Melanoma recognition framework based on expert

definition of ABCD for dermoscopic images

Qaisar Abbas1,2, M. Emre Celebi3, Irene Fondon Garcia4 and Waqar Ahmad1,2

1Department of Computer Science, National Textile University, Faisalabad, 37610, Pakistan,2Center for Biomedical Imaging and Bioinformatics, Key Laboratory of Image Processing, Faisalabad, Pakistan,

3Department of Computer Science, Louisiana State University, Shreveport, LA, USA and 4Department of Signal Theory and Communications,School of Engineering Path of Discovery, s/n C. P., 41092, Seville, Spain

Background/purpose: Melanoma Recognition based on clini-

cal ABCD rule is widely used for clinical diagnosis of pig-

mented skin lesions in dermoscopy images. However, the

current computer-aided diagnostic (CAD) systems for classifi-

cation between malignant and nevus lesions using the ABCD

criteria are imperfect due to use of ineffective computerized

techniques.

Methods: In this study, a novel melanoma recognition system

(MRS) is presented by focusing more on extracting features

from the lesions using ABCD criteria. The complete MRS sys-

tem consists of the following six major steps: transformation to

the CIEL*a*b* color space, preprocessing to enhance the

tumor region, black-frame and hair artifacts removal, tumor-

area segmentation, quantification of feature using ABCD crite-

ria and normalization, and finally feature selection and classifi-

cation.

Results: The MRS system for melanoma-nevus lesions is

tested on a total of 120 dermoscopic images. To test the per-

formance of the MRS diagnostic classifier, the area under the

receiver operating characteristics curve (AUC) is utilized. The

proposed classifier achieved a sensitivity of 88.2%, specificity

of 91.3%, and AUC of 0.880.

Conclusions: The experimental results show that the pro-

posed MRS system can accurately distinguish between malig-

nant and benign lesions. The MRS technique is fully automatic

and can easily integrate to an existing CAD system. To

increase the classification accuracy of MRS, the CASH pattern

recognition technique, visual inspection of dermatologist, con-

textual information from the patients, and the histopathological

tests can be included to investigate the impact with this sys-

tem.

Key words: melanoma – computer-aided diagnostic – der-

moscopy – pattern recognition – ABCD criteria

� 2012 John Wiley & Sons A/SAccepted for publication 26 April 2012

MALIGNANT MELANOMA (MM) is one of therare skin cancers (1) with an increasing

incidence rate and it is highly developed inhuman skin (2) because of uncontrolled growthof pigment cells, known as melanocytes. In theouter layer of the skin toward the lower part,melanocytes are present in the basal layer ofthe epidermis. In fact, the melanocytes constructa protein called melanin, which helps the skinto defend it by captivating ultraviolet (UV) radi-ation. The melanocytes are found in both whiteand non-white people (3). According to differ-ent skin places of humans, MM lesions (4) arecategorized as: Superficial spreading melanoma(SSM), Lentigo maligna melanoma (LMM),Acral lentiginous melanoma (ALM), Nodularmelanoma (NM), Mucosal melanoma (UM), andDesmoplastic melanoma (DM).

To diagnose melanoma and other pigmentedskin lesions (PSLs), dermoscopy is a non-inva-sive procedure (5). In clinical practices, derma-tologists use it to visualize the lesions moreclearly. Even though this device showed clearfeatures of skin lesions, but still, it is very hardto identify among different types of melanomasand PSLs, and even experienced dermatologists(6) have accuracy below 85%. Due to this rea-son, many melanoma cases are not diagnosedproperly. To clinically diagnose lesions, experi-enced dermatologists rely initially on patternrecognition, second on history, and later labora-tory parameters. Generally, physicians use clini-cal (7, 8) ABCD; Menzies’s method; 7-pointchecklist; and patterns classification methods todiagnose and classify the PSLs. Moreover, skinbiopsies (9) at an early stage are also necessary

1

Skin Research and Technology 2012; 0: 1–10Printed in Singapore � All rights reserveddoi: 10.1111/j.1600-0846.2012.00614.x

© 2012 John Wiley & Sons A/SSkin Research and Technology

to identify melanoma. Nowadays, medical diag-nostic cost is rapidly increasing, especially thecost of skin biopsies. Many better ways aresorted out to classify melanoma at an earlystage, which also provide the assistance ofdecreasing the number of skin biopsies. It wasexpected that computer-aided diagnostic (CAD)tools would help in identifying early MM (10)and support to trim down the death rate causedby this type of skin cancer.Automate CAD systems (11, 12) are developed

for providing ‘second opinion’ to dermatologists.Preprocessing, artifact removal, segmentation oflesion, ABCD (A: Asymmetry, B: Border, C:Color, and D: Differential structure) rule, patternrecognition, feature selection, and finally classifi-cation are the major stages of CAD system. Incurrent CAD systems, the features are extractedand defined according to the ABCD rule. TheABCD rule (13) is effectively researched in thelast few decades. It should be noticed that mela-noma could be recognized using effective fea-tures from ABCD, and pattern extractiontechniques (14) from lesion. The literature reviewsuggested that a lot of attention is paid todevelop a melanoma recognition system usingABCD rule, pattern analysis, or a combination ofboth techniques. However, the effectiveness ofcurrent systems is limited due to a variety oflesion shapes, irregular boundaries, specularreflection, low contrast, artifacts, and ineffectiveextraction of ABCD attributes. In fact, the ABCDcriterion is indicated as a general risk of malig-nancy that imposed strict boundary for decision.To distinguish between melanoma and benignlesions, this criterion appears less valuable to beutilized as features in the computer-based classi-fication algorithms. As a result, the detection ofmelanomas using ABCD criteria for computer-based algorithms remains a challenging task.In this study, effective computerized methods

are presented to extract ABCD parameters forthe classification of skin lesions. Our mainobjective is to develop an effective and com-plete melanoma recognition system (MRS) solu-tion based on the attributes of ABCD. The MRSsystem consists of six consecutive major steps:color space transforms, preprocessing, black-frame and hair artifacts removal, lesion-areasegmentation, feature quantification and nor-malization, and finally the feature selection andclassification. In the color space transform step,first RGB color dermoscopy image is converted

into the approximately uniform CIEL*a*b* colorspace (15). In the preprocessing step, unevenillumination correction and contrast enhance-ment procedures are applied using homomor-phic transform filtering (HTF) (16) technique. Inthe black-frame and hair artifacts removal step,a linear scan technique is developed and hairare detected by derivate of gaussian (DOG),morphological edge detection, and laterrepaired by fast marching schemes (17). After-wards, the lesion area is segmented using skintumor area extraction (STEA) (18) technique. Inthe feature quantification and normalizationstep, the lesion segmented region is used toextract the features of ABCD and then constructthe feature vector, which is normalized usingnormal probability-density function. The nor-malized feature vector is calculated for eachdermoscopy image. After that the optimal fea-tures are selected using sequential floating for-ward selection (SFFS) technique. Finally, theclassification decision is performed using binarysupport vector machine (SVM) algorithm. Toperform experiments, the dataset of 120 imagesare obtained as a CD resource (19). The classifi-cation accuracy of MRS system is measured bythe area under the receiver operating character-istics (AUC) curve (20).

Materials and Methods

DatasetThe proposed MRS system based on clinicalABCD attributes is effectively performed on atotal of 120 lesions. In this dataset, each imageconsists of exactly 1 lesion and there are 60malignant and 60 nevi. The dataset wasacquired as a CD resource from the two Euro-pean university hospitals as part of the EDRA-CDROM, 2002 (19). In this dataset, there are1064 color dermoscopic images with spatial res-olution of 768 9 512 pixels. During routine clin-ical assessments, all the images in this datasetwere captured to imitate a priori probabilitiesof the diagnosis. The detailed description of theselected 120 images is presented in Table 1.Light specular reflection and hair artifact arepresented in most of the images in the dataset.

MethodThe overall proposed melanoma recognitionsystem (MRS) consists of six consecutive stages:

2

Abbas et al.

color space transforms, preprocessing, black-frame and hair artifacts removal, lesion-areasegmentation, feature quantification and nor-malization, and finally feature selection andclassification. This system is mainly based ondefining the attributes from clinical ABCD rulethat is detailed in the following paragraphs.

Color space transformTo implement a melanoma recognition system(MRS) in the wide varieties of dermoscopy con-text, a uniform CIEL*a*b* color space is uti-lized. In 1976, CIE developed two color spacesthat approximately exhibited these characteris-tics which are L*a*b* and L*u*v*. Euclidean dis-tances in those spaces were shown to beapproximately correlated with perceptual colordifferences. In addition to that the white pointadaptation present in the CIEL*u*v* system,with a subtractive shift instead of a multiplica-tive normalization, can result in poor visual cor-respondence and also in predictedcorresponding colors lying outside the realiz-able gamut (15). Therefore, CIEL*a*b* perceptu-ally adaptive color space was used to accuratelypreprocess the skin lesions and features extrac-tion for effectively recognizing of melanomas.All RGB dermoscopy images were transformedfirst to CIEL*a*b* color space.

PreprocessingIn any type of dermoscopy image, the impor-tant step is to preprocess the skin image tofacilitate the subsequent stages. The preprocess-ing step is used to correct specular reflectionand enhance the contrast that facilitates the rec-ognition of skin lesion from dermoscopyimages. As an illumination problem is vital in

dermoscopy images, the contrast of somelesions is not clear, which results in unsatisfac-tory lesion segmentation. To correct unevenillumination and contrast enhancement, homo-morphic transform filtering (HTF) technique isused on the luminance (L*) plane of CIEL*a*b*color space. This technique is described in (16).Basically, it is based on the idea that illumina-tion and reflectance are combined through amultiplication operation to form an entireimage. Applying logarithms to the expression,the relationship between illumination andreflectance becomes additive and thus thesetwo components of the image can be linearlyseparated in the frequency domain. As the illu-mination component tends to vary slowlyacross the image while the reflectance tends tochange rapidly, the intensity variation acrossthe image can be reduced while highlightingthe details considering high frequency compo-nents as reflectance and low frequency compo-nents as illumination and then applying a highpass filter in the logarithmic domain. As aresult, an enhanced image is obtained. Theresult of dermoscopy image enhancement forinput image [Fig. 1(a)] is represented inFig. 1(b).

Black-frame and hair artifacts removalCircular or rectangular black-frame and hairartifacts are often presented in the dermoscopyimages. Therefore, to get an accurate lesion areaand effective features, these should be removedfrom the skin lesion images. To remove black-frame, a simple scanning method is developedbased on a linear search. This linear searchmethod is non-restricted and search across theboundaries of the image. From this scanning,the black pixels with zero intensity value areneglected so another image without black-frameis constructed. Figure 1(c) illustrated the resultof a black-frame removal algorithm.After removing the black-frame, the hair

pixels are removed using derivative of gauss-ian, morphological function, and fast marchingtechniques. In this study, an effective hair-removal algorithm was utilized (17) due to itsspeed, robustness and without affecting thelesion texture. Removing hair stage consists oftwo steps: the detection of hair and repairingof detected hair using image inpainting tech-niques. In the hair detection step, the deriva-

TABLE 1. . The detailed description of selected dataset that consists of atotal of 120 dermoscopy images of two major categories melanoma andnevus

Lesions type

Attributes

Number of lesions Diameter

Melanoma 60 8.78

Blue nevus 10 6.1

Clark nevus 15 5.2

Combined nevus 10 4.3

Congenital nevus 15 5.4

Dermal nevus 10 3.78

3

Melanoma recognition system

tive of Gaussian (DOG) is used for initial seg-mentation, and after that the refinement ofdetected lines are performed using morpho-logical edge-based techniques. To detect hairregions, luminance component of CIEL*a*b*enhanced image was used. Afterwards, non-iterative inpainting technique is utilized torepair the hair-occluded information from der-moscopy images. The effective hair detectionand removal steps are illustrated in Fig. 1(d)and 1(e), respectively.

Lesion-area segmentationTo recognize melanoma or to analyze theproperties of ABCD, the lesion area should bedetected. In this process, the lesion is sepa-rated from the surrounding skin. In the litera-ture, there are many lesion segmentationalgorithms developed that are based on edge,region, or active contour or simple threshold-ing technique. However, an automatic lesionsegmentation step is very difficult due toirregular or fuzzy tumor border. Therefore, inthis study, the lesion segmentation step is per-formed using skin tumor area extraction(STEA) (18) algorithm. In STEA algorithm, animproved dynamic programming (IDP)method is developed for color image segmen-tation based on color and texture features. Ini-tially, the rough segmentation is performedusing minimum error thresholding technique

on the luminance (L*) image. After that IDPtechnique is applied to refine this segmenta-tion. For a detailed explanation of this step,the readers may refer to (18). The segmenta-tion results using IDP algorithm is shown inFig. 2.

Feature quantification and normalizationThe main objective of this research is to investi-gate the influence of clinical ABCD attributeson the computerized recognition of melanoma.For this reason, effective image processing algo-rithms are used to measure the attributes ofABCD from segmented tumor region. Fromeach lesion, asymmetry, border quantification,color, and differential structure features aremeasured and construct a feature to describelesion. After that normalization on this featurevector is performed based on normal probabil-ity-density function.

Asymmetry featureThe asymmetry feature is examined based onboth x-axis and y-axis. In terms of borders, col-ors, and differential structures, the lesion isdivided by these axes that are positioned to cre-ate the smallest possible asymmetry. To mea-sure the lesion asymmetry (11), first centroidlocation, geometric and central moments of thelesion features are calculated. In this study, first

(a)

(d) (e)

(b) (c)

Fig. 1. Black-frame removal, enhancement, and hair-pixels repair results of malignant melanoma dermoscopy image, which show (a) the originalimage, (b) illumination corrected image, (c) black-frame removal image, (d) hair detected, and (e) repaired images.

4

Abbas et al.

order moments are used to find the centralmoments along x and y-axes. Two values arecalculated from Asymmetry measurement alongboth axes. First value comes from the major axisorientation of the segmented lesion (h), whichwas calculated from tangent inverse angle byaveraging the geometrical and central moments.After that the mass region was rotated hdegrees clockwise to align the ellipse (majorand minor) axes with the image (x and y) axes.The lesion area was then folded about the x-axisand the area difference between the overlap-ping folds was taken as the amount of asymme-try about the x-axis. Accordingly, to calculateasymmetry along y-axis, the same techniquewas applied.

Border quantification featureDifferent attributes related to border quantifi-cations or descriptions from the segmentedregion are measured. Border description anal-ysis (21), B in the ABCD rule of dermatosco-py, can be easily evaluated to extract the setof numerical features based on the lesionstructure related attributes and border irregu-larity.The attributes of lesion structure, such as

equivalent circular diameter (ECD), mean inten-sity variance (Mv), area (Ar), and Centroid (Cn)with radius (Rd) of each lesion border regionare effectively measured. To measure the Area(Ar) in pixels and centroid (Cn) location oflesion, the method of bit quads (23) is followed.The equivalent of circular diameter (ECD) oflesion area is calculated by the equation:ECD = [(4 9 Ar)1/2/p], and radius by RD =ECD/2.0. However, circularity measure is thecritical step for measuring ECD value. In thepast studies, the circularity is not perfectlyobtained although this measure is critical toquantify the border-irregularity, which highlydepends on the image resolution. Therefore, thepresent approach uses the circularity measure

from (24), which improves the state of the artmeasures. The mean intensity variance (Mv) oflesion is extracted from the (L*) plane of theCIEL*a*b* color space.Border-irregularity measure is one of most

the important clinical features to differentiatebetween malignant melanomas and benignnevi. Four parameters related to border irreg-ularity, such as compactness indexes (CI),fractal dimension (FD), edge abruptness (EA),and pigmentation transition (PT) are com-puted. The Compactness Index (CI) measuredthe roundness of lesion, which is calculatedbased on the circularity (22). Fractal Dimen-sion (FD) parameter is used to measure themalignancy of lesions by the fact that if thevalue of FD increases, the probability of thelesions of being melanoma increases too. Forthis purpose, the box counting technique (25)is used. Pigmentation Transition (PT) parame-ter is calculated from L* plane of CIEL*a*b*color space. Then, the maximum gradientmagnitude is calculated along the lesionboundary. A benign lesion often exhibits grad-ual fading of the pigmentation, whereas sharpabrupt edge suggests malignancy (26). The EAparameter is calculated by measuring the gra-dient magnitude in the combined image of a*and b* planes.

Color featureTo recognize early melanoma or other pig-mented skin lesions (PSLs), the extraction ofcolor feature (11) plays an important role. Thesix shades of color and their symmetry are cal-culated over the lesion region for effective dif-ferentiation between melanomas and nevi. Infact, light-brown, dark-brown, white, red, blue,and black shades of color are extracted by thedermatologists to diagnosis the PSLs lesions.Therefore, the combination of six colors andtheir symmetry parameters are used to measurethe color features.

(a) (b) (c)

Fig. 2. The skin tumor area extract (STEA) algorithm by an improved dynamic programming, where (a) represents the preprocess image, (b) thesegmentation image, and (c) border mask image.

5

Melanoma recognition system

The spatially adaptive six shades of domi-nant colors are measured first from the seg-mented lesion area. The iterative k-meansclustering algorithm is utilized with a fixedsix number of clusters to measure the domi-nant colors. Every pixel of the image is repre-sented by a color in this clustering technique,which is identical to the average color of thepixels in its neighborhood that belong to thatclass. As a result, this k-means algorithm pro-vided the six locally adapted dominant colorsand the corresponding percentage of occur-rence (%occr) of each color within a certainneighborhood. In addition to that color sym-metry of lesion area is also calculated becauseit is an important feature to measure pigmentdistribution in a certain neighborhood. Forcalculating this feature, a technique based onperceived color differences is utilized that cor-responds to distances (#dist) among color pix-els in the CIEL*a*b* uniform color space. Anaverage value (distc) is used, which is calcu-lated from this distances (#dist) measure.Finally, %occr and distc values are used forthe definition of color feature.

Differential-structure featureA number of differential-structural componentsare usually presented in a lesion, which is also animportant feature. Pigment network, dots, brownglobules, and branched streaks are the structuralcomponents that are determined in this study. Itshould be noticed that greater the number of dif-ferential structures, the higher the probability ofthe lesion being a melanoma. Pigment networkand dots type differential structures are deter-mined by Radon-like transform function (27),whereas the brow globules and branched streaksare analyzed using local binary pattern (LBP)(28) technique. The Radon-like transform func-tion and LBP techniques are applied on the lumi-nance (L*) plane of the CIEL*a*b* color space toquantify the differential structures.Radon-Like features are derived from image

intensities to present a way of aggregating anydesired information. For any input imageregion R(x,y), the knots (derived using cannyedge map) and the extraction function use thefollowing transform function T(x,y):

Tðx; yÞ ¼ maxr;u

DGðr;uÞ � Rðx; yÞ ð1Þ

where, r and u are the scales and orienta-tion of the second derivate of gaussian filterDG(r, u) with a convolution operator (⊗ ).The knots for Radon-Like features are definedusing an edge map of T(x, y) and the extrac-tion function, which is defined in the study.The transformation T(x, y) captures responseof the most dominant GSD filter at each pixel,across various scales and orientations. Forquantification of dots and pigment network,three scales and four orientations (0o, 45o, 90o,and 135o) are used, and then the mean ofRadon-like features are accumulated. Afterthat morphological edge detection and con-nected-component labeling algorithms are usedto obtain the quantification values of thesestructures. The diameter (Dia), mean (Mean),and variance intensity (Var) variances areextracted on each connected object on the seg-mented area of lesion.The local binary pattern (LBP) is widely

used for pattern analysis. In this technique, abinary code is measured by thresholding acircularly symmetric neighborhood with thevalue of the center pixel for each pixel in animage. After that a histogram is created toaggregate all the occurrences of different bin-ary patterns. We are using the eight -neigh-bors of a pixel for determining the LBPoperator. To quantify local texture interactions,the LBP histogram approach is applied toeach decomposed part of the segmentedlesion. The mean LBPk

l and variance LBPkr of

LBP histograms of sliding window (w) of size16 9 16 are calculated to combine texturecharacteristics that are given in Eq. (2).Finally, the mean value of each histogram is

extracted.

f localLBPl;r¼ [k

i¼1 [ki¼1 LBPk

l;LBPkr

n oð2Þ

All the features which are extracted from thelesion area are combined into a single featurevector, such as:fi ¼major;min or;ECD;Mv;Ar;Cn;RD;CI; FD;EA;PT;

%occr;distc; [ki¼1 [k

i¼1 LBPkl; LBP

kr

n o;Dia;Mean;Var

" #

ð3Þ

To normalize the feature vector (fi), the nor-mal probability-density function is used totransform this vector into zero mean and unit

6

Abbas et al.

variance. This normalized feature vector isdenoted by f inorm.

Feature selection and classificationThe normalized feature vector (f inorm) consists ofsome potentially related features to differentiatebetween benign and malginant lesions. In thisfeature vector, there are some attributes that arenot equally valuable for this classification task(11) due to irrelevancy or redundancy. In addi-tion to that it is a very complicated task toexplicitly identify the patterns in the large data-set. Therefore, it is necessary to use some fea-ture selection (FS) algorithm to determine theoptimal and most relevant features for classifi-cation. In feature selection, a subset of featurescan be selected either using sequential floatingforward or backward selection (SFFS, SBFS)methods with cross-validation (XVAL) or leave-one-out schemes, incremental stepwise tech-nique with the hypothesis test of Wilks’ lambdaand plus-2 minus-1 method with objective func-tion.In this study, the SFFS method was adopted

for feature selection. Sequential ForwardSelection is the simplest greedy search algo-rithm. Sequential Floating Forward Selection(SFFS) starts from the empty set and after eachforward step, SFFS performs backward steps aslong as the objective function increases. Thisstep was performed only on the trainingdataset. When the most reliable and rele-vant features ðf1norm; f2norm; f3norm; . . .; fnnormÞ areselected, a classification function classifyðf1norm; f2norm; f3norm; . . .; fnnormÞ computes the decisioni.e., class 2 malignant; nevusf g lesion. To performclassification decision for melanoma-nevus, asupport vector machine (SVM) learning algo-rithm was used. For training and testing of mel-anoma and nevus classes, the SVM classifierused the normalized feature vector. The 40% ofdataset are used for training while 60% of themare used for testing the SVM classifier. Inpractice, the SVM classifier is used to assign theclass label to which the unknown instancebelongs to using binary SVM classification algo-rithm. To evaluate the performance of the SVMclassifier, leave-one-out cross-validation and 10-fold cross-validation are used that divide thedataset in the above mentioned ratio. Duringthis calculation, the sensitivity: SE and specific-ity: SP values are also measured and ROC

curve is also plotted on the basis of SE and SPvalues. Accordingly, the area under the receiveroperating curve (AUC) was also calculated. Thevalue of AUC is ranges from 0 to 1 and thehigher the AUC value the greater the classifica-tion accuracy.

Results and Discussion

To measure the performance of the MRS sys-tem, the experiments are performed on a data-set of 120 skin lesions. In this dataset, eachskin lesion image is arranged with known cate-gories. For experiments, 120 dermoscopyimages are consisted of malignant: 60, Bluenevus: 10, Clark nevus: 15, Combined nevus:10, Congenital nevus: 15, and Dermal nevus: 10lesions. To perform an effective melanoma-nevus classification, this collection is first pre-processed to enhance the contrast by homomor-phic transform filtering (HTF). After that hairpixels are detected by derivative of gaussian,morphological edge detection operations, andreplaced by a fast marching inpainting tech-nique. A skin tumor area extraction (STAE)algorithm is then applied based on improveddynamic programming (IDP) algorithm fordetecting of skin tumor region only. Next, theexpert features are extracted to define the clini-cal ABCD attributes using effective image pro-cessing techniques and then construct thefeature vector, which is normalized using nor-mal probability-density function. The normal-ized feature vector is calculated for eachdermoscopy image. After that the optimal fea-tures are selected using sequential floating for-ward selection (SFFS) technique. Finally, theclassification decision is performed using sup-port vector machine (SVM) algorithm.To make classification decision between mela-

noma-nevus, the entire collection of 120 imagesis divided into 40% of training and 60% of test-ing sets. The effectiveness of the MRS systemcould be tested using a query skin lesion. Aftertraining the melanoma-nevus SVM classifier, asearch lesion is considered to be a correct matchif it belongs to the same category to which thequery image belongs. To evaluate the perfor-mance of the SVM classifier, leave-one-outcross-validation and 10-fold cross-validation areused that divide the dataset in the above men-tioned ratio. During this calculation, the sensi-tivity: SE and specificity: SP values are also

7

Melanoma recognition system

measured and ROC curve is also plotted on thebasis of SE and SP values. Accordingly, the areaunder the receiver operating curve (AUC) wasalso calculated.The MRS technique was preliminarily imple-

mented in Matlab 7.6.0.324® (The Mathworks,Natick, MA, USA). The detailed description ofaverage running time consumed on each com-putation step is presented in Table 1. Toenhance the contrast of skin lesion, on average,the preprocessing step took 3.5 s, the hair-detec-tion technique took 4.3 s and repairing steptook 2.8 s. The lesion skin-area segmentationtechnique took 1.4 s. The feature extraction andnormalization steps on average took 6.3 and 0.6s, respectively. For training the melanoma-nevus SVM classifier on this dataset, a total of4.65 s is consumed for two classes. However, incase of testing a new input skin lesion, on aver-age 5.49 s is taken. For the most part of thecomputation, time is spent on the featureextraction and training the SVM classifier formelanoma-nevus classes. The computation timecan be decreased by means of using optimizedC/C++ implementation. All computations wereexecuted on a 2.0 GHz Core to Duo 32-bit Intelprocessor with 1 GB DDR2 RAM, running Win-dows 7 operating system.Melanoma-nevus binary classifier is applied

to 120 lesions in division of 40% of training and60% of testing tests. The classification results ofboth categories are shown in Table 2. In thistable, the sensitivity, specificity, and AUC mea-sures are shown. It should be noticed that thebest classification results are obtained i.e., incase of nevus (AUC: 0.824) and melanoma(AUC: 0.880) type lesions. The area under thecurve (AUC) value is high, which indicates thatthe melanoma-nevus classifier has a good classi-fication accuracy. The Fig. 3 has shown the cor-responding receiving operating characteristiccurve (ROC) of SVM and K-NN machine learn-ing algorithms in terms of melanoma and nevusclasses. The comparisons between SVM andK-NN are also performed to compare the classi-fication accuracy using these two classifiers. Incase of K-NN classifier, the classification resultsare obtained, such as AUC of 0.730 and AUC of0.74 for nevus and melanoma lesions, respec-tively.The melanoma-nevus classification accuracy

for the SVM classifier based on ABCD is notvery high, which might be due to the absence

of CASH attributes from the sample dataset.The attributes of ABCD and CASH may be inte-grated into a single system to increase the per-formance level of melanoma recognition.Although, the ABCD rule is only one of theimportant components in distinguishingbetween these two groups, we obtained classifi-cation accuracy that is almost 90%. The mainobjective of this study is to develop effectivecomputerized methods for melanoma recogni-tion based on ABCD attributes for the skinlesions based on dermoscopy images. In thisstudy, all six stages for MRS are integrated andtested, such as color space transforms, prepro-cessing, black-frame and hair-artifacts removal,lesion-area segmentation, feature extraction andnormalization, and finally feature selection andclassification. If we combined the attributes ofABCD and CASH then it could increase theperformance level of Computer-aided detection(CAD) of lesions or content-based image retrie-val (CBIR) systems. As a result, for melanomarecognition, the combination of attributes of

TABLE 2. . Performance measure of melanoma-nevus classifier usingAUC of SVM machine learning algorithm based on 120 dermoscopyimages

Skin lesions

Performance measures

Sensitivity Specificity AUC1

Melanoma 88.2 91.30 0.880

Nevus 86.5 88.2 0.824

AUC = Area under the receiver operating curve.

Fig. 3. Receiver operating characteristics (ROC) curve for the mela-noma-nevus SVM and KNN learning algorithm on 120 dermoscopyimages.

8

Abbas et al.

ABCD and CASH into a single classificationsystem is our future plan.

Conclusions

In this study, melanoma recognition system(MRS) based on clinical ABCD technique is pro-posed for the automatic classification of pig-mented skin lesions. The MRS system involvesvarious steps: transformation to the CIEL*a*b*color space, preprocessing to enhance the tumorregion, artifact removal, tumor-area segmenta-tion, quantification of feature using ABCD attri-butes and normalization, and finally featureselection and classification. The MRS system istested and evaluated for melanoma-nevus clas-sification using dermoscopic images. The exper-imental results show that the proposedmelanoma recognition system is highly accurateto classify between melanoma and nevuslesions. The MRS method is fully automatic andcan easily integrate into an existing CAD sys-tem. The main objective of this research is todemonstrate how the effective computerized

methods can be integrated into a single classifi-cation system to assist the dermatologists forthe melanoma recognition. However, it is pre-dictable that many other advanced featuresusing CASH and features from other sourcescould be important to develop a complete com-puter-aided diagnostic (CAD) system. In thefuture study, we plan to incorporate more fea-tures related to the diagnostic relevance intoMRS system. Ideally, the presence of an expertdermatologist is also required, who can performoverall visual inspection of the skin lesion andsuggests the final diagnosis which is composedof objective evaluation by the MRS system andfrom the patient data and the histopathologicaltests.

Acknowledgments

This study was supported by the National Tex-tile University Faisalabad-37610 and highereducation commission (HEC) of Pakistan withgrant no. PM-IPFP/HRD/HEC/2010/2395.

References

1. Rigel DS. Malignant melanoma:perspectives on incidence and itseffects on awareness, diagnosis,and treatment. CA Cancer J Clin1996; 46: 195–198.

2. Kobayashi N, Muramatsu T, Ya-mashina Y, Shirai T, Ohnishi T,Mori T. Melanin reduces Ultravio-let-Induced DNA damage forma-tion and killing rate in culturedhuman melanoma cells. J InvestDermatol 1993; 101: 685–689.

3. Crombie IK. Racial differences inmelanoma incidence. Br J Cancer1979; 40: 185–193.

4. Swetter SM, Boldrick JC, Jung SY,Egbert BM, Harvell JD. IncreasingIncidence of Lentigo Maligna mela-noma subtypes: Northern Califor-nia and National trends 1990–2000.J Invest Dermatol 2005; 125: 685–691.

5. Ishihara Y, Saida T, Miyazaki A,et al. Early acral melanoma in situ:correlation between the parallelridge pattern on dermoscopy andmicroscopic features. Am J Derma-topathol 2006; 28: 21–27.

6. Argenziano G, Soyer HP, ChimentiS, et al. Dermoscopy of pigmentedskin lesions: results of a consensus

meeting via the Internet. J AmAcad Dermatol 2003; 48: 679–693.

7. Johr RH. Dermoscopy: alternativemelanocytic algorithms—the ABCDrule of dermatoscopy, menzies scor-ing method, and 7-point checklist.Clin Dermatol 2002; 20: 240–247.

8. Blum A, Luedtke H, Ellwanger U,Schwabe R, Rassner G, Garbe C.Digital image analysis for diagno-sis of Cutaneous melanoma. Devel-opment of a highly effectivecomputer algorithm based on anal-ysis of 837 melanocytic Lesions. BrJ Dermatol 2004; 151: 1029–1038.

9. Friedman RJ, Rigel DS, Kopf AW.Early detection of malignant mela-noma: the role of physician exami-nation and self-examination of theskin. CA Cancer J Clin 1985; 35:130–151.

10. Andreassi U, Perotti R, Rubegni P,et al. Digital dermoscopy analysisfor the differentiation of atypicalNevi and early melanoma. ArchDermatol 1999; 135: 1459–1465.

11. Celebi ME, Kingravi HA, Uddin B,et al. A methodological approachto the classification of dermoscopyimages. Comput Med Imag Grap2007; 31: 362–373.

12. Rubegni P, Risulo M, Burroni M.Reply to digital dermoscopy analy-sis and Internet-based program for

discrimination of pigmented skinlesion dermoscopic images. Br JDermatol 2006; 154: 571–572.

13. Nachbar F, Stolz W, Merkle T,Cognetta AB, Vogt T, LandthalerM, Bilek P, Braun-Falco O, PlewigG. The ABCD rule of dermatosco-py: high prospective value in thediagnosis of doubtful melanocyticskin lesions. J Am Acad Dermatol1994; 30: 551–559.

14. She Z, Liu Y, Damatoa A. Combi-nation of features from skin pat-tern and ABCD analysis for lesionclassification. Skin Res Technol2007; 13: 25–33.

15. Green P, MacDonald L. Color engi-neering. Achieving device inde-pendent colour. West Sussex,England: Wiley; 2003.

16. Abbas Q, Celebi ME, Fondon I,Rashid M. Lesion border detectionin dermoscopy images usingdynamic programming. Skin ResTechnol 2011; 17: 91–100.

17. Abbas Q, Celebi ME, Garcıa IF.Hair removal methods: a compara-tive study for dermoscopy images.Biomed Signal Process Control2011; 6: 395–404.

18. Abbas Q, Celebi ME, Garcıa IF.Skin tumor area extraction usingan improved dynamic program-

9

Melanoma recognition system

ming approach. Skin Res Technol2011; 18: 133–142.

19. Argeniano G, Soyer PH, De VG,Carli P, Delfino M. Interactive atlasof dermoscopy CD. Milan: EDRAmedical publishing and Newmedia, 2002.

20. Bradley AP. The use of the areaunder the ROC curve in the evalu-ation of machine learning algo-rithms. Pattern Recogn 1997; 30:1145–1159.

21. Grana C, Pellacani G, Cucchiara R,Seidenari S. A new algorithm forborder description of polarizedlight surface Microscopic Imagesof Pigmented skin lesions. IEEE TMed Imaging 2003; 25: 959–964.

22. Lee TK, McLean DI, Atkins MS.Irregularity Index - a new borderirregularity measure for cutaneousmelanocytic lesions. Med ImageAnal 2003; 7: 47–64.

23. Yang L, Albregtsen F, LonnestadT, Grottum P. Proceedings of the

IAPR workshop on machine visionapplications. Kawasaki, Japan; 1994:272–276.

24. Haralick RM. A measure for circu-larity of digital figures. IEEE TransSyst Man Cybern 1974; SMC-4: 394–396.

25. Sarkar N, Chaudhuri BB. An effi-cient differential box-countingapproach to compute fractaldimension of image. IEEE TransSyst Man Cybern 1994; 24: 115–120.

26. Hintz-Madsen M, Hansen LK, Lar-sen J, Drzewiecki KT. A probabilis-tic neural network framework fordetection of malignant melanoma.In: Naguib RNG and Sherbet GV,eds. Artificial neural networks incancer diagnosis, prognosis, andpatient management. Boca Raton:CRC Press; 2001: 141–183.

27. Kumar R, Reina AV, Pfister H.Proceedings of the IEEE ComputerSociety Workshop on Mathemati-

cal Methods in Biomedical ImageAnalysis (MMBIA). San Francisco,CA: IEEE Computer Society; June;2010.

28. Ojala T, Pietikainen M, MaenpaaTT. Multiresolution gray-scale androtation invariant texture classifica-tion with local binary pattern.IEEE Trans PAMI 2002; 24: 971–987.

Address:Dr Qaisar AbbasDepartment of Computer ScienceNational Textile University37610, FaisalabadPakistanTel: +92 41 9230081 Ext: 140Fax: +92 (41) 9200764e-mail: qaisarabbasphd@gmail.com;drqaisar@ntu.edu.pk

10

Abbas et al.