Feature and Contrast Enhancement of Mammographic Image ... Feat… · It is obvious that features and contrast enhancement ... 3.1. Contrast Evaluation Criterion for Image. The contrast

Hindawi Publishing CorporationComputational and Mathematical Methods in MedicineVolume 2013 Article ID 716948 8 pageshttpdxdoiorg1011552013716948

Research ArticleFeature and Contrast Enhancement of MammographicImage Based on Multiscale Analysis and Morphology

Shibin Wu12 Shaode Yu12 Yuhan Yang12 and Yaoqin Xie12

1 Shenzhen Institute of Advanced Technology Chinese Academy of Science China2 Shenzhen Key Laboratory for Low-Cost Healthcare China

Correspondence should be addressed to Yaoqin Xie yqxiesiataccn

Received 26 June 2013 Revised 17 October 2013 Accepted 24 October 2013

Academic Editor Liang Li

Copyright copy 2013 Shibin Wu et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

A new algorithm for feature and contrast enhancement of mammographic images is proposed in this paperThe approach bases onmultiscale transform and mathematical morphology First of all the Laplacian Gaussian pyramid operator is applied to transformthemammography into different scale subband images In addition the detail or high frequency subimages are equalized by contrastlimited adaptive histogram equalization (CLAHE) and low-pass subimages are processed by mathematical morphology Finallythe enhanced image of feature and contrast is reconstructed from the Laplacian Gaussian pyramid coefficients modified at one ormore levels by contrast limited adaptive histogram equalization and mathematical morphology respectively The enhanced imageis processed by global nonlinear operator The experimental results show that the presented algorithm is effective for feature andcontrast enhancement of mammogramThe performance evaluation of the proposed algorithm is measured by contrast evaluationcriterion for image signal-noise-ratio (SNR) and contrast improvement index (CII)

1 Introduction

Breast cancer has been a significant public health problem forwomen in the world and early detection of breast cancer isvery essential in the field of medicine before the means toprevent breast cancer have not yet been foundHowever thereare new cases 234580 and death rate 171 from the NationalCancer Institute in the United States in 2013 [1] Breast canceraccounted for about more than 38 of cancer incidence anda significant percentage of cancer mortality in the developingand developed countries in 2009 [2] Thus it is well knownthat the early detection and treatment of breast cancer are themost effective keymeans of reducingmortality Furthermoremammography is widely recognized as being the only effec-tive and primary imagingmodality for the early detection anddiagnosis of breast cancer [3ndash5] In mammography low doseX-ray is used for imaging Hence themammographic imagesare poor in contrast and contaminated due to the low dose X-ray for imaging In low contrast mammograms it is difficultto interpret between the normal tissue and malignant tissueIn addition [6] introduced that mammographers miss about

10 of all cancerous lesions when using the poor contrastmammograms

In recent years there are many researchers that proposedall kinds of contrast enhancement algorithms to solve theseproblems produced by poor contrast images Sundaram etal [6] proposed the contrast enhancement method based onhistogram to improve the mammography quality but thismethod has not suppressed the amplified noise in histogramequalization progress Mohideen et al [2] used multiwaveletwith hard threshold to denoise and enhance mammographicimage contrast Kumar et al [7] proposed the algorithmbasedon morphology and wavelet transform for enhancement ofmammographic images Morrow et al [8] designed a region-based contrast enhancement algorithm for mammogramsThis method uses each pixel in the image as a seed togrow a region Contrast is then enhanced by applying anempirical transformation based on each regionrsquos seed pixelvalue its contrast and background information Stojic et al[9] developed an algorithm using mathematical morphologyto enhance local contrast of mammography Stahl et al [10]applied the method of nonlinear multiscale processing based

2 Computational and Mathematical Methods in Medicine

on Laplace pyramid for digital radiography enhancementHowever the binomial filter was used at each pyramid levelwith a nonlinear factor for contrast enhancement whichwas sensitive to noise Comparing with Stahlrsquos algorithmGaussian filter is used in the proposed method As far aswe know most essential image processing algorithms appliedthe Gaussian filter while the binomial filter applications arenot popular in state-of-the-art image processing algorithmsMultiscale analysis technology based on the wavelet trans-form for mammogram contrast enhancement was applied in[2 7 11ndash15] However the wavelet transform leads to unde-sirable artifacts in the enhanced image [16] To overcome thisshortcoming we propose an algorithm framework in whichthe wavelet transform is replaced with the Laplacian Gaus-sian pyramid transform for multiscale analysis Comparingwith the wavelet technique the Laplacian Gaussian pyramidtechnique seems to be amore suitable decompositionmethodfor multiscale contrast enhancement of mammogram [16]

We compare the suitability of thesemethods for enhance-ment of mammographic images in general We proposethe approach which seems to be suitable for contrastenhancement of mammograms Traditional multiscale anal-ysis methods only enhance the high frequency componentsof mammogram The proposed approach decomposes theimage by the Laplacian Gaussian pyramid transform firstlySecondly the CLAHE is adopted to enhance contrast anddetail information of each level high frequency subimagedecomposed by Laplacian pyramid transform and Gaussianfilter is used to restrain the CLAHE to enlarge the noise ofeach subimage Furthermore mathematical morphology isapplied to enhance local features and contrast at each pyra-mid levelrsquos subimage Finally the inverse Laplacian Gaussianpyramid transform is applied to reconstruct the decomposedsubimages to obtain the features and contrast enhancementimage The enhanced image was adjusted the entire contrastby a global non-linear operator for nature visualization

The rest of this paper is organized as follows Theoryof the Laplacian Gaussian pyramid transform and the keycharacteristics of the CLAHE andmathematical morphologyare detailed in Section 2 Section 3 presents the experimentalresults and discussion The conclusion of this paper is statedin Section 4

2 Materials and Method

21 The Proposed Method Traditional image enhancementtechniques cannot adapt to the varying characteristic ofimages The application of a global transform or a fixedoperator to an entire image often yields poor result at leastin some parts of the given image [10] In order to solvethese problems we propose a novel approach based on themultiscale analysis and mathematical morphology whichcannot only enhance the local detail information and edgesbut also restrict the modified noise effectively

To begin with the original image is decomposed bythe Laplacian Gaussian pyramid transform to obtain lowfrequency subband images and different scale coefficientsof the high frequency components At present all kinds

of filter techniques including gray-scale operator medianfiltering edge enhancement operator frequency enhance-ment operator spatial-frequency filtering and anisotropicadaptive filtering are applied to reduce the noises of imagefor achieving image enhancementHowever these techniquesare suitable for features enhancement of the whole imageIn fact we do need to enhance the local features and edgesof medical image in some practical clinical applicationMathematical morphology operations fit these requirementsThe morphology operators can enrich the low frequencyinformation of the image and enhance this part contrastin terms of the principle of mathematical morphologyTherefore noise in differentmultiscale subimages can be sup-pressed by morphology operations Vessels fibroglandulartissues masses and calcification points are reinforced at eachpyramid level Furthermoremorphology operators canmakethe feature details become smoother and the edges sharperMost essential image processing algorithmcanbe representedin the form of morphological operations In addition thelow frequency subbands contain basic information and partscontrast of imageTherefore the processing of low frequencydata is important and cannot be ignored For the low-passfiltered subband images we apply themathematical morpho-logical operations combining opening operationwith closingoperation to enrich the basic information and enhance imagecontrast

Secondly histogram equalization is used for enrichingdetail information and sharping edges of the whole imageThe detail information and image edges belong to the highfrequency components Furthermore the high frequencysubbands contain noises We adopt the CLAHE to enhancethe high frequency subbands coefficients which can not onlyenhance the features and image contrast and enrich the detailinformation and image edges but also effectively suppressesthe enlarged noise Comparing with the AHE the CLAHEcan reduce computational time and effectively suppress theenlarged noise Comparing with the traditional histogramequalization and the modified histogram equalization theCLAHE can effectively enhance the local features edges andimage contrast

Finally we can reconstruct an enhanced image and itssize is the same as the original image We apply the contrastlow frequency coefficients adjusted through mathematicalmorphology and the high frequency subbands processed bythe CLAHE to extract the enhanced feature and contrastimage due to the Laplacian Gaussian pyramid transform isof the property of reversal Besides a global gain operation isused to adjust the contrast of reconstructed image in order tomake the enhanced image more natural and smoother Theflowchart of the proposed algorithm is shown in Figure 1

22 The Laplacian Gaussian Pyramid Transform The Lapla-cian Gaussian pyramid technique was developed by Burtand Adelson for the context of compression of images [1617] The Laplacian Gaussian pyramid transform has beenused to analyze images at multiscale analysis for a broadrange of application [17] The purpose of multiscale imagecontrast enhancement is that the original image is divided

Computational and Mathematical Methods in Medicine 3

Input image

Laplacian gaussian pyramid decomposition

Laplacian gaussian pyramid reconstruction

Global contrast gain adjustment

Enhanced image

Gaussian low filter

Low frequency coefficients

Morphology

High pass filter

High frequency coefficients

CLAHE

Figure 1 Flowchart of the proposed method

into several different multiscale levelsrsquo subband images whichare enhanced by the CLAHE and morphology operationsrespectively All levelsrsquo subband images are reconstructedto extract the enhanced image The flow diagram of theLaplacian Gaussian pyramid transform for the decompo-sition and reconstruction processes of image is shown asFigure 2 The original image is filtered by Gaussian low-pass filter and subsampled to produce 119892

1 The image 119892

1is

then interpolated or made the convolution operations withkernel width 5 to reproduce the original array size andsubtracted pixelwise from the original image to produce 119887

0

This subband image 1198870 which is produced by an equivalent

process of high-pass filter is the finest level of the LaplacianGaussian pyramid The decimated low-pass filtering image1198921is further filtered by the Gaussian low-pass filter and

subsampled producing 1198922 The 119892

2is interpolated or made the

convolution operations with kernel width 5 and subtractedfrom the 119892

1 which results in the second pyramid layer 119887

1

All subsequent layers of the Laplacian Gaussian pyramid arecomputed by repeating these operations to the subsampledGaussian low-pass filtering images from the previous iter-ation until the setting pyramid level image 119887

119871minus1and the

last level pyramid image 119892119871are obtained The flowchart of

the reconstruction process is drawn on the right hand ofthe Figure 2 The subimage 119892


convolution operations with kernel width 5 to the array sizeof the next finer pyramid level 119887

119871minus1 and the enhanced 1198871015840

119871minus1

is added in this pixelwise to produce 1198921015840

119871minus1 Interpolation

contrast enhancement convolution and addition operationsare repeated until the reconstructed image at the originalresolution level is obtained The reconstruction is completelyreversible if the interpolation filters used in decompositionand reconstruction are identical [10 16ndash19]

It is obvious that features and contrast enhancementprocesses are implemented between the 119887

119871minus1and 119887

1015840

119871minus1 The

image 119892119871is adjusted according to mathematical morphologi-

cal operations to obtain contrast enhancement 1198921015840119871 Repeating

the reconstruction process can extract the enhanced image

23 Contrast Limited Adaptive Histogram Equalization His-togram equalization is a specific case of themore general classof histogram remapping methods Histogram equalizationbecause of its advantages of high speed and better effect iswidely applied to enhance the contrast of mammographicimagesHistogram is the function of gray level which denotesthe gray level of every pixel Therefore the contrast ratiowill be improved by a gray nonlinear transform to adjust theaccumulation function and the gray in small range will betransformed in the whole field

The histogram is a discrete function and defined asfollows

119901119903(119903119896) =

119899119896

119899 (1)

where 119899 is the total pixels of a mammogram and 119899119896is pixel

number of corresponding the 119903119896gray level

Hypothesis that the gray transfer function is 119904 = 119879(119903)whose slope is limited to non-minus continuum monotoneincreasing function and it can transform input image 119868(119909 119910)to output image 1198681015840(119909 119910) Let 119901

119903(119903) and 119901

119904(119904) represent the

probability density function of the random variable 119903 and 119904respectively 119903 is the gray level of the input image and 119904 is thegray level of the output image According to the definition ofthe histogram and the cumulate density function the originalimage histogram and processed histogram areas are equal to

119901119904(119904) = 119901

119903(119903)

119889119903

119889119904 (2)

Let 119904 belong to [0 119871 minus 1] the gray transfer function isexpressed as follows

119904 = 119879 (119903) = (119871 minus 1) int

119903

0

119901119903(119908) 119889119908 (3)

where 119908 is an integral dummy variableAccording to the properties of integral we can make the

derivation of formula (3) to get a new equation as follows

119889119904

119889119903=119889119879 (119903)

119889119903(119871 minus 1)

119889

119889119903[int

119903

0

119901119903(119908) 119889119908] = (119871 minus 1) 119901

119903(119903)

(4)

Further we can put the formula (4) into (2) and obtain anew formula as follows

119901119904(119904) = 119901

119903(119903)

10038161003816100381610038161003816100381610038161003816

119889119903

119889119904

10038161003816100381610038161003816100381610038161003816= 119901119903(119903)

10038161003816100381610038161003816100381610038161003816

1

(119871 minus 1) 119901119903(119903)

10038161003816100381610038161003816100381610038161003816

=1

119871 minus 1 0 le 119904 le 119871 minus 1

(5)


LP LP

LP LP

LP LP

LP LP

Quantizationprocessing

Original image Enhanced image

LP

LP

LP

LP

g1

g2

g3

gL

minus

minus

minus

minusminus

minus

minus

minus+

+

+

+

+

+

+

+

Σ

Σ

Σ

Σ

Σ

Σ

Σ

Σ

b0

b1

b2

bLminus1

g9984001

g9984002

g9984003

g998400L

b9984000

b9984001

b9984002

b998400Lminus1

Figure 2 Flow diagram of the Laplacian Gaussian pyramid transform

Limitedfunction Pixels

Gray level

(a) (b)

Gray level

Pixels

Figure 3 Limited function chart (a) Original gray chart and (b) limited contrast gray chart

According to (1) and the gray transfer function (3)transform we can induce a discrete form given as follows

119904119896= 119879 (119903

119896) = (119871 minus 1)

119896

sum

119895=0

119901119903(119903119895) =

119871 minus 1

119899

119896

sum

119895=0

119899119895

119896 = 0 1 2 119871 minus 1

(6)

In generally (6) is the gray level remapping functionComparing with the whole histogram equalization the

adaptive histogram equalization (AHE) has the advantageof good local contrast enhancement But the AHE needs tocompute the local histogram and accumulate distributionfunction of every pixel it is extremely computational inten-sive Besides the AHE is sensitive to noise

The AHE enhances the image contrast and enlarges thenoise In some cases enhancement process results in imagedistortion in some detail area which affects the visual diag-nosis of clinicians for the enhanced image [20 21] We needto not only enhance the features and image contrast but alsorestrict the magnified noise Thus limiting contrast functionto AHE in every block is needed to generate transformfunction respectively How much percent of contrast will be

restricted and enhanced It is need to preliminarily adjust thecontrast of each block at each pyramid level image Thus wedefine a limited function to limit the gray level probabilitydensity and control the exceed histogram The adjustableprocessing is shown as Figure 3

24 Mathematical Morphology Mathematical morphologyoriginated in set theory geometry and topology establishesthe relationship between the geometry of physical systemand some of its property Most essential image processingalgorithms can be represented in the form of morphologicaloperations For example morphology offers a unified andpowerful approach to different image processing problems[9 11 22] We apply the morphological opening and clos-ing operations to process the different multiscale subbandimages The opening and closing operations are producedby combining dilation and erosion operator respectivelyFurthermore a structuring element SE of rectangle shape isused in dilation and erosion operator respectively

The erosion of a gray-scale digital image 119868(119909 119910) by astructural element SE(119894 119895) is defined as follows [7 9 11]

(119868 otimes SE) (119898 119899) = min 119868 (119898 minus 119894 119899 + 119895) minus SE (119894 119895) (7)


(a) (b)

(c) (d)

(e) (f)

Figure 4 Comparison of contrast enhancement of mammograms (a) Original mammogram (b) Enhanced through histogram equalization(c) Enhanced through adaptive histogram equalization (d) Enhanced through nonlinear multiscale processing based on Laplace pyramid(e) Enhanced through the method based on wavelet transform and morphology (f) Enhanced through the proposed method

The gray-scale dilation can be described as

(119868 oplus SE) (119898 119899) = max 119868 (119898 minus 119894 119899 minus 119895) + SE (119894 119895) (8)

The morphological opening and closing operations havethe same form as the binary counterparts The openingoperation of image 119868(119909 119910) using the structure elements SEis defined as erosion followed by dilation and expressed as(9) Utilizing the structure element SE the closing operationof image 119868(119909 119910) is defined as dilation followed by erosion andgiven by (10) as follows

119868 ∘ SE = (119868 otimes SE) oplus SE (9)

119868 ∙ SE = (119868 oplus SE) otimes SE (10)

The opening and closing operations can be interpreted asfollows the opening can remove all of the pixels (light details)in a region that are smaller than the structure element Forthe corresponding opposite sequence the closing can fill inholes and concavities smaller than the structure elementThetwo operations can remove the noise in the image and cannotmake the image distortion

In practical application morphological opening and clos-ing pairs are combined in sequence for various image pro-cessing operations There are two morphological operationsknown as top-hat (TH) and bottom-hat (BH) transforma-tions The top-hat transformation by opening is defined as

the difference between the original image and its gray scaleopening using structuring element SE and it is defined as

TH = 119868 minus (119868 ∘ SE) (11)

Similarly dual bottom-hat transformation by closing isthe difference between the gray-scale closing image andoriginal image as described by

BH = (119868 ∙ SE) minus 119868 (12)

Based on the previous theoretical analysis the TH trans-formation is an effective technique for enhancing small brightdetails from the background Conversely the dark featurescan be extracted from a brighter background by the BHtransformation In order to enhance the local contrast of themammograms the processing procedure is adding originalimage to the top-hat transformed image and subtractingthe bottom-hat image Furthermore its efficiency in imagecontrast enhancement has been proved by [9]The calculatedformula is given as follows

119862 = 119868 + TH minus BH (13)

Equation (13) has been used to process the low-passfrequency coefficients in the proposed algorithm whichenhances the features and contrast of mammographic image


(a) (b)

(c) (d)

(e) (f)


25 Global Gain Adjustment We adopt a global gain adjust-ment technique in our investigation for histogram equaliza-tion enhancing part pixels uniformly at each pyramid levelwhich makes the visualization of the enhanced image tobecome more natural We employ the mean introduced in[3 22] to accomplish this nonlinear operation The globalgain adjustment function can be expressed as

119891 (119911) = 119886 [sigm (119888 (119911 minus 119887)) minus sigm (minus119888 (119911 + 119887))] (14)

where 119886 = (1(sigm(119888(1 minus 119887)) minus sigm(minus119888(1 + 119887)))) subject to0 lt 119887 lt 1 b and 119888 are ratio coefficients of enhancement andcan be set different values according to the practical testingexperiment to adaptively adjust

The sigm(119911) is described as follows

sigm (119911) =1

1 + 119890minus119911 (15)

3 Results and Discussion

The proposed algorithm has been applied to more than 10mammographic images being 1944 lowast 3072 in sizes fromthe Angell Medical ADM-600MG We had readjusted the

size of each tested original image to be 1280 lowast 3072 inorder to decrease the cost times The testing procedure hasbeen implemented in MATLAB2012a To demonstrate theeffectiveness of our method we compared its results withthe existing popularmethods of histogramequalization (HE)AHE the algorithm of nonlinear multiscale processing basedon Laplace pyramid proposed in [10] and the method basedonmorphology andwavelet transform Furthermore we takethe advantage of the metrics of contrast evaluation crite-rion SNR and contrast improvement index (CII) to mea-sure the quantitative performance analysis of the proposedmethod

31 Contrast Evaluation Criterion for Image The contrastof enhanced image is evaluated by employing the metricfunction which was proposed in [4] and given as follows

119862119888=

1

119872119873

119872

sum

119894=1

119873

sum

119895=1

11989110158402(119894 119895) minus

10038161003816100381610038161003816100381610038161003816100381610038161003816

1

119872119873

119872

sum

119894=1

119873

sum

119895=1

1198911015840(119894 119895)

10038161003816100381610038161003816100381610038161003816100381610038161003816

2

(16)

where 119872 and 119873 are height and width of the image respec-tively and 1198911015840(119894 119895) is the enhanced imageThe larger the valueof (16) is the better the contrast of the image is


32 Contrast Improvement Index A quantization measure ofcontrast enhancement can be defined by a contrast improve-ment index and its formula can be expressed by the following[8 14 15 23]

CII =119862processed

119862original (17)

where119862processed and119862original are the contrasts of the processedand original images respectively C is the average value of thelocal region contrast in the processed or original imageThusthe CII value of original image is equal to oneThe local con-trast at each pixel is measured as (XmaxminusXmin)(Xmax+Xmin)in its local window sizeWe applied a version of the optimizeddefinition of contrast demonstrated in [8] The contrast 119862 ofan image is described as follows

119862 =119898119891minus 119898119887

119898119891+ 119898119887

(18)

where119898119891is themean luminance value of the foreground and

the corresponding 119898119887is equal to the mean luminance value

of background In our experiment we use the 5 lowast 5 localwindows The greater value of CII indicates that the givenimage quality is better

As shown in Figures 4 and 5 (a) represents one of two dif-ferent womenrsquos mammographic original images respectivelyand ((b)ndash(f)) are produced byHE AHE nonlinearmultiscaleprocessing based on Laplace pyramid the method basedon morphology and wavelet transform and the proposedmethod respectively In the histogram equalization thepixels are spreading uniformlyThe traditional HE method isadopted to enhance the original image The produced resultsthrough HE improve the contrast a little but we find itdifficult to identify tissue nodes and features and the HEover amplifies the noise which results in the fibroglandularalbefaction and the enhanced results cannot be applied toclinic It seems that the enhanced contrast images throughAHE method are degraded in visual quality In addition theAHEmethod over enhances the original imagesThe contrastand features of the enhanced images are improved by usingStahlrsquos algorithm However it can be seen that the noiseare enlarged The low frequency components are ignored Itis difficult to identify the fibroglandular and tissue nodesand the results cannot be applied in clinical applicationAlthough the algorithm based on wavelet transform andmorphology has made the original image become smootherand made the noise of image decrease and enhanced themammograms contrast but we can almost not distinguishthe details and tissue nodes Hence the processed imagesdo not play a significant role in practical clinic The resultsenhanced by the proposed algorithm are not only savinggood details but also the edges are preserved and enhancedFeatures and fibroglandular are enhanced the tissue nodescan be identified clearly in the apparent visual quality andthe contrast of mammogram has been improved a lot

In Tables 1 and 2 comparison of values of contrast andcontrast improvement index shows that the proposedmethodoutperforms the HE AHE the approach of nonlinear

Table 1 Measurement evaluation of contrast CII and SNR forFigure 4

Method Contrast CII SNROriginal image 00503 10000HE 01607 25140 141344AHE 00756 35981 40076Nonlinear method basedon Laplace pyramid 01230 40467 69851

Wavelet method 01137 54673 181565The proposed method 02677 145514 129723




multiscale processing based on Laplace pyramid and themethod based on wavelet transform and morphology inall the experimental images which indicates better contrastenhancement of images Corresponding to SNR values theenhanced results of the wavelet method are the largest inour experiments SNR is closely related to the luminance ofimage The enhanced mammographic images of the waveletmethod are the brightest in visualization but the results donot play a significant role in the clinical application We canclearly identify that the proposed method can better enhanceimage contrast and preserve the detailed information of theimage Both visual quality and quantitative measurementshave shown that the proposed method is more suitablefor features and contrast enhancement of mammographicimages

4 Conclusion

A method based on the Laplacian Gaussian pyramid trans-form mathematical morphology and contrast limited adap-tive histogram equalization is proposed for features andcontrast enhancement of mammographic images whichemploys the penalty terms to adjust the various aspects ofcontrast enhancementThus the proposed method enhancesthe image contrast at the same time as it effectively restrictsthe enlarged noise The proposed method has been testedon mammograms and compared with the existing popularapproaches of histogram equalization and adaptive his-togram equalization nonlinear multiscale processing basedon Laplace pyramid and the method based on wavelettransform and morphology Furthermore the experimentalresults show that the proposed method seems more suitable


formammographic images enhancement in both visual qual-ity and qualitative measurements Besides the experimentalresults yielded by the proposed method still have a littleunnatural in visualization Therefore the next step we willimprove the proposed method and test it using mammo-graphic images from the standard Database MammographicImage Analysis Society (MIAS) For performance evaluationof the proposed algorithm SNR contrast evaluation criterionand contrast improvement index are adopted The experi-mental results and metric data tables show that the proposedmethod seems to yield significantly the enhanced imagesComparing with other tested methods the results of theproposed method are more suitable for clinic application

Acknowledgments

This work is supported in part by Grants from the NationalNatural Science Foundation of China (NSFC 81171402)NSFC Joint Research Fund for Overseas Research ChineseHong Kong andMacao Young Scholars (30928030) NationalBasic Research Program 973 (2010CB732606) from Ministryof Science and Technology of China Guangdong InnovativeResearch Team Program (no 2011S013) of China Scienceand Technological Program for Dongguanrsquos Higher Educa-tion Science and Research Health Care Institutions (Grantno 2011108101001) Comprehensive Strategic CooperationProject of Guangdong province and Chinese academy ofsciences (Grant no 2011B090300079)

References

[1] National Cancer Institute Breast Cancer 2013 httpwwwcancergovcancertopicstypesbreast

[2] K Mohideen A Perumal Krishnan and M Sathik ldquoImagedenoising and enhancement using multiwavelet with hardthreshold in digital mammographic imagesrdquo International ArabJournal of e-Technology vol 2 no 1 pp 49ndash55 2011

[3] A F Laine S Schuler J Fan and W Huda ldquoMammographicfeature enhancement by multiscale analysisrdquo IEEE Transactionson Medical Imaging vol 13 no 4 pp 725ndash740 1994

[4] M Sundaram K Ramar N Arumugam and G Prabin ldquoHis-togram modified local contrast enhancement for mammogramimagesrdquoApplied Soft Computing Journal vol 11 no 8 pp 5809ndash5816 2011

[5] MAdelD ZuwalaMRasigni et al ldquoNoise reduction onmam-mographic phantom imagesrdquo Electronic Letters on ComputerVision and Image Analysis vol 5 no 4 pp 64ndash74 2006

[6] M Sundaram K Ramar N Arumugam and G Prabin ldquoHis-togram based contrast enhancement for mammogram imagesrdquoin Proceedings of the International Conference on Signal Process-ing Communication Computing and Networking Technologies(ICSCCN rsquo11) pp 842ndash846 Thuckafay India July 2011

[7] N H Kumar S Amutha and D R R Babu ldquoEnhancementof mammographic images using morphology and wavelettransformrdquo Computer Technology Application vol 3 no 1 pp192ndash198 2012

[8] W M Morrow R B Paranjape R M Rangayyan and J EL Desautels ldquoRegion-based contrast enhancement of mammo-gramsrdquo IEEE Transactions on Medical Imaging vol 11 no 3 pp392ndash406 1992

[9] T Stojic I Reljin and B Reljin ldquoLocal contrast enhancement indigital mammography by using mathematical morphologyrdquo inProceedings of the International Symposium on Signals Circuitsand Systems (ISSCS rsquo05) vol 2 pp 609ndash612 July 2005

[10] M Stahl T Aach and S Dippel ldquoDigital radiography enhance-ment by nonlinear multiscale processingrdquoMedical Physics vol27 no 1 pp 56ndash65 2000

[11] S Amutha D R R Babu M R Shankar and N H KumarldquoMammographic image enhancement using modified mathe-matical morphology and Bi-orthogonal waveletrdquo in Proceedingsof the IEEE International Symposium on IT in Medicine andEducation (ITME rsquo11) vol 1 pp 548ndash553 Cuangzhou ChinaDecember 2011

[12] A Mencattini M Salmeri R Lojacono M Frigerio and FCaselli ldquoMammographic images enhancement and denoisingfor breast cancer detection using dyadic wavelet processingrdquoIEEETransactions on Instrumentation andMeasurement vol 57no 7 pp 1422ndash1430 2008

[13] P Gorgel A Sertbas and O N Ucan ldquoA wavelet-based mam-mographic image denoising and enhancement with homomor-phic filteringrdquo Journal ofMedical Systems vol 34 no 6 pp 993ndash1002 2010

[14] A Laine J Fan and W Yang ldquoWavelets for contrast enhance-ment of digital mammographyrdquo IEEE Engineering in Medicineand Biology Magazine vol 14 no 5 pp 536ndash550 1995

[15] A Laine W Huda B G Steinbach and J C HoneymanldquoMammographic image processing using wavelet processingtechniquesrdquo European Radiology vol 5 no 5 pp 518ndash523 1995

[16] S Dippel M Stahl R Wiemker and T Blaffert ldquoMultiscalecontrast enhancement for radiographies Laplacian pyramidversus fast wavelet transformrdquo IEEE Transactions on MedicalImaging vol 21 no 4 pp 343ndash353 2002

[17] P J Burt and E H Adelson ldquoThe Laplacian pyramid as acompact image coderdquo IEEE Transactions on Communicationsvol 31 no 4 pp 532ndash540 1983

[18] P Vuylsteke and E Schoeters ldquoMultiscale image contrastamplification (MUSICA)rdquo Image Processing vol 2167 pp 551ndash560 1994

[19] F Sattar L Floreby G Salomonsson and B Lovstrom ldquoImageenhancement based on a nonlinear multiscale methodrdquo IEEETransactions on Image Processing vol 6 no 6 pp 888ndash895 1997

[20] SM Pizer E P Amburn J D Austin et al ldquoAdaptive histogramequalization and its variationsrdquo Computer Vision Graphics andImage Processing vol 39 no 3 pp 355ndash368 1987

[21] S A Ahmad M N Taib N E A Khalid and H TaibldquoAn analysis of image enhancement techniques for dental X-ray image interpretationrdquo International Journal of MachineLearning and Computing vol 2 no 3 pp 292ndash297 2012

[22] D Giordano I Kavasidis and C Spampinato ldquoAdaptive localcontrast enhancement combined with 2D discrete wavelettransform for mammographic mass detection and classifica-tionrdquo Communications in Computer and Information Sciencevol 166 no 1 pp 209ndash218 2011

[23] M Trivedi A Jaiswal and V Bhateja ldquoA new contrast improve-ment index based on logarithmic image processing modelrdquoAdvances in Intelligent System and Computing vol 199 pp 715ndash723 2013


on Laplace pyramid for digital radiography enhancementHowever the binomial filter was used at each pyramid levelwith a nonlinear factor for contrast enhancement whichwas sensitive to noise Comparing with Stahlrsquos algorithmGaussian filter is used in the proposed method As far aswe know most essential image processing algorithms appliedthe Gaussian filter while the binomial filter applications arenot popular in state-of-the-art image processing algorithmsMultiscale analysis technology based on the wavelet trans-form for mammogram contrast enhancement was applied in[2 7 11ndash15] However the wavelet transform leads to unde-sirable artifacts in the enhanced image [16] To overcome thisshortcoming we propose an algorithm framework in whichthe wavelet transform is replaced with the Laplacian Gaus-sian pyramid transform for multiscale analysis Comparingwith the wavelet technique the Laplacian Gaussian pyramidtechnique seems to be amore suitable decompositionmethodfor multiscale contrast enhancement of mammogram [16]

We compare the suitability of thesemethods for enhance-ment of mammographic images in general We proposethe approach which seems to be suitable for contrastenhancement of mammograms Traditional multiscale anal-ysis methods only enhance the high frequency componentsof mammogram The proposed approach decomposes theimage by the Laplacian Gaussian pyramid transform firstlySecondly the CLAHE is adopted to enhance contrast anddetail information of each level high frequency subimagedecomposed by Laplacian pyramid transform and Gaussianfilter is used to restrain the CLAHE to enlarge the noise ofeach subimage Furthermore mathematical morphology isapplied to enhance local features and contrast at each pyra-mid levelrsquos subimage Finally the inverse Laplacian Gaussianpyramid transform is applied to reconstruct the decomposedsubimages to obtain the features and contrast enhancementimage The enhanced image was adjusted the entire contrastby a global non-linear operator for nature visualization

The rest of this paper is organized as follows Theoryof the Laplacian Gaussian pyramid transform and the keycharacteristics of the CLAHE andmathematical morphologyare detailed in Section 2 Section 3 presents the experimentalresults and discussion The conclusion of this paper is statedin Section 4

2 Materials and Method

21 The Proposed Method Traditional image enhancementtechniques cannot adapt to the varying characteristic ofimages The application of a global transform or a fixedoperator to an entire image often yields poor result at leastin some parts of the given image [10] In order to solvethese problems we propose a novel approach based on themultiscale analysis and mathematical morphology whichcannot only enhance the local detail information and edgesbut also restrict the modified noise effectively

To begin with the original image is decomposed bythe Laplacian Gaussian pyramid transform to obtain lowfrequency subband images and different scale coefficientsof the high frequency components At present all kinds

of filter techniques including gray-scale operator medianfiltering edge enhancement operator frequency enhance-ment operator spatial-frequency filtering and anisotropicadaptive filtering are applied to reduce the noises of imagefor achieving image enhancementHowever these techniquesare suitable for features enhancement of the whole imageIn fact we do need to enhance the local features and edgesof medical image in some practical clinical applicationMathematical morphology operations fit these requirementsThe morphology operators can enrich the low frequencyinformation of the image and enhance this part contrastin terms of the principle of mathematical morphologyTherefore noise in differentmultiscale subimages can be sup-pressed by morphology operations Vessels fibroglandulartissues masses and calcification points are reinforced at eachpyramid level Furthermoremorphology operators canmakethe feature details become smoother and the edges sharperMost essential image processing algorithmcanbe representedin the form of morphological operations In addition thelow frequency subbands contain basic information and partscontrast of imageTherefore the processing of low frequencydata is important and cannot be ignored For the low-passfiltered subband images we apply themathematical morpho-logical operations combining opening operationwith closingoperation to enrich the basic information and enhance imagecontrast

Secondly histogram equalization is used for enrichingdetail information and sharping edges of the whole imageThe detail information and image edges belong to the highfrequency components Furthermore the high frequencysubbands contain noises We adopt the CLAHE to enhancethe high frequency subbands coefficients which can not onlyenhance the features and image contrast and enrich the detailinformation and image edges but also effectively suppressesthe enlarged noise Comparing with the AHE the CLAHEcan reduce computational time and effectively suppress theenlarged noise Comparing with the traditional histogramequalization and the modified histogram equalization theCLAHE can effectively enhance the local features edges andimage contrast

Finally we can reconstruct an enhanced image and itssize is the same as the original image We apply the contrastlow frequency coefficients adjusted through mathematicalmorphology and the high frequency subbands processed bythe CLAHE to extract the enhanced feature and contrastimage due to the Laplacian Gaussian pyramid transform isof the property of reversal Besides a global gain operation isused to adjust the contrast of reconstructed image in order tomake the enhanced image more natural and smoother Theflowchart of the proposed algorithm is shown in Figure 1

22 The Laplacian Gaussian Pyramid Transform The Lapla-cian Gaussian pyramid technique was developed by Burtand Adelson for the context of compression of images [1617] The Laplacian Gaussian pyramid transform has beenused to analyze images at multiscale analysis for a broadrange of application [17] The purpose of multiscale imagecontrast enhancement is that the original image is divided


Input image




Enhanced image

Gaussian low filter


Morphology

High pass filter


CLAHE



1 The image 119892

1is


0







1


119871minus1and the






119871minus1





119871minus1and 119887

1015840

119871minus1 The






119901119903(119903119896) =

119899119896

119899 (1)




119903(119903) and 119901



119901119904(119904) = 119901

119903(119903)

119889119903

119889119904 (2)


119904 = 119879 (119903) = (119871 minus 1) int

119903

0

119901119903(119908) 119889119908 (3)



119889119904

119889119903=119889119879 (119903)

119889119903(119871 minus 1)

119889

119889119903[int

119903

0

119901119903(119908) 119889119908] = (119871 minus 1) 119901

119903(119903)

(4)


119901119904(119904) = 119901

119903(119903)

10038161003816100381610038161003816100381610038161003816

119889119903

119889119904

10038161003816100381610038161003816100381610038161003816= 119901119903(119903)

10038161003816100381610038161003816100381610038161003816

1

(119871 minus 1) 119901119903(119903)

10038161003816100381610038161003816100381610038161003816

=1

119871 minus 1 0 le 119904 le 119871 minus 1

(5)


LP LP

LP LP

LP LP

LP LP



LP

LP

LP

LP

g1

g2

g3

gL

minus

minus

minus

minusminus

minus

minus

minus+

+

+

+

+

+

+

+

Σ

Σ

Σ

Σ

Σ

Σ

Σ

Σ

b0

b1

b2

bLminus1

g9984001

g9984002

g9984003

g998400L

b9984000

b9984001

b9984002

b998400Lminus1



Gray level

(a) (b)

Gray level

Pixels



119904119896= 119879 (119903

119896) = (119871 minus 1)

119896

sum

119895=0

119901119903(119903119895) =

119871 minus 1

119899

119896

sum

119895=0

119899119895

119896 = 0 1 2 119871 minus 1

(6)









(a) (b)

(c) (d)

(e) (f)










TH = 119868 minus (119868 ∘ SE) (11)


BH = (119868 ∙ SE) minus 119868 (12)


119862 = 119868 + TH minus BH (13)



(a) (b)

(c) (d)

(e) (f)






sigm (119911) =1

1 + 119890minus119911 (15)





119862119888=

1

119872119873

119872

sum

119894=1

119873

sum

119895=1

11989110158402(119894 119895) minus

10038161003816100381610038161003816100381610038161003816100381610038161003816

1

119872119873

119872

sum

119894=1

119873

sum

119895=1

1198911015840(119894 119895)

10038161003816100381610038161003816100381610038161003816100381610038161003816

2

(16)





119862original (17)


119862 =119898119891minus 119898119887

119898119891+ 119898119887

(18)













4 Conclusion




Acknowledgments


References

























Input image




Enhanced image

Gaussian low filter


Morphology

High pass filter


CLAHE



1 The image 119892

1is


0







1


119871minus1and the






119871minus1





119871minus1and 119887

1015840

119871minus1 The






119901119903(119903119896) =

119899119896

119899 (1)




119903(119903) and 119901



119901119904(119904) = 119901

119903(119903)

119889119903

119889119904 (2)


119904 = 119879 (119903) = (119871 minus 1) int

119903

0

119901119903(119908) 119889119908 (3)



119889119904

119889119903=119889119879 (119903)

119889119903(119871 minus 1)

119889

119889119903[int

119903

0

119901119903(119908) 119889119908] = (119871 minus 1) 119901

119903(119903)

(4)


119901119904(119904) = 119901

119903(119903)

10038161003816100381610038161003816100381610038161003816

119889119903

119889119904

10038161003816100381610038161003816100381610038161003816= 119901119903(119903)

10038161003816100381610038161003816100381610038161003816

1

(119871 minus 1) 119901119903(119903)

10038161003816100381610038161003816100381610038161003816

=1

119871 minus 1 0 le 119904 le 119871 minus 1

(5)


LP LP

LP LP

LP LP

LP LP



LP

LP

LP

LP

g1

g2

g3

gL

minus

minus

minus

minusminus

minus

minus

minus+

+

+

+

+

+

+

+

Σ

Σ

Σ

Σ

Σ

Σ

Σ

Σ

b0

b1

b2

bLminus1

g9984001

g9984002

g9984003

g998400L

b9984000

b9984001

b9984002

b998400Lminus1



Gray level

(a) (b)

Gray level

Pixels



119904119896= 119879 (119903

119896) = (119871 minus 1)

119896

sum

119895=0

119901119903(119903119895) =

119871 minus 1

119899

119896

sum

119895=0

119899119895

119896 = 0 1 2 119871 minus 1

(6)









(a) (b)

(c) (d)

(e) (f)










TH = 119868 minus (119868 ∘ SE) (11)


BH = (119868 ∙ SE) minus 119868 (12)


119862 = 119868 + TH minus BH (13)



(a) (b)

(c) (d)

(e) (f)






sigm (119911) =1

1 + 119890minus119911 (15)





119862119888=

1

119872119873

119872

sum

119894=1

119873

sum

119895=1

11989110158402(119894 119895) minus

10038161003816100381610038161003816100381610038161003816100381610038161003816

1

119872119873

119872

sum

119894=1

119873

sum

119895=1

1198911015840(119894 119895)

10038161003816100381610038161003816100381610038161003816100381610038161003816

2

(16)





119862original (17)


119862 =119898119891minus 119898119887

119898119891+ 119898119887

(18)













4 Conclusion




Acknowledgments


References

























LP LP

LP LP

LP LP

LP LP



LP

LP

LP

LP

g1

g2

g3

gL

minus

minus

minus

minusminus

minus

minus

minus+

+

+

+

+

+

+

+

Σ

Σ

Σ

Σ

Σ

Σ

Σ

Σ

b0

b1

b2

bLminus1

g9984001

g9984002

g9984003

g998400L

b9984000

b9984001

b9984002

b998400Lminus1



Gray level

(a) (b)

Gray level

Pixels



119904119896= 119879 (119903

119896) = (119871 minus 1)

119896

sum

119895=0

119901119903(119903119895) =

119871 minus 1

119899

119896

sum

119895=0

119899119895

119896 = 0 1 2 119871 minus 1

(6)









(a) (b)

(c) (d)

(e) (f)










TH = 119868 minus (119868 ∘ SE) (11)


BH = (119868 ∙ SE) minus 119868 (12)


119862 = 119868 + TH minus BH (13)



(a) (b)

(c) (d)

(e) (f)






sigm (119911) =1

1 + 119890minus119911 (15)





119862119888=

1

119872119873

119872

sum

119894=1

119873

sum

119895=1

11989110158402(119894 119895) minus

10038161003816100381610038161003816100381610038161003816100381610038161003816

1

119872119873

119872

sum

119894=1

119873

sum

119895=1

1198911015840(119894 119895)

10038161003816100381610038161003816100381610038161003816100381610038161003816

2

(16)





119862original (17)


119862 =119898119891minus 119898119887

119898119891+ 119898119887

(18)













4 Conclusion




Acknowledgments


References

























(a) (b)

(c) (d)

(e) (f)










TH = 119868 minus (119868 ∘ SE) (11)


BH = (119868 ∙ SE) minus 119868 (12)


119862 = 119868 + TH minus BH (13)



(a) (b)

(c) (d)

(e) (f)






sigm (119911) =1

1 + 119890minus119911 (15)





119862119888=

1

119872119873

119872

sum

119894=1

119873

sum

119895=1

11989110158402(119894 119895) minus

10038161003816100381610038161003816100381610038161003816100381610038161003816

1

119872119873

119872

sum

119894=1

119873

sum

119895=1

1198911015840(119894 119895)

10038161003816100381610038161003816100381610038161003816100381610038161003816

2

(16)





119862original (17)


119862 =119898119891minus 119898119887

119898119891+ 119898119887

(18)













4 Conclusion




Acknowledgments


References

























(a) (b)

(c) (d)

(e) (f)






sigm (119911) =1

1 + 119890minus119911 (15)





119862119888=

1

119872119873

119872

sum

119894=1

119873

sum

119895=1

11989110158402(119894 119895) minus

10038161003816100381610038161003816100381610038161003816100381610038161003816

1

119872119873

119872

sum

119894=1

119873

sum

119895=1

1198911015840(119894 119895)

10038161003816100381610038161003816100381610038161003816100381610038161003816

2

(16)





119862original (17)


119862 =119898119891minus 119898119887

119898119891+ 119898119887

(18)













4 Conclusion




Acknowledgments


References



























119862original (17)


119862 =119898119891minus 119898119887

119898119891+ 119898119887

(18)













4 Conclusion




Acknowledgments


References


























Acknowledgments


References
























Documents

Feature and Contrast Enhancement of Mammographic Image ... Feat… · It is obvious that features and contrast enhancement ... 3.1. Contrast Evaluation Criterion for Image. The contrast