Progress in Electromagnetics Research vol 109 pp 325ndash343 2010

[8] Y Zhang Z Dong L Wu and S Wang ldquoA hybrid method forMRI brain image classificationrdquo Expert Systems with Applica-tions vol 38 no 8 pp 10049ndash10053 2011

[9] M E Celebi H Iyatomi G Schaefer and W V StoeckerldquoLesion border detection in dermoscopy imagesrdquoComputerizedMedical Imaging and Graphics vol 33 no 2 pp 148ndash153 2009

[10] Q AbbasM E Celebi and I F Garcıa ldquoSkin tumor area extrac-tion using an improved dynamic programming approachrdquo SkinResearch and Technology vol 18 pp 133ndash142 2012

[11] Q Li F Li and K Doi ldquoComputerized detection of lungnodules in thin-section CT images by use of selective enhance-ment filters and an automated rule-based classifierrdquo AcademicRadiology vol 15 no 2 pp 165ndash175 2008

[12] A El-Bazl M Nitzken E Vanbogaertl G Gimelrsquojarb RFalfi and M Abo El-Ghar ldquoA novel shaped-based diagnosticapproach for early diagnosis of lung nodulesrdquo in Proceedingsof the IEEE International Symposium in Biomedical Imaging(ISBI rsquo11) pp 137ndash140 Chicago Ill USA 2011

[13] M T Coimbra and J P S Cunha ldquoMPEG-7 visual descriptorsmdashcontributions for automated feature extraction in capsuleendoscopyrdquo IEEE Transactions on Circuits and Systems for VideoTechnology vol 16 no 5 pp 628ndash636 2006

[14] B Li and M Q H Meng ldquoTexture analysis for ulcer detectionin capsule endoscopy imagesrdquo Image andVision Computing vol27 no 9 pp 1336ndash1342 2009

[15] C K Chui An Introduction to Wavelets Academic Press SanDiego Calif USA 1992

[16] M Vetterli andCHerley ldquoWavelets and filter banks theory anddesignrdquo IEEE Transactions on Signal Processing vol 40 no 9pp 2207ndash2232 1992

12 Journal of Medical Engineering

[17] J G Daugman ldquoUncertainty relation for resolution inspace spatial frequency and orientation optimized by two-dimensional visual cortical filtersrdquo Journal of the Optical Societyof America A vol 2 no 7 pp 1160ndash1169 1985

[18] I W Selesnick R G Baraniuk and N G Kingsbury ldquoThedual-tree complex wavelet transformrdquo IEEE Signal ProcessingMagazine vol 22 no 6 pp 123ndash151 2005

[19] E Candes and D Donoho ldquoRidgelets a key to higher-dimensional intermittencyrdquo Philosophical Transactions of theLondon Royal Society vol 357 pp 2495ndash2509 1999

[20] E J Candes and D L Donoho ldquoContinuous curvelettransformmdashI Resolution of the wavefront setrdquo Applied andComputational Harmonic Analysis vol 19 no 2 pp 162ndash1972005

[21] M N Do and M Vetterli ldquoThe contourlet transform an effi-cient directional multiresolution image representationrdquo IEEETransactions on Image Processing vol 14 no 12 pp 2091ndash21062005

[22] R J Ferrari R M Rangayyan J E L Desautels and A FFrere ldquoAnalysis of asymmetry in mammograms via directionalfiltering with Gabor waveletsrdquo IEEE Transactions on MedicalImaging vol 20 no 9 pp 953ndash964 2001

[23] Z Cui and G Zhang ldquoA novel medical image dynamic fuzzyclassification model based on ridgelet transformrdquo Journal ofSoftware vol 5 no 5 pp 458ndash465 2010

[24] T Geback and P Koumoutsakos ldquoEdge detection inmicroscopyimages using curveletsrdquo BMC Bioinformatics vol 10 article 752009

[25] J Ma andG Plonka ldquoThe curvelet transform a review of recentapplicationsrdquo IEEE Signal ProcessingMagazine vol 27 no 2 pp118ndash133 2010

[26] N Kingsbury ldquoComplex wavelets and shift invariancerdquo inProceedings of the IEEE Seminar on Time-Scale and Time-Frequency Analysis and Applications pp 501ndash510 London UK2000

[27] Y L Qiao C Y Song and C H Zhao ldquoM-band ridgelettransform based texture classificationrdquo Pattern RecognitionLetters vol 31 no 3 pp 244ndash249 2010

[28] F Gomez and E Romero ldquoTexture characterization using acurvelet based descriptorrdquo Lecture Notes in Computer Sciencevol 5856 pp 113ndash120 2009

[29] H Shan and J Ma ldquoCurvelet-based geodesic snakes for imagesegmentation with multiple objectsrdquo Pattern Recognition Let-ters vol 31 no 5 pp 355ndash360 2010

[30] R Eslami and H Radha ldquoNew image transforms using hybridwavelets and directional filter banks analysis and designrdquo inProceedings of the IEEE International Conference on ImageProcessing (ICIP rsquo05) pp 733ndash736 Genova Italy September2005

[31] O O V Villegas and V G C Sanchez ldquoThe wavelet basedcontourlet transform and its application to feature preservingimage codingrdquo Lecture Notes in Computer Science vol 4827 pp590ndash600 2007

[32] S Lahmir and M Boukadoum ldquoClassification of brain MRIusing the LH and HL wavelet transform sub-bandsrdquo in Pro-ceedings of the IEEE International Symposium on Circuits andSystems (ISCAS rsquo11) pp 1025ndash1028 Rio de Janeiro Brazil May2009 2011

[33] S Lahmir and M Boukadoum ldquoBrain MRI classificationusing an ensemble system and LH and HL wavelet Sub-bandsFeaturesrdquo in Proceedings of the IEEE Symposium Series on

Computational Intelligence (SSCI rsquo11) pp 1ndash7 Paris FranceApril2011

[34] S Lahmir and M Boukadoum ldquoHybrid Cosine and RadonTransform-based processing for Digital Mammogram FeatureExtraction and Classification with SVMrdquo in Proceedings of the33rd IEEE Annual International Conference on Engineering inMedecine and Biology Society (EMBS rsquo11) pp 5104ndash5107 BostonMass USA 2011

[35] S Lahmir andM Boukadoum ldquoDWT and RT-Based Approachfor Feature Extraction and classification of Mammograms withSVMrdquo in Proceedings of the IEEE Biomedical Circuits andSystems Conference (BioCAS rsquo11) pp 412ndash415 San Diego CalifUSA November 2011

[36] VN VapnikTheNature of Statistical LearningTheory Springer1995

[37] S Lahmiri and M Boukadoum ldquoHybrid discret wavelet trans-form and Gabor filter banks processing for mammogramfeatures extractionrdquo in Proceedings of the IEEE New Circuits andSystems (NEWCAS rsquo11) pp 53ndash56 Bordeaux France June 2011

[38] L M Bruce and N Shanmugam ldquoUsing neural networkswith wavelet transforms for an automated mammographicmass classifierrdquo in Proceedings of the 22nd Annual InternationalConference of the IEEE Engineering in Medicine and BiologySociety pp 985ndash987 Chicago Ill USA July 2000

[39] S M H Jamarani G Rezai-rad and H Behnam ldquoA novelmethod for breast cancer prognosis using wavelet packet basedneural networkrdquo in Proceedings of the 27th Annual InternationalConference of the Engineering in Medicine and Biology Society(IEEE-EMBS rsquo05) pp 3414ndash3417 Shanghai China September2005

[40] httpmedical-dictionarythefreedictionarycomcircinate+retinopathy

[41] C I O Martins F N S Medeiros R M S Veras F N Bezerraand R M Cesar Jr ldquoEvaluation of retinal vessel segmentationmethods for microaneurysms detectionrdquo in Proceedings of theIEEE International Conference on Image Processing (ICIP rsquo09)pp 3365ndash3368 Cairo Egypt November 2009

[42] C Plant S J Teipel A Oswald et al ldquoAutomated detectionof brain atrophy patterns based on MRI for the prediction ofAlzheimerrsquos diseaserdquo NeuroImage vol 50 no 1 pp 162ndash1742010

[43] J Meier R Bock L G Nyul and G Michelson ldquoEye fun-dus image processing system for automated glaucoma clas-sificationrdquo in Proceedings of the 52nd Internationales Wis-senschaftliches Kolloquium Technische Universitat Ilmenau2007

[44] E Sakka A Prentza I E Lamprinos and D KoutsourisldquoMicrocalcification detection using multiresolution analysisbased on wavelet transformrdquo in Proceedings of the IEEE Inter-national Special Topic Conference on Information Technology inBiomedicine Ioannina Greece October 2006

[45] S G Mallat ldquoTheory for multiresolution signal decompositionthe wavelet representationrdquo IEEE Transactions on Pattern Anal-ysis and Machine Intelligence vol 11 no 7 pp 674ndash693 1989

[46] J Suckling J Parker and D R Dance ldquoThe mammographicimage analysis society digital mammogram databaserdquo in Pro-ceedings of the the 2nd International Workshop on DigitalMammography A G Gale S M Astley D D Dance and AY Cairns Eds pp 375ndash378 Elsevier York UK 1994

[47] H J Chiou C Y Chen T C Liu et al ldquoComputer-aideddiagnosis of peripheral soft tissue masses based on ultrasound

Journal of Medical Engineering 13

imagingrdquo Computerized Medical Imaging and Graphics vol 33no 5 pp 408ndash413 2009

[48] J K Kim J M Park K S Song and H W Park ldquoAdaptivemammographic image enhancement using first derivative andlocal statisticsrdquo IEEE Transactions on Medical Imaging vol 16no 5 pp 495ndash502 1997

[49] H S Sheshadri andAKandaswamy ldquoBreast tissue classificationusing statistical feature extraction of mammogramsrdquo MedicalImaging and Information Sciences vol 23 no 3 pp 105ndash1072006

[50] N Cristianini and J Shawe-Taylor Introduction to SupportVector Machines and Other Kernel-Based Learning MethodsCambridge University Press Cambridge UK 2000

[51] L Chen X Mao Y Xue and L L Cheng ldquoSpeech emotionrecognition features and classification modelsrdquo Digital SignalProcessing vol 22 pp 1154ndash1160 2012

[52] F Palmieri U Fiore A Castiglione and A De Santis ldquoOn thedetection of card-sharing traffic through wavelet analysis andSupport Vector Machinesrdquo Applied Soft Computing vol 13 no1 pp 615ndash627 2013

[53] A Azadeh M Saberi A Kazem V Ebrahimipour A Nour-mohammadzadeh and Z Saberi ldquoA flexible algorithm forfault diagnosis in a centrifugal pump with corrupted data andnoise based on ANN and support vector machine with hyper-parameters optimizationrdquoApplied Soft Computing vol 13 no 3pp 1478ndash1485 2013

[54] R J Martis U R Acharya K M Mandana A K Ray andC Chakraborty ldquoCardiac decision making using higher orderspectrardquo Biomedical Signal Processing and Control vol 8 pp193ndash203 2013

[55] M R Mohammad ldquoChi-square distance kernel of the gaitsfor the diagnosis of Parkinsonrsquos diseaserdquo Biomedical SignalProcessing and Control vol 8 pp 66ndash70 2013

[56] R Vandenberghe N Nelissen E Salmon et al ldquoBinary clas-sification of 18F-flutemetamol PET using machine learningcomparison with visual reads and structuralMRIrdquoNeuroImagevol 64 no 1 pp 517ndash525 2013

[57] httpmarathoncseeusfeduMammographyDatabasehtml[58] httpwwwcesclemsonedusimahooverstare[59] httpwwwmedharvardeduaanlib

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

10 Journal of Medical Engineering

Table 3 SVM classification sensitivity as a function of feature extraction method and level of DWT decompositionlowast

DWT DWT-Gabor DWT DWT-GaborDecomposition level One One Two TwoMammograms 9414 (plusmn006) 9333 (plusmn006) 9029 (plusmn0039) 8978 (plusmn004)Retina 100 100 100 100Brain MRI 9384 (plusmn000) 8251 (plusmn016) 100 5384 (plusmn023)lowastTenfold cross-validation used for mammograms and retina images leave-one-out used for brain MRIs

Table 4 SVM classification performance measures obtained withGabor-based features

Accuracy Specificity SensitivityMammograms 6803 (plusmn001) 100 0Retina 5000 (plusmn000) 100 0Brain MRI 8661 (plusmn003) 100 0

The obtained correct classification rate of mammogramsretina and brain magnetic resonance images is respectively6803 (plusmn001) 5000 (plusmn000) and 8661 (plusmn003) Theaverage results for classifier specificity and sensitivity for allimages are 100 and 0 This finding indicates that Gabor-based features are suitable to detect pathological imagesbut fails to detect normal images In sum the results showthat Gabor-based features do not perform better than DWTand DWT-Gabor-based features These findings confirm thesuperiority of combining theDWT andGabor filter banks forfeature extraction

Based on the previous results it appears that the DWT-Gabor approach for feature extraction is effective for detect-ing the abrupt changes in biological tissue that characterizethe pathological patterns that were investigated and it yieldsbetter classification accuracy and specificity than the DWT-only approach It also offers equal of better sensitivity exceptfor brain MRIs For brain MRIs the obtained specificityand sensitivity results with the DWT-Gabor approach showimproved true negative detection but lower true positiveperformance Finally the obtained results reveal also thatlevel-one DWT decomposition is preferable to level-twodecomposition

Finally Table 5 compares the results obtained with theDWT-Gabor approach to other work that we surveyed Inmany cases the DWT-Gabor method yields higher clas-sification rates particularly for mammograms and retinaFor the problem of brain MRI classification our obtainedperformance is better than the results of [38] but less thanwhat is reported in [6 8] However these comparisons shouldbe viewed with caution as not all the results stem froma common image database and the different authors usedifferent sample and image sizesMoreovermany authors useno cross-validation and simply performa single arbitrary splitof their data into training and test sets to obtain their accuracyresults Obviously one cannot generalize or draw definiteconclusion from such efforts and comparisons between

works cannot be made other than in general terms In thisrespect it can only be concluded from our results that theDWT-Gabor for feature extraction is effective for obtaininghigh image classification accuracy by an SVM and that itmay outperform other feature extraction and classificationtechniques reported in the literature at least those basedon DWT-only image decomposition Unfortunately a moredefinite conclusion is impossiblewithout gaining access to theimage databases used by the other authors

5 Computational Complexity

Finally the computational complexity of the DWT Gaborcontourlet and curvelet for an 119873 times 119873 image is respectively119874(119873) 119874(1198732 times 119872

2) with 119872 being the width of Gabor

(Gaussian) mask filter119874(1198732) and119874(1198732 log(119873)) As a resultthe computational complexity of the combination of theDWT and Gabor filter is 119874(119873) + 119874(1198732 times 1198722) In terms offeatures extraction processing time the average time requiredto process a brain a mammogram and a retina imagewith the DWT approach (DWT-Gabor) was respectively019 (031) 017 (032) and 015 seconds (035) using MatlabR2009a on a 15 GHz Core2 Duo processor

6 Conclusion

We proposed a supervised system for biomedical imagesclassification that uses statistical features obtained from thecombination of the discrete wavelet transform and Gaborfilter to classify normal images versus cancer images usingsupport vector machines as classifiers Our experimentalresults show that such a hybrid processing model achieveshigher accuracy in comparison to using DWT or Gaborfilter banks alone Therefore the proposed image processingand features extraction approach seem to be very promis-ing for the detection of certain pathologies in biomedicalimages

For future works it is recommended to consider a largerset of features and a selection process to identify the mostdiscriminant ones In addition the Gabor parameters willbe adjusted for each type of image separately to improvethe accuracy Furthermore the DWT-Gabor will be directlycompared to the dual-tree complex wavelet curvelet andcontourlet using the same databases and images in orderto draw general conclusions Also multilabels classificationswill be considered in future works to investigate the discrim-inative power of our approach for each type of pathology

Journal of Medical Engineering 11

Table 5 Comparison with the literature

Features Classifier Accuracylowast

Mammograms[3] Gabor k-NN 80[4] DT-CWT SVM 8864[5] Contourlet SVM 966Our approach DWT-Gabor SVM 9667 (plusmn005)

Retina[1] DWT + GLCM LDA 822

[2] Morphological + GLCM Probabilistic boosting algorithm 813ndash922717ndash852

[39] Gabor SVM 83Our approach DWT-Gabor SVM 100

Brain[6] DWT SVM 98[8] DWT + PCA BPNN 100

SVM 90[38] Voxels Bayes 92

VFI 78Our approach DWT-Gabor SVM 9736 (plusmn002)

lowastCorrect classification rate

Finally more experiments on the effect of kernel choice andits parameter on classification accuracy will be investigated

References

[1] A Khademi and S Krishnan ldquoShift-invariant discrete wavelettransform analysis for retinal image classificationrdquoMedical andBiological Engineering and Computing vol 45 no 12 pp 1211ndash1222 2007

[2] N Lee A F Laine andT R Smith ldquoLearning non-homogenoustextures and the unlearning problemwith application to drusendetection in retinal imagesrdquo in Proceedings of the 5th IEEEInternational Symposium on Biomedical Imaging From Nano toMacro (ISBI rsquo08) pp 1215ndash1218 Paris France May 2008

[3] A Dong and B Wang ldquoFeature selection and analysis onmammogram classificationrdquo in Proceedings of the IEEE PacificRim Conference on Communications Computers and SignalProcessing (PACRIM rsquo09) pp 731ndash735 Victoria BC CanadaAugust 2009

[4] A Tirtajaya and D D Santika ldquoClassification of microcalcifi-cation using dual-tree complex wavelet transform and supportvector machinerdquo in Proceedings of the 2nd International Confer-ence onAdvances in Computing Control and TelecommunicationTechnologies (ACT rsquo10) pp 164ndash166 Jakarta Indonesia Decem-ber 2010

[5] F Moayedi Z Azimifar R Boostani and S Katebi ldquoContour-let-based mammography mass classification using the SVMfamilyrdquo Computers in Biology and Medicine vol 40 no 4 pp373ndash383 2010

[6] S Chaplot LM Patnaik andN R Jagannathan ldquoClassificationof magnetic resonance brain images using wavelets as input tosupport vector machine and neural networkrdquo Biomedical SignalProcessing and Control vol 1 no 1 pp 86ndash92 2006

[7] Y Zhang S Wang and L Wu ldquoA novel method for magneticresonance brain image classification based on adaptive chaoticPSOrdquo Progress in Electromagnetics Research vol 109 pp 325ndash343 2010

[8] Y Zhang Z Dong L Wu and S Wang ldquoA hybrid method forMRI brain image classificationrdquo Expert Systems with Applica-tions vol 38 no 8 pp 10049ndash10053 2011

[9] M E Celebi H Iyatomi G Schaefer and W V StoeckerldquoLesion border detection in dermoscopy imagesrdquoComputerizedMedical Imaging and Graphics vol 33 no 2 pp 148ndash153 2009

[10] Q AbbasM E Celebi and I F Garcıa ldquoSkin tumor area extrac-tion using an improved dynamic programming approachrdquo SkinResearch and Technology vol 18 pp 133ndash142 2012

[11] Q Li F Li and K Doi ldquoComputerized detection of lungnodules in thin-section CT images by use of selective enhance-ment filters and an automated rule-based classifierrdquo AcademicRadiology vol 15 no 2 pp 165ndash175 2008

[12] A El-Bazl M Nitzken E Vanbogaertl G Gimelrsquojarb RFalfi and M Abo El-Ghar ldquoA novel shaped-based diagnosticapproach for early diagnosis of lung nodulesrdquo in Proceedingsof the IEEE International Symposium in Biomedical Imaging(ISBI rsquo11) pp 137ndash140 Chicago Ill USA 2011

[13] M T Coimbra and J P S Cunha ldquoMPEG-7 visual descriptorsmdashcontributions for automated feature extraction in capsuleendoscopyrdquo IEEE Transactions on Circuits and Systems for VideoTechnology vol 16 no 5 pp 628ndash636 2006

[14] B Li and M Q H Meng ldquoTexture analysis for ulcer detectionin capsule endoscopy imagesrdquo Image andVision Computing vol27 no 9 pp 1336ndash1342 2009

[15] C K Chui An Introduction to Wavelets Academic Press SanDiego Calif USA 1992

[16] M Vetterli andCHerley ldquoWavelets and filter banks theory anddesignrdquo IEEE Transactions on Signal Processing vol 40 no 9pp 2207ndash2232 1992

12 Journal of Medical Engineering

[17] J G Daugman ldquoUncertainty relation for resolution inspace spatial frequency and orientation optimized by two-dimensional visual cortical filtersrdquo Journal of the Optical Societyof America A vol 2 no 7 pp 1160ndash1169 1985

[18] I W Selesnick R G Baraniuk and N G Kingsbury ldquoThedual-tree complex wavelet transformrdquo IEEE Signal ProcessingMagazine vol 22 no 6 pp 123ndash151 2005

[19] E Candes and D Donoho ldquoRidgelets a key to higher-dimensional intermittencyrdquo Philosophical Transactions of theLondon Royal Society vol 357 pp 2495ndash2509 1999

[20] E J Candes and D L Donoho ldquoContinuous curvelettransformmdashI Resolution of the wavefront setrdquo Applied andComputational Harmonic Analysis vol 19 no 2 pp 162ndash1972005

[21] M N Do and M Vetterli ldquoThe contourlet transform an effi-cient directional multiresolution image representationrdquo IEEETransactions on Image Processing vol 14 no 12 pp 2091ndash21062005

[22] R J Ferrari R M Rangayyan J E L Desautels and A FFrere ldquoAnalysis of asymmetry in mammograms via directionalfiltering with Gabor waveletsrdquo IEEE Transactions on MedicalImaging vol 20 no 9 pp 953ndash964 2001

[23] Z Cui and G Zhang ldquoA novel medical image dynamic fuzzyclassification model based on ridgelet transformrdquo Journal ofSoftware vol 5 no 5 pp 458ndash465 2010

[24] T Geback and P Koumoutsakos ldquoEdge detection inmicroscopyimages using curveletsrdquo BMC Bioinformatics vol 10 article 752009

[25] J Ma andG Plonka ldquoThe curvelet transform a review of recentapplicationsrdquo IEEE Signal ProcessingMagazine vol 27 no 2 pp118ndash133 2010

[26] N Kingsbury ldquoComplex wavelets and shift invariancerdquo inProceedings of the IEEE Seminar on Time-Scale and Time-Frequency Analysis and Applications pp 501ndash510 London UK2000

[27] Y L Qiao C Y Song and C H Zhao ldquoM-band ridgelettransform based texture classificationrdquo Pattern RecognitionLetters vol 31 no 3 pp 244ndash249 2010

[28] F Gomez and E Romero ldquoTexture characterization using acurvelet based descriptorrdquo Lecture Notes in Computer Sciencevol 5856 pp 113ndash120 2009

[29] H Shan and J Ma ldquoCurvelet-based geodesic snakes for imagesegmentation with multiple objectsrdquo Pattern Recognition Let-ters vol 31 no 5 pp 355ndash360 2010

[30] R Eslami and H Radha ldquoNew image transforms using hybridwavelets and directional filter banks analysis and designrdquo inProceedings of the IEEE International Conference on ImageProcessing (ICIP rsquo05) pp 733ndash736 Genova Italy September2005

[31] O O V Villegas and V G C Sanchez ldquoThe wavelet basedcontourlet transform and its application to feature preservingimage codingrdquo Lecture Notes in Computer Science vol 4827 pp590ndash600 2007

[32] S Lahmir and M Boukadoum ldquoClassification of brain MRIusing the LH and HL wavelet transform sub-bandsrdquo in Pro-ceedings of the IEEE International Symposium on Circuits andSystems (ISCAS rsquo11) pp 1025ndash1028 Rio de Janeiro Brazil May2009 2011

[33] S Lahmir and M Boukadoum ldquoBrain MRI classificationusing an ensemble system and LH and HL wavelet Sub-bandsFeaturesrdquo in Proceedings of the IEEE Symposium Series on

Computational Intelligence (SSCI rsquo11) pp 1ndash7 Paris FranceApril2011

[34] S Lahmir and M Boukadoum ldquoHybrid Cosine and RadonTransform-based processing for Digital Mammogram FeatureExtraction and Classification with SVMrdquo in Proceedings of the33rd IEEE Annual International Conference on Engineering inMedecine and Biology Society (EMBS rsquo11) pp 5104ndash5107 BostonMass USA 2011

[35] S Lahmir andM Boukadoum ldquoDWT and RT-Based Approachfor Feature Extraction and classification of Mammograms withSVMrdquo in Proceedings of the IEEE Biomedical Circuits andSystems Conference (BioCAS rsquo11) pp 412ndash415 San Diego CalifUSA November 2011

[36] VN VapnikTheNature of Statistical LearningTheory Springer1995

[37] S Lahmiri and M Boukadoum ldquoHybrid discret wavelet trans-form and Gabor filter banks processing for mammogramfeatures extractionrdquo in Proceedings of the IEEE New Circuits andSystems (NEWCAS rsquo11) pp 53ndash56 Bordeaux France June 2011

[38] L M Bruce and N Shanmugam ldquoUsing neural networkswith wavelet transforms for an automated mammographicmass classifierrdquo in Proceedings of the 22nd Annual InternationalConference of the IEEE Engineering in Medicine and BiologySociety pp 985ndash987 Chicago Ill USA July 2000

[39] S M H Jamarani G Rezai-rad and H Behnam ldquoA novelmethod for breast cancer prognosis using wavelet packet basedneural networkrdquo in Proceedings of the 27th Annual InternationalConference of the Engineering in Medicine and Biology Society(IEEE-EMBS rsquo05) pp 3414ndash3417 Shanghai China September2005

[40] httpmedical-dictionarythefreedictionarycomcircinate+retinopathy

[41] C I O Martins F N S Medeiros R M S Veras F N Bezerraand R M Cesar Jr ldquoEvaluation of retinal vessel segmentationmethods for microaneurysms detectionrdquo in Proceedings of theIEEE International Conference on Image Processing (ICIP rsquo09)pp 3365ndash3368 Cairo Egypt November 2009

[42] C Plant S J Teipel A Oswald et al ldquoAutomated detectionof brain atrophy patterns based on MRI for the prediction ofAlzheimerrsquos diseaserdquo NeuroImage vol 50 no 1 pp 162ndash1742010

[43] J Meier R Bock L G Nyul and G Michelson ldquoEye fun-dus image processing system for automated glaucoma clas-sificationrdquo in Proceedings of the 52nd Internationales Wis-senschaftliches Kolloquium Technische Universitat Ilmenau2007

[44] E Sakka A Prentza I E Lamprinos and D KoutsourisldquoMicrocalcification detection using multiresolution analysisbased on wavelet transformrdquo in Proceedings of the IEEE Inter-national Special Topic Conference on Information Technology inBiomedicine Ioannina Greece October 2006

[45] S G Mallat ldquoTheory for multiresolution signal decompositionthe wavelet representationrdquo IEEE Transactions on Pattern Anal-ysis and Machine Intelligence vol 11 no 7 pp 674ndash693 1989

[46] J Suckling J Parker and D R Dance ldquoThe mammographicimage analysis society digital mammogram databaserdquo in Pro-ceedings of the the 2nd International Workshop on DigitalMammography A G Gale S M Astley D D Dance and AY Cairns Eds pp 375ndash378 Elsevier York UK 1994

[47] H J Chiou C Y Chen T C Liu et al ldquoComputer-aideddiagnosis of peripheral soft tissue masses based on ultrasound

Journal of Medical Engineering 13

imagingrdquo Computerized Medical Imaging and Graphics vol 33no 5 pp 408ndash413 2009

[48] J K Kim J M Park K S Song and H W Park ldquoAdaptivemammographic image enhancement using first derivative andlocal statisticsrdquo IEEE Transactions on Medical Imaging vol 16no 5 pp 495ndash502 1997

[49] H S Sheshadri andAKandaswamy ldquoBreast tissue classificationusing statistical feature extraction of mammogramsrdquo MedicalImaging and Information Sciences vol 23 no 3 pp 105ndash1072006

[50] N Cristianini and J Shawe-Taylor Introduction to SupportVector Machines and Other Kernel-Based Learning MethodsCambridge University Press Cambridge UK 2000

[51] L Chen X Mao Y Xue and L L Cheng ldquoSpeech emotionrecognition features and classification modelsrdquo Digital SignalProcessing vol 22 pp 1154ndash1160 2012

[52] F Palmieri U Fiore A Castiglione and A De Santis ldquoOn thedetection of card-sharing traffic through wavelet analysis andSupport Vector Machinesrdquo Applied Soft Computing vol 13 no1 pp 615ndash627 2013

[53] A Azadeh M Saberi A Kazem V Ebrahimipour A Nour-mohammadzadeh and Z Saberi ldquoA flexible algorithm forfault diagnosis in a centrifugal pump with corrupted data andnoise based on ANN and support vector machine with hyper-parameters optimizationrdquoApplied Soft Computing vol 13 no 3pp 1478ndash1485 2013

[54] R J Martis U R Acharya K M Mandana A K Ray andC Chakraborty ldquoCardiac decision making using higher orderspectrardquo Biomedical Signal Processing and Control vol 8 pp193ndash203 2013

[55] M R Mohammad ldquoChi-square distance kernel of the gaitsfor the diagnosis of Parkinsonrsquos diseaserdquo Biomedical SignalProcessing and Control vol 8 pp 66ndash70 2013

[56] R Vandenberghe N Nelissen E Salmon et al ldquoBinary clas-sification of 18F-flutemetamol PET using machine learningcomparison with visual reads and structuralMRIrdquoNeuroImagevol 64 no 1 pp 517ndash525 2013

[57] httpmarathoncseeusfeduMammographyDatabasehtml[58] httpwwwcesclemsonedusimahooverstare[59] httpwwwmedharvardeduaanlib

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Journal of Medical Engineering 11

Table 5 Comparison with the literature

Features Classifier Accuracylowast

Mammograms[3] Gabor k-NN 80[4] DT-CWT SVM 8864[5] Contourlet SVM 966Our approach DWT-Gabor SVM 9667 (plusmn005)

Retina[1] DWT + GLCM LDA 822

[2] Morphological + GLCM Probabilistic boosting algorithm 813ndash922717ndash852

[39] Gabor SVM 83Our approach DWT-Gabor SVM 100

Brain[6] DWT SVM 98[8] DWT + PCA BPNN 100

SVM 90[38] Voxels Bayes 92

VFI 78Our approach DWT-Gabor SVM 9736 (plusmn002)

lowastCorrect classification rate

Finally more experiments on the effect of kernel choice andits parameter on classification accuracy will be investigated

References

[1] A Khademi and S Krishnan ldquoShift-invariant discrete wavelettransform analysis for retinal image classificationrdquoMedical andBiological Engineering and Computing vol 45 no 12 pp 1211ndash1222 2007

[2] N Lee A F Laine andT R Smith ldquoLearning non-homogenoustextures and the unlearning problemwith application to drusendetection in retinal imagesrdquo in Proceedings of the 5th IEEEInternational Symposium on Biomedical Imaging From Nano toMacro (ISBI rsquo08) pp 1215ndash1218 Paris France May 2008

[3] A Dong and B Wang ldquoFeature selection and analysis onmammogram classificationrdquo in Proceedings of the IEEE PacificRim Conference on Communications Computers and SignalProcessing (PACRIM rsquo09) pp 731ndash735 Victoria BC CanadaAugust 2009

[4] A Tirtajaya and D D Santika ldquoClassification of microcalcifi-cation using dual-tree complex wavelet transform and supportvector machinerdquo in Proceedings of the 2nd International Confer-ence onAdvances in Computing Control and TelecommunicationTechnologies (ACT rsquo10) pp 164ndash166 Jakarta Indonesia Decem-ber 2010

[5] F Moayedi Z Azimifar R Boostani and S Katebi ldquoContour-let-based mammography mass classification using the SVMfamilyrdquo Computers in Biology and Medicine vol 40 no 4 pp373ndash383 2010

[6] S Chaplot LM Patnaik andN R Jagannathan ldquoClassificationof magnetic resonance brain images using wavelets as input tosupport vector machine and neural networkrdquo Biomedical SignalProcessing and Control vol 1 no 1 pp 86ndash92 2006

[7] Y Zhang S Wang and L Wu ldquoA novel method for magneticresonance brain image classification based on adaptive chaoticPSOrdquo Progress in Electromagnetics Research vol 109 pp 325ndash343 2010

[8] Y Zhang Z Dong L Wu and S Wang ldquoA hybrid method forMRI brain image classificationrdquo Expert Systems with Applica-tions vol 38 no 8 pp 10049ndash10053 2011

[9] M E Celebi H Iyatomi G Schaefer and W V StoeckerldquoLesion border detection in dermoscopy imagesrdquoComputerizedMedical Imaging and Graphics vol 33 no 2 pp 148ndash153 2009

[10] Q AbbasM E Celebi and I F Garcıa ldquoSkin tumor area extrac-tion using an improved dynamic programming approachrdquo SkinResearch and Technology vol 18 pp 133ndash142 2012

[11] Q Li F Li and K Doi ldquoComputerized detection of lungnodules in thin-section CT images by use of selective enhance-ment filters and an automated rule-based classifierrdquo AcademicRadiology vol 15 no 2 pp 165ndash175 2008

[12] A El-Bazl M Nitzken E Vanbogaertl G Gimelrsquojarb RFalfi and M Abo El-Ghar ldquoA novel shaped-based diagnosticapproach for early diagnosis of lung nodulesrdquo in Proceedingsof the IEEE International Symposium in Biomedical Imaging(ISBI rsquo11) pp 137ndash140 Chicago Ill USA 2011

[13] M T Coimbra and J P S Cunha ldquoMPEG-7 visual descriptorsmdashcontributions for automated feature extraction in capsuleendoscopyrdquo IEEE Transactions on Circuits and Systems for VideoTechnology vol 16 no 5 pp 628ndash636 2006

[14] B Li and M Q H Meng ldquoTexture analysis for ulcer detectionin capsule endoscopy imagesrdquo Image andVision Computing vol27 no 9 pp 1336ndash1342 2009

[15] C K Chui An Introduction to Wavelets Academic Press SanDiego Calif USA 1992

[16] M Vetterli andCHerley ldquoWavelets and filter banks theory anddesignrdquo IEEE Transactions on Signal Processing vol 40 no 9pp 2207ndash2232 1992

12 Journal of Medical Engineering

[17] J G Daugman ldquoUncertainty relation for resolution inspace spatial frequency and orientation optimized by two-dimensional visual cortical filtersrdquo Journal of the Optical Societyof America A vol 2 no 7 pp 1160ndash1169 1985

[18] I W Selesnick R G Baraniuk and N G Kingsbury ldquoThedual-tree complex wavelet transformrdquo IEEE Signal ProcessingMagazine vol 22 no 6 pp 123ndash151 2005

[19] E Candes and D Donoho ldquoRidgelets a key to higher-dimensional intermittencyrdquo Philosophical Transactions of theLondon Royal Society vol 357 pp 2495ndash2509 1999

[20] E J Candes and D L Donoho ldquoContinuous curvelettransformmdashI Resolution of the wavefront setrdquo Applied andComputational Harmonic Analysis vol 19 no 2 pp 162ndash1972005

[21] M N Do and M Vetterli ldquoThe contourlet transform an effi-cient directional multiresolution image representationrdquo IEEETransactions on Image Processing vol 14 no 12 pp 2091ndash21062005

[22] R J Ferrari R M Rangayyan J E L Desautels and A FFrere ldquoAnalysis of asymmetry in mammograms via directionalfiltering with Gabor waveletsrdquo IEEE Transactions on MedicalImaging vol 20 no 9 pp 953ndash964 2001

[23] Z Cui and G Zhang ldquoA novel medical image dynamic fuzzyclassification model based on ridgelet transformrdquo Journal ofSoftware vol 5 no 5 pp 458ndash465 2010

[24] T Geback and P Koumoutsakos ldquoEdge detection inmicroscopyimages using curveletsrdquo BMC Bioinformatics vol 10 article 752009

[25] J Ma andG Plonka ldquoThe curvelet transform a review of recentapplicationsrdquo IEEE Signal ProcessingMagazine vol 27 no 2 pp118ndash133 2010

[26] N Kingsbury ldquoComplex wavelets and shift invariancerdquo inProceedings of the IEEE Seminar on Time-Scale and Time-Frequency Analysis and Applications pp 501ndash510 London UK2000

[27] Y L Qiao C Y Song and C H Zhao ldquoM-band ridgelettransform based texture classificationrdquo Pattern RecognitionLetters vol 31 no 3 pp 244ndash249 2010

[28] F Gomez and E Romero ldquoTexture characterization using acurvelet based descriptorrdquo Lecture Notes in Computer Sciencevol 5856 pp 113ndash120 2009

[29] H Shan and J Ma ldquoCurvelet-based geodesic snakes for imagesegmentation with multiple objectsrdquo Pattern Recognition Let-ters vol 31 no 5 pp 355ndash360 2010

[30] R Eslami and H Radha ldquoNew image transforms using hybridwavelets and directional filter banks analysis and designrdquo inProceedings of the IEEE International Conference on ImageProcessing (ICIP rsquo05) pp 733ndash736 Genova Italy September2005

[31] O O V Villegas and V G C Sanchez ldquoThe wavelet basedcontourlet transform and its application to feature preservingimage codingrdquo Lecture Notes in Computer Science vol 4827 pp590ndash600 2007

[32] S Lahmir and M Boukadoum ldquoClassification of brain MRIusing the LH and HL wavelet transform sub-bandsrdquo in Pro-ceedings of the IEEE International Symposium on Circuits andSystems (ISCAS rsquo11) pp 1025ndash1028 Rio de Janeiro Brazil May2009 2011

[33] S Lahmir and M Boukadoum ldquoBrain MRI classificationusing an ensemble system and LH and HL wavelet Sub-bandsFeaturesrdquo in Proceedings of the IEEE Symposium Series on

Computational Intelligence (SSCI rsquo11) pp 1ndash7 Paris FranceApril2011

[34] S Lahmir and M Boukadoum ldquoHybrid Cosine and RadonTransform-based processing for Digital Mammogram FeatureExtraction and Classification with SVMrdquo in Proceedings of the33rd IEEE Annual International Conference on Engineering inMedecine and Biology Society (EMBS rsquo11) pp 5104ndash5107 BostonMass USA 2011

[35] S Lahmir andM Boukadoum ldquoDWT and RT-Based Approachfor Feature Extraction and classification of Mammograms withSVMrdquo in Proceedings of the IEEE Biomedical Circuits andSystems Conference (BioCAS rsquo11) pp 412ndash415 San Diego CalifUSA November 2011

[36] VN VapnikTheNature of Statistical LearningTheory Springer1995

[37] S Lahmiri and M Boukadoum ldquoHybrid discret wavelet trans-form and Gabor filter banks processing for mammogramfeatures extractionrdquo in Proceedings of the IEEE New Circuits andSystems (NEWCAS rsquo11) pp 53ndash56 Bordeaux France June 2011

[38] L M Bruce and N Shanmugam ldquoUsing neural networkswith wavelet transforms for an automated mammographicmass classifierrdquo in Proceedings of the 22nd Annual InternationalConference of the IEEE Engineering in Medicine and BiologySociety pp 985ndash987 Chicago Ill USA July 2000

[39] S M H Jamarani G Rezai-rad and H Behnam ldquoA novelmethod for breast cancer prognosis using wavelet packet basedneural networkrdquo in Proceedings of the 27th Annual InternationalConference of the Engineering in Medicine and Biology Society(IEEE-EMBS rsquo05) pp 3414ndash3417 Shanghai China September2005

[40] httpmedical-dictionarythefreedictionarycomcircinate+retinopathy

[41] C I O Martins F N S Medeiros R M S Veras F N Bezerraand R M Cesar Jr ldquoEvaluation of retinal vessel segmentationmethods for microaneurysms detectionrdquo in Proceedings of theIEEE International Conference on Image Processing (ICIP rsquo09)pp 3365ndash3368 Cairo Egypt November 2009

[42] C Plant S J Teipel A Oswald et al ldquoAutomated detectionof brain atrophy patterns based on MRI for the prediction ofAlzheimerrsquos diseaserdquo NeuroImage vol 50 no 1 pp 162ndash1742010

[43] J Meier R Bock L G Nyul and G Michelson ldquoEye fun-dus image processing system for automated glaucoma clas-sificationrdquo in Proceedings of the 52nd Internationales Wis-senschaftliches Kolloquium Technische Universitat Ilmenau2007

[44] E Sakka A Prentza I E Lamprinos and D KoutsourisldquoMicrocalcification detection using multiresolution analysisbased on wavelet transformrdquo in Proceedings of the IEEE Inter-national Special Topic Conference on Information Technology inBiomedicine Ioannina Greece October 2006

[45] S G Mallat ldquoTheory for multiresolution signal decompositionthe wavelet representationrdquo IEEE Transactions on Pattern Anal-ysis and Machine Intelligence vol 11 no 7 pp 674ndash693 1989

[46] J Suckling J Parker and D R Dance ldquoThe mammographicimage analysis society digital mammogram databaserdquo in Pro-ceedings of the the 2nd International Workshop on DigitalMammography A G Gale S M Astley D D Dance and AY Cairns Eds pp 375ndash378 Elsevier York UK 1994

[47] H J Chiou C Y Chen T C Liu et al ldquoComputer-aideddiagnosis of peripheral soft tissue masses based on ultrasound

Journal of Medical Engineering 13

imagingrdquo Computerized Medical Imaging and Graphics vol 33no 5 pp 408ndash413 2009

[48] J K Kim J M Park K S Song and H W Park ldquoAdaptivemammographic image enhancement using first derivative andlocal statisticsrdquo IEEE Transactions on Medical Imaging vol 16no 5 pp 495ndash502 1997

[49] H S Sheshadri andAKandaswamy ldquoBreast tissue classificationusing statistical feature extraction of mammogramsrdquo MedicalImaging and Information Sciences vol 23 no 3 pp 105ndash1072006

[50] N Cristianini and J Shawe-Taylor Introduction to SupportVector Machines and Other Kernel-Based Learning MethodsCambridge University Press Cambridge UK 2000

[51] L Chen X Mao Y Xue and L L Cheng ldquoSpeech emotionrecognition features and classification modelsrdquo Digital SignalProcessing vol 22 pp 1154ndash1160 2012

[52] F Palmieri U Fiore A Castiglione and A De Santis ldquoOn thedetection of card-sharing traffic through wavelet analysis andSupport Vector Machinesrdquo Applied Soft Computing vol 13 no1 pp 615ndash627 2013

[53] A Azadeh M Saberi A Kazem V Ebrahimipour A Nour-mohammadzadeh and Z Saberi ldquoA flexible algorithm forfault diagnosis in a centrifugal pump with corrupted data andnoise based on ANN and support vector machine with hyper-parameters optimizationrdquoApplied Soft Computing vol 13 no 3pp 1478ndash1485 2013

[54] R J Martis U R Acharya K M Mandana A K Ray andC Chakraborty ldquoCardiac decision making using higher orderspectrardquo Biomedical Signal Processing and Control vol 8 pp193ndash203 2013

[55] M R Mohammad ldquoChi-square distance kernel of the gaitsfor the diagnosis of Parkinsonrsquos diseaserdquo Biomedical SignalProcessing and Control vol 8 pp 66ndash70 2013

[56] R Vandenberghe N Nelissen E Salmon et al ldquoBinary clas-sification of 18F-flutemetamol PET using machine learningcomparison with visual reads and structuralMRIrdquoNeuroImagevol 64 no 1 pp 517ndash525 2013

[57] httpmarathoncseeusfeduMammographyDatabasehtml[58] httpwwwcesclemsonedusimahooverstare[59] httpwwwmedharvardeduaanlib

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

12 Journal of Medical Engineering

[17] J G Daugman ldquoUncertainty relation for resolution inspace spatial frequency and orientation optimized by two-dimensional visual cortical filtersrdquo Journal of the Optical Societyof America A vol 2 no 7 pp 1160ndash1169 1985

[18] I W Selesnick R G Baraniuk and N G Kingsbury ldquoThedual-tree complex wavelet transformrdquo IEEE Signal ProcessingMagazine vol 22 no 6 pp 123ndash151 2005

[19] E Candes and D Donoho ldquoRidgelets a key to higher-dimensional intermittencyrdquo Philosophical Transactions of theLondon Royal Society vol 357 pp 2495ndash2509 1999

[20] E J Candes and D L Donoho ldquoContinuous curvelettransformmdashI Resolution of the wavefront setrdquo Applied andComputational Harmonic Analysis vol 19 no 2 pp 162ndash1972005

[21] M N Do and M Vetterli ldquoThe contourlet transform an effi-cient directional multiresolution image representationrdquo IEEETransactions on Image Processing vol 14 no 12 pp 2091ndash21062005

[22] R J Ferrari R M Rangayyan J E L Desautels and A FFrere ldquoAnalysis of asymmetry in mammograms via directionalfiltering with Gabor waveletsrdquo IEEE Transactions on MedicalImaging vol 20 no 9 pp 953ndash964 2001

[23] Z Cui and G Zhang ldquoA novel medical image dynamic fuzzyclassification model based on ridgelet transformrdquo Journal ofSoftware vol 5 no 5 pp 458ndash465 2010

[24] T Geback and P Koumoutsakos ldquoEdge detection inmicroscopyimages using curveletsrdquo BMC Bioinformatics vol 10 article 752009

[25] J Ma andG Plonka ldquoThe curvelet transform a review of recentapplicationsrdquo IEEE Signal ProcessingMagazine vol 27 no 2 pp118ndash133 2010

[26] N Kingsbury ldquoComplex wavelets and shift invariancerdquo inProceedings of the IEEE Seminar on Time-Scale and Time-Frequency Analysis and Applications pp 501ndash510 London UK2000

[27] Y L Qiao C Y Song and C H Zhao ldquoM-band ridgelettransform based texture classificationrdquo Pattern RecognitionLetters vol 31 no 3 pp 244ndash249 2010

[28] F Gomez and E Romero ldquoTexture characterization using acurvelet based descriptorrdquo Lecture Notes in Computer Sciencevol 5856 pp 113ndash120 2009

[29] H Shan and J Ma ldquoCurvelet-based geodesic snakes for imagesegmentation with multiple objectsrdquo Pattern Recognition Let-ters vol 31 no 5 pp 355ndash360 2010

[30] R Eslami and H Radha ldquoNew image transforms using hybridwavelets and directional filter banks analysis and designrdquo inProceedings of the IEEE International Conference on ImageProcessing (ICIP rsquo05) pp 733ndash736 Genova Italy September2005

[31] O O V Villegas and V G C Sanchez ldquoThe wavelet basedcontourlet transform and its application to feature preservingimage codingrdquo Lecture Notes in Computer Science vol 4827 pp590ndash600 2007

[32] S Lahmir and M Boukadoum ldquoClassification of brain MRIusing the LH and HL wavelet transform sub-bandsrdquo in Pro-ceedings of the IEEE International Symposium on Circuits andSystems (ISCAS rsquo11) pp 1025ndash1028 Rio de Janeiro Brazil May2009 2011

[33] S Lahmir and M Boukadoum ldquoBrain MRI classificationusing an ensemble system and LH and HL wavelet Sub-bandsFeaturesrdquo in Proceedings of the IEEE Symposium Series on

Computational Intelligence (SSCI rsquo11) pp 1ndash7 Paris FranceApril2011

[34] S Lahmir and M Boukadoum ldquoHybrid Cosine and RadonTransform-based processing for Digital Mammogram FeatureExtraction and Classification with SVMrdquo in Proceedings of the33rd IEEE Annual International Conference on Engineering inMedecine and Biology Society (EMBS rsquo11) pp 5104ndash5107 BostonMass USA 2011

[35] S Lahmir andM Boukadoum ldquoDWT and RT-Based Approachfor Feature Extraction and classification of Mammograms withSVMrdquo in Proceedings of the IEEE Biomedical Circuits andSystems Conference (BioCAS rsquo11) pp 412ndash415 San Diego CalifUSA November 2011

[36] VN VapnikTheNature of Statistical LearningTheory Springer1995

[37] S Lahmiri and M Boukadoum ldquoHybrid discret wavelet trans-form and Gabor filter banks processing for mammogramfeatures extractionrdquo in Proceedings of the IEEE New Circuits andSystems (NEWCAS rsquo11) pp 53ndash56 Bordeaux France June 2011

[38] L M Bruce and N Shanmugam ldquoUsing neural networkswith wavelet transforms for an automated mammographicmass classifierrdquo in Proceedings of the 22nd Annual InternationalConference of the IEEE Engineering in Medicine and BiologySociety pp 985ndash987 Chicago Ill USA July 2000

[39] S M H Jamarani G Rezai-rad and H Behnam ldquoA novelmethod for breast cancer prognosis using wavelet packet basedneural networkrdquo in Proceedings of the 27th Annual InternationalConference of the Engineering in Medicine and Biology Society(IEEE-EMBS rsquo05) pp 3414ndash3417 Shanghai China September2005

[40] httpmedical-dictionarythefreedictionarycomcircinate+retinopathy

[41] C I O Martins F N S Medeiros R M S Veras F N Bezerraand R M Cesar Jr ldquoEvaluation of retinal vessel segmentationmethods for microaneurysms detectionrdquo in Proceedings of theIEEE International Conference on Image Processing (ICIP rsquo09)pp 3365ndash3368 Cairo Egypt November 2009

[42] C Plant S J Teipel A Oswald et al ldquoAutomated detectionof brain atrophy patterns based on MRI for the prediction ofAlzheimerrsquos diseaserdquo NeuroImage vol 50 no 1 pp 162ndash1742010

[43] J Meier R Bock L G Nyul and G Michelson ldquoEye fun-dus image processing system for automated glaucoma clas-sificationrdquo in Proceedings of the 52nd Internationales Wis-senschaftliches Kolloquium Technische Universitat Ilmenau2007

[44] E Sakka A Prentza I E Lamprinos and D KoutsourisldquoMicrocalcification detection using multiresolution analysisbased on wavelet transformrdquo in Proceedings of the IEEE Inter-national Special Topic Conference on Information Technology inBiomedicine Ioannina Greece October 2006

[45] S G Mallat ldquoTheory for multiresolution signal decompositionthe wavelet representationrdquo IEEE Transactions on Pattern Anal-ysis and Machine Intelligence vol 11 no 7 pp 674ndash693 1989

[46] J Suckling J Parker and D R Dance ldquoThe mammographicimage analysis society digital mammogram databaserdquo in Pro-ceedings of the the 2nd International Workshop on DigitalMammography A G Gale S M Astley D D Dance and AY Cairns Eds pp 375ndash378 Elsevier York UK 1994

[47] H J Chiou C Y Chen T C Liu et al ldquoComputer-aideddiagnosis of peripheral soft tissue masses based on ultrasound

Journal of Medical Engineering 13

imagingrdquo Computerized Medical Imaging and Graphics vol 33no 5 pp 408ndash413 2009

[48] J K Kim J M Park K S Song and H W Park ldquoAdaptivemammographic image enhancement using first derivative andlocal statisticsrdquo IEEE Transactions on Medical Imaging vol 16no 5 pp 495ndash502 1997

[49] H S Sheshadri andAKandaswamy ldquoBreast tissue classificationusing statistical feature extraction of mammogramsrdquo MedicalImaging and Information Sciences vol 23 no 3 pp 105ndash1072006

[50] N Cristianini and J Shawe-Taylor Introduction to SupportVector Machines and Other Kernel-Based Learning MethodsCambridge University Press Cambridge UK 2000

[51] L Chen X Mao Y Xue and L L Cheng ldquoSpeech emotionrecognition features and classification modelsrdquo Digital SignalProcessing vol 22 pp 1154ndash1160 2012

[52] F Palmieri U Fiore A Castiglione and A De Santis ldquoOn thedetection of card-sharing traffic through wavelet analysis andSupport Vector Machinesrdquo Applied Soft Computing vol 13 no1 pp 615ndash627 2013

[53] A Azadeh M Saberi A Kazem V Ebrahimipour A Nour-mohammadzadeh and Z Saberi ldquoA flexible algorithm forfault diagnosis in a centrifugal pump with corrupted data andnoise based on ANN and support vector machine with hyper-parameters optimizationrdquoApplied Soft Computing vol 13 no 3pp 1478ndash1485 2013

[54] R J Martis U R Acharya K M Mandana A K Ray andC Chakraborty ldquoCardiac decision making using higher orderspectrardquo Biomedical Signal Processing and Control vol 8 pp193ndash203 2013

[55] M R Mohammad ldquoChi-square distance kernel of the gaitsfor the diagnosis of Parkinsonrsquos diseaserdquo Biomedical SignalProcessing and Control vol 8 pp 66ndash70 2013

[56] R Vandenberghe N Nelissen E Salmon et al ldquoBinary clas-sification of 18F-flutemetamol PET using machine learningcomparison with visual reads and structuralMRIrdquoNeuroImagevol 64 no 1 pp 517ndash525 2013

[57] httpmarathoncseeusfeduMammographyDatabasehtml[58] httpwwwcesclemsonedusimahooverstare[59] httpwwwmedharvardeduaanlib

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

Journal of Medical Engineering 13

imagingrdquo Computerized Medical Imaging and Graphics vol 33no 5 pp 408ndash413 2009

[48] J K Kim J M Park K S Song and H W Park ldquoAdaptivemammographic image enhancement using first derivative andlocal statisticsrdquo IEEE Transactions on Medical Imaging vol 16no 5 pp 495ndash502 1997

[49] H S Sheshadri andAKandaswamy ldquoBreast tissue classificationusing statistical feature extraction of mammogramsrdquo MedicalImaging and Information Sciences vol 23 no 3 pp 105ndash1072006

[50] N Cristianini and J Shawe-Taylor Introduction to SupportVector Machines and Other Kernel-Based Learning MethodsCambridge University Press Cambridge UK 2000

[51] L Chen X Mao Y Xue and L L Cheng ldquoSpeech emotionrecognition features and classification modelsrdquo Digital SignalProcessing vol 22 pp 1154ndash1160 2012

[52] F Palmieri U Fiore A Castiglione and A De Santis ldquoOn thedetection of card-sharing traffic through wavelet analysis andSupport Vector Machinesrdquo Applied Soft Computing vol 13 no1 pp 615ndash627 2013

[53] A Azadeh M Saberi A Kazem V Ebrahimipour A Nour-mohammadzadeh and Z Saberi ldquoA flexible algorithm forfault diagnosis in a centrifugal pump with corrupted data andnoise based on ANN and support vector machine with hyper-parameters optimizationrdquoApplied Soft Computing vol 13 no 3pp 1478ndash1485 2013

[54] R J Martis U R Acharya K M Mandana A K Ray andC Chakraborty ldquoCardiac decision making using higher orderspectrardquo Biomedical Signal Processing and Control vol 8 pp193ndash203 2013

[55] M R Mohammad ldquoChi-square distance kernel of the gaitsfor the diagnosis of Parkinsonrsquos diseaserdquo Biomedical SignalProcessing and Control vol 8 pp 66ndash70 2013

[56] R Vandenberghe N Nelissen E Salmon et al ldquoBinary clas-sification of 18F-flutemetamol PET using machine learningcomparison with visual reads and structuralMRIrdquoNeuroImagevol 64 no 1 pp 517ndash525 2013

[57] httpmarathoncseeusfeduMammographyDatabasehtml[58] httpwwwcesclemsonedusimahooverstare[59] httpwwwmedharvardeduaanlib

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of

International Journal of

AerospaceEngineeringHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

RoboticsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Active and Passive Electronic Components

Control Scienceand Engineering

Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

RotatingMachinery

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

Journal ofEngineeringVolume 2014

Submit your manuscripts athttpwwwhindawicom

VLSI Design

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Shock and Vibration

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Civil EngineeringAdvances in

Acoustics and VibrationAdvances in

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Electrical and Computer Engineering

Journal of

Advances inOptoElectronics

Hindawi Publishing Corporation httpwwwhindawicom

Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

SensorsJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Modelling amp Simulation in EngineeringHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Chemical EngineeringInternational Journal of Antennas and

Propagation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Navigation and Observation

International Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

DistributedSensor Networks

International Journal of