Research ArticleComputer-Aided Diagnosis of Micro-Malignant MelanomaLesions Applying Support Vector Machines

Joanna Jaworek-Korjakowska

Department of Automatics and Biomedical Engineering AGH University of Science and TechnologyAleja Mickiewicza 30 30-059 Krakow Poland

Correspondence should be addressed to Joanna Jaworek-Korjakowska jaworekaghedupl

Received 23 December 2015 Revised 15 April 2016 Accepted 26 April 2016

Academic Editor Yudong Cai

Copyright copy 2016 Joanna Jaworek-KorjakowskaThis is an open access article distributed under theCreativeCommonsAttributionLicense which permits unrestricted use distribution and reproduction in anymedium provided the originalwork is properly cited

Background One of the fatal disorders causing death is malignant melanoma the deadliest form of skin cancer The aim of themodern dermatology is the early detection of skin cancer which usually results in reducing the mortality rate and less extensivetreatment This paper presents a study on classification of melanoma in the early stage of development using SVMs as a usefultechnique for data classification Method In this paper an automatic algorithm for the classification of melanomas in their earlystage with a diameter under 5mm has been presented The system contains the following steps image enhancement lesionsegmentation feature calculation and selection and classification stage using SVMs Results The algorithm has been tested on200 images including 70 melanomas and 130 benign lesions The SVM classifier achieved sensitivity of 90 and specificity of 96The results indicate that the proposed approach captured most of the malignant cases and could provide reliable information foreffective skin mole examination Conclusions Micro-melanomas due to the small size and low advancement of development createenormous difficulties during the diagnosis even for experts The use of advanced equipment and sophisticated computer systemscan help in the early diagnosis of skin lesions

1 Introduction

Melanoma is a less common butmost dangerous form of skincancer It starts in the melanocytes or pigment producingcells found in the outer layer of the skin when unrepairedDNA damage to skin cells (most often caused by ultravioletradiation from sunshine or tanning beds) triggers mutations(genetic defects) that lead the skin cells to multiply rapidlyand form malignant tumors (Figure 1) These cells grow outof control and form a tumorMelanomas are fast-growing andhighlymalignant tumors often spreading to the nearby lymphnodes lungs and brain [1] Malignant melanoma is likely tobecome one of the most common malignant tumors in thefuture with even a ten times higher incidence rate Sincethe early 1970s malignantmelanoma incidence has increasedsignificantly for example in the USA it grows approximatelyby 4 percent every year [1] The main environmental reasonfor the rapid growth of melanoma cases is the excessive expo-sure to ultraviolet (UV) from the sun and sunbeds Experts

estimate about 90 of melanomas are associated with severeUV exposure and sunburns over a lifetime [2]

Due to the high skin cancer incidence dermatologi-cal oncology has become a quickly developing branch ofmedicine Nowadays the progress is visible both in primaryresearch concerning pathogenesis of tumors (the role of genesor viruses in tumor development) and in the development ofnew more efficient methods of computer-aided diagnosis [1]

This paper is organized in 4 sections as follows Sec-tion 1 Introduction presents the skin cancer awareness anddescribes the clinical definition of micro-melanoma skinlesions motivation of the undertaken research and relatedworks Section 2 Materials andMethods specifies the steps ofthe designed computer system including dermoscopy imagepreprocessing skin lesion segmentation feature calculationand selection and SVM classification method In Section 3Results and Discussion the conducted tests and resultsare described Section 4 Conclusions closes the paper andhighlights future directions

Hindawi Publishing CorporationBioMed Research InternationalVolume 2016 Article ID 4381972 8 pageshttpdxdoiorg10115520164381972

2 BioMed Research International

Four stages of melanoma

Epidermis

Dermis

Subcutaneoustissue

Stage IDisease only inskin very thin

Stage IIISpread to lymph nodes

Stage 0Melanoma confined toepidermal region of skin

Stage IILocalized diseasethicker than stage I

Stage IVDisease spread toother organs

Figure 1 Melanoma is staged in four segments depending upon the spread of the disease Stage 0 is also called melanoma in situ themelanoma cells are only in the outer layer of the skin (the epidermis) Stages I and II mean there are melanoma cells in the layer directlyunder the epidermis or touching on the next layer down but there is no sign that it has spread to lymph nodes or other parts of the body [11]

(1) Clinical Importance and Motivation The most importantparameter which predicts the stage of melanoma is the thick-ness of the examined lesion (Figure 1) Skin moles with thethickness less than 1mm are near 100 curable [1] Seidenariand coworkers reported the direct correlation between diam-eter and thickness and as the diameter of in situ melanomasis smaller than that of invasive lesions it is reasonable tobelieve that small melanomas are usually in an initial growthphase [3] Therefore the aim of each clinician is to detectmalignant melanomas when they are still small and thin

Micro-melanomas represent a minority of diagnosedlesions their frequency ranging from 1 to 17 The meandiameter of in situ melanomas is around 1 cm and invasivemelanomas are usually gt6mm [3ndash6] In our research we con-centrate on micro-melanomas with the diameter lower than5mm

In the light of the above data prevention and earlydiagnosis of melanoma become extremely important issuesMicro-melanomas due to the small size and low advancementof development create enormous difficulties during the diag-nosis even for experts Such lesions are so little that they areclose to the limit of clinical relevance

Nowadays themedical checkup of smallmoles is possiblethrough a dermoscope (also called digital epiluminescencemicroscopy or ELM) which consists of a magnifier (typicallytimes10) a nonpolarised or polarised light source a transparentplate and a liquidmedium (for nonpolarised light) It enablesperforming a noninvasive in vivo examination making sub-surface structures of the skin visible (Figure 2)

The dermoscopic diagnosis of pigmented skin lesionis based on the assessment of the presence or absence ofdifferent global and local features Various analytic methodsand algorithms have been set forth in the last 25 years Themost important and widely used are pattern analysis ABCDrule of dermoscopy 7-point checklist and Menzies method[1 7]The classification of the obtained parameters is the finalstep of the diagnosis process where the extracted features are

Figure 2 Dermoscopic images of micro-melanoma lesions (basedon [1])

utilized to ascertain whether the lesion is malignant suspi-cious or benign

(2) Related Works Different groups of researchers havefocused on the extraction and classification of significantfeatures to provide distinction between benign andmalignantskin lesions however when dealing with micro-melanomasonly few works have been published but only in the scientificliterature ofmedicine In 2004Bono and coworkers describeda clinical study of 22 cases of melanoma with the diameterlower than 3mmThey extracted geometric and dermoscopicfeatures A lesion was considered suspicious for melanomawhen at least one parameter has been positive [8] In theirwork they notice that the ABCD diagnostic algorithmbecomes redundant while evaluating micro-melanomas andthat traditional dermoscopic diagnostic methods often failIn 2006 Bono et al published another work about micro-melanoma lesions based on more than 200 medical cases [9]The clinical evaluation gave a sensitivity of 43 and speci-ficity of 91 However using a dermoscope the sensitivityincreased to 83 while the specificity dropped to 69 Thedescribed above papers approach the subject only from theclinical point of view

During patient examinations researchers observe thatyoung inexperienced dermatologists and family physicianshave huge difficulties in the correct visual assessment of skin

BioMed Research International 3

Micro-melanomaimage

Preprocessing Segmentation Feature extractionFeature selection

Classification

Figure 3 Flowchart of the proposed automated image analysis system

lesions Due to this obstacle researchers try to implementand build computer-aided diagnosis (CAD) systems forautomated diagnosis of melanoma to increase the specificityand sensitivity and make the assessment of the skin molemore simple [10] Although such computer programs aredeveloped for different diagnostic algorithms to the best ofour knowledge a CAD system to classify micro-melanomalesions has not been proposed yet In our approach we pro-pose to divide skinmoles into twomain groups in terms of thediameter In this research we concentrate on the evaluationof significant and discriminant features as well as an accurateclassification process of only micro-melanoma skin lesions

2 Materials and Methods

After the introduction to the aim of this research we wouldlike to introduce the technical aswell as algorithmic solutionsThe first part presents the implemented diagnostic systemwith its particular steps In the second part we concentrateon the feature extraction and selection as well as the SVMclassification method

The implemented application is divided into four stagespreprocessing (image enhancement) skin lesion segmen-tation feature extraction and selection and classification(Figure 3) Since themain aimof this research is to present thefeature extraction and classification step the preprocessingand segmentation stages will be described shortly A detaileddescription of these steps can be found in our works [12ndash14]

21 Image Enhancement Dermoscopy images are often dete-riorated by noise due to various sources of interference andother phenomena that affect the image acquisition processImage enhancement techniques are used to refine the ana-lyzed image so that desired image features become easier toperceive and to be detected by the automated image analysissystem [15] Dermoscopic images are inhomogeneous andcomplex and furthermore they contain extraneous artifactssuch as skin lines air bubbles and hairs which appear inalmost every image Therefore the preprocessing step isessential for dermoscopic images to improve the quality andit is responsible for reducing the amount of artifacts andnoise The preprocessing step includes black frame removalsmoothing of the air bubbles and black hair inpainting

The black frame is detected on the basis of the light-ness component value in the HSL color space In order to

determine the darkness of a pixel with (119877 119866 119861) coordinatesthe lightness component of the HSL color space is calculated[12]

119871 =

max (119877 119866 119861) +min (119877 119866 119861)

2

(1)

A pixel is considered to be black when the lightness value isless than 15 (range of the lightness value [0 255]) For thesmoothing of air bubbles and light hairs we use the Gaussianfilter

The removing of thick black hair is the most importantstep in dermoscopy image preprocessing since it influencesthe segmentation and feature extraction step Firstly the der-moscopic RGB image is being converted into grayscale withthe NTSC 1953 standard Secondly the black top-hat trans-form which is a morphological image processing techniqueis used to detect thick dark hairsThe result of this step is thedifference between the closing operation and the input image

119879119908(119868) = 119868 ∘ 119887 minus 119868 (2)

where ∘ denotes the closing operation 119868 is the grayscale inputimage and 119887 is a grayscale structuring element [15] The laststep is the hair inpainting Hair line pixels are replaced withvalues calculated on the basis of the neighborhood pixels(Figure 4)

22 Segmentation Algorithm The principal goal of the seg-mentation process is to partition an image into regions thatare homogeneous in terms of pixel intensity or specific fea-tures [15] Inmedical imaging segmentation is important andcrucial for feature extraction and image measurements andclassification Precision of classification of a tumor dependsgreatly on the accuracy of the extracted tumor area as mostof the features are calculated on this basis During our workwe have compared different segmentation methods [13 14]Based on the previous research the applied segmentationalgorithm for the skin lesion extraction is based on seededregion-growing algorithm The region-growing algorithmalso called region merging looks for groups of pixels thathave similar intensities whereas thresholding methods focuson the difference of pixel intensities [15]Themain advantageof the region-growing algorithm is that it is capable ofcorrectly segmenting regions that have the same propertiesand are spatially separated Because the healthy skin in der-moscopy images is mostly homogenous the region-growing

4 BioMed Research International

(a) (b)

Figure 4 Outcome of the preprocessing step (a) dermoscopic input image and (b) image after black frame removal and hair inpainting

(a) (b) (c)

Figure 5 Dermoscopic image segmentation using region-growing method (a) dermoscopic input image (b) image after enhancement and(c) segmented skin mole

algorithm can be applied to detect it For the skin lesion seg-mentation we select one seed which is located in the left cor-ner on the healthy skin It gives us the certainty that we willseparate the homogeneous background from the skin moleThe region is iteratively grown by comparing all unallocatedneighboring pixels to the regionThe region-growing processconsists of picking a seed from the set investigating all 4-connectedneighbors of this seed andmerging suitable neigh-bors to the seed The seed is then removed and all mergedneighbors are added to the seed set The region-growingprocess continues until the seed set is empty Figure 5presents the results of preprocessing and segmentation stepsfor micro-melanoma skin lesion

23 Dermoscopic ImageAnalysis After the preprocessing andsegmentation steps the dermoscopy image is prepared for adetailed analysisNow it is necessary to analyze the image andrecognize critical parameters that can indicate melanomaThese properties are basically the same ones that a physician

looks for In this research we concentrate on two of the mostinformative visual cues in dermoscopy image interpretationincluding shape and texture Shape parameters correspondwith the diagnostic algorithm called the ABCD rule of der-moscopy while the texture parameters correspond with thediagnostic algorithm 7-point checklist

231 Feature Extraction Many different image features havebeen proposed for the analysis of medical images includ-ing intensity geometrical morphological fractal dimensionbased and texture featuresThe feature extraction step can beconsidered as data compression that removes additional andirrelevant information while preserving relevant informationin a data vector Features selected for the calculation ofdetected skin moles are based on the knowledge and expe-rience about the properties and differences in appearancebetween the benign and malignant skin lesions However tofind out the most redundant and effective features we extracta large number of initial features The proposed features are

BioMed Research International 5

identifiable translatable rotation and scale invariant noiseresistant and reliable

Shape Features (Geometrical Features) Shape-based imageretrieval consists of measuring of similarity between shapesrepresented by their features Shapes that have large differ-ences can be identifiedwith the geometrical features howevereven small changes can affect the final value of the featureWe calculated parameters including major and minor axeslength eccentricity (measure of aspect ratio it is the ratioof the major axis to the minor axis) elongation (based onthe bounding box) compactness (normalized version byHaralick) ellipse variance rectangularity (ratio area of shapeand area of the minimum bounding box) convexity (ratio ofperimeter of convex hull and contour) and solidity (ratio ofarea of convex hull and shape area)

Asymmetry Asymmetry is the major parameter in the diag-nostic algorithm ABCD rule of dermoscopy We quantifyasymmetry calculating the symmetry distance (SD) proposedby NG in 1997 [16 17] It is the calculated result of typical dis-placement among the number of verticeswhen the skin lesionshape is transformed into its symmetric shape [18] The SDvalue is measured as

SD =

1

119899

119899minus1

sum

119894=0

10038171003817100381710038171003817119875119894 minus 119894

10038171003817100381710038171003817 (3)

Color Variegation A skin lesion may have a variety of colorsincluding tan dark brown black blue red and occasionallylight gray Melanomas are usually characterized by three ormore colors and in about 40 of melanomas even five or sixcolors are present We analyze the color variation throughfour different channels including red green blue and inten-sity The intensity channel for each pixel 119868(119909 119910) is defined

119868 (119909 119910) = radic1198772(119909 119910) + 119866

2(119909 119910) + 119861

2(119909 119910) (4)

The following parameters are calculated for assessing thecolor variegation [7]

cv1= log 120590 (119877)

120583 (119877)

cv2= log 120590 (119866)

120583 (119866)

cv3= log 120590 (119861)

120583 (119861)

cv4= log 120590 (119868)

120583 (119868)

(5)

where 120583 is the mean value and 120590 denotes standard deviation

Texture Features Examination of dermoscopy images oftenrequires the interpretation of global pattern appearancewhich can be generally described with terms such as regular-ity homogeneity smoothness or grain These attributes are

related to the local intensity variations and can be quantifiedby using texture metrics For the analysis of texture parame-ters we use cooccurrence matrix measures that are computedfrom a 2D histogram which preserves spatial information

A cooccurrence matrix is a matrix that represents thedistribution of cooccurring values at a given offset A cooc-currence matrix 119862 is defined over an 119899 times 119898 image 119868 param-eterized by an offset (Δ119909 Δ119910) as [7]

119862Δ119909Δ119910

(119894 119895)

=

119899

sum

119901=1

119898

sum

119902=1

1 if 119868 (119901 119902) = 119894 119868 (119901 + Δ119909 119902 + Δ119910) = 119895

0 otherwise

(6)

The values 119894 and 119895 and size of 119862 are determined by the valuesin 119868

The texture analysis for dermoscopic images is performedfor 8-bit grayscale images quantized to 256 values and isreferred to as gray-level cooccurrence matrices (GLCM)

Textural features are frequently used in advanced melan-ocytic lesion analysis [7] Celebi et al suggested uniformlyquantizing the dermoscopy image to 64 gray levels and thenextracting 8 Haralick features for each one of the fourorientations using single pixel offsets [7 19ndash21]The followingparameters have been calculated

maximum probability = max119894119895

119888 (119894 119895)

energy = sum

119894

sum

119895

119888 (119894 119895)2

entropy = minussum

119894

sum

119895

119901 (119894 119895) log (119888 (119894 119895))

dissimilarity =1003816100381610038161003816119894 minus 119895

1003816100381610038161003816119888 (119894 119895)

contrast =119873minus1

sum

119899=0

sum

119894

sum

119895

119888 (119894 119895) |1003816100381610038161003816119894 minus 119895

1003816100381610038161003816= 119899

inverse difference = sum

119888 (119894 119895)

1 +1003816100381610038161003816119894 minus 119895

1003816100381610038161003816

inverse difference moment = sum

119888 (119894 119895)

1 + (119894 minus 119895)2

correlation =

(119894119895) 119901 (119894 119895) minus 120583119909120583119910

120590119909120590119910

(7)

where 119888(119894 119895) is the value of the normalized cooccurrencematrix at indexes (119894 119895)119873 is the number of gray levels 120583

119909and

120583119910are the mean values of the rows and columns of 119888 120590

119909and

120590119910are the respective standard deviations and the symbol |

indicates a condition that must be validSince the range of values of raw data varies widely the

feature scaling step has to be done before feature selection andclassification Feature scaling is a method used to standardizethe range of independent variables or features of data In this

6 BioMed Research International

work we rescale the range of features to scale the range in[minus1 1] The formula is given as

1199091015840

= 2 sdot

119909 minusmin (119909)

max (119909) minusmin (119909)

minus 1 (8)

232 Feature Selection In the next step we select the mostsuitable features from the set of all features described in theprevious sectionThere are many criteria to describe effectivefeatures but they should possess the following advantageslarge interclass mean distance small interclass varianceinsensitivity to extraneous variables and low correlationwithother features [15]

In this work we propose to use the correlation-basedfeature selection (CFS) algorithm CFS relies on a heuristicfor evaluating the worth or merit of a subset of featuresThis heuristic takes into account the usefulness of individualfeatures for predicting the class label along with the level ofintercorrelation among them

If the correlation between each of the components in atest and the outside variable is known and the intercorre-lation between each pair of components is given then thecorrelation between a composite consisting of the summedcomponents and the outside variable can be predicted from[22]

119903119911119888

=

119896119903119911119894

radic119896 + 119896 minus (119896 minus 1) 119903119894119894

(9)

where 119903119911119888

is correlation between the summed componentsand the outside variable 119896 is number of components (fea-tures) 119903

119911119894is average of the correlations between the com-

ponents and the outside variable and 119903119894119894is average intercor-

relation between componentsEquation (9) shows that the correlation between a com-

posite and an outside variable is a function of the numberof component variables in the composite and the magnitudeof the intercorrelations among them together with the mag-nitude of the correlations between the components and theoutside variable [22]

The following features have been chosen for the classifica-tion process compactness solidity symmetry distance colorvariegation features variance of entropy contrast dissimilar-ity and ellipse variance

24 Classification Using Support Vector Machine (SVM)After the set of discriminative features has been selected aclassifier can be designed for the evaluation of micro skinmoles The Support Vector Machine (SVM) is a supervisedlearning model used for data analysis pattern recognitionclassification and regression analysis An SVM trainingalgorithmbuilds amodel that assigns new samples to a relatedclass An SVMmodel is a representation of samples as pointsin space a linear function is used so that the examples of theseparate classes are divided by a clear gap

Given a training set of instance-label pairs (x119894 y119894) 119894 = 1

119897 where x119894isin 119877119899 and y isin 1 minus1

119897 the SVM requires thesolution of the following optimization problem [23]

minw119887120585

1

2

w119879w + 119862

119897

sum

119894=1

120585119894

subject to y119894(w119879120601 (x

119894) + 119887) ge 1 minus 120585

119894

120585119894ge 0

(10)

where training vectors x119894are mapped into a higher dimen-

sional space by the function 120601 The SVM finds a linear sep-arating hyperplane with the maximal margin in this higherdimensional space119862 gt 0 is the penalty parameter of the errorterm

Moreover 119870(x119894 x119895) equiv 120601(x

119894)119879

120601(x119895) is called the kernel

function [23] We use the commonly implemented RadialBasis Function (RBF) defined as

119870(x119894 x119895) = exp (minus120574

1003817100381710038171003817x119894minus x119895

10038171003817100381710038171003817

2

) 120574 gt 0 (11)

25 Model Evaluation Method For the evaluation of thedesigned algorithm we use cross-validation which is a tech-nique for assessing how the results of a statistical analysis willgeneralize to an independent dataset Cross-validation (CV)is primarily a way ofmeasuring the predictive performance ofa statisticalmodel In this approach the so called 119896-fold cross-validation method has been used In 119896-fold cross-validationthe original sample is randomly split into 119896 smaller setsThe following steps are undertaken a model is trained using119896 minus 1 of the folds as training data and the resulting model isvalidated on the remaining part of the data

The performance measure reported by 119896-fold cross-validation is then the average of the values computed in theloop

3 Results and Discussion

31 Database Specification The described algorithm for theautomatic classification of micro-melanocytic lesions hasbeen tested on dermoscopic images from a widely usedInteractive Atlas of Dermoscopy [1] and a private databaseImages for this atlas have been provided by two universityhospitals (University of Naples Italy and University of GrazAustria) and stored on aCD-ROMin a JPEG formatThedoc-umentation of each dermoscopic image was performed usinga Dermaphot apparatus (Heine Optotechnik HerrschingGermany) and a photo camera (Nikon F3) mounted on astereomicroscope (Wild M650 Heerbrugg AG Switzerland)in order to produce digitized ELM images of skin lesions Allof the images have been assessed manually by a dermoscopicexpert with extensive clinical experience

Furthermore all of the descriptions of skin cases werebased on histopathological examination of the biopsy mate-rial In order to develop and test the automatic procedurefor the classification of micro-melanocytic skin lesions 200images with different resolutions ranging from 0033 to

BioMed Research International 7

Table 1 List of parameters evaluated by the proposed method

Desired result Output result TotalBenign Malignant

Benign 65 5 70Malignant 7 123 130

Sensitivity 90 Specificity 96 200

05mmpixel were chosen The database included 70 malig-nant melanomas and 130 benign cases

The preprocessing step (black frame removal and hairremoval) as well as the segmentation step (border error lessthan 6) did not affect the further research [13 14]

32 Statistical Analysis Diagnostic exactness is often used tomean the fraction of cases on which the automated imageanalysis system gives a correct answer where correctness isdetermined by comparing the diagnostic decision to somedefinition of ldquotruthrdquo [15] There are different ways to definethe ldquotruthrdquo In our case we depend on the separate goldstandard which is determined by the results of biopsy for thedermoscopic images

The way to assess the accuracy of the designed classifica-tion system is to compare its results against the gold standardThe analyzed task is a binary problem where the answer canbe ldquo1rdquo when the lesion is malignant or ldquo0rdquo when it is benignTomeasure the diagnostic performanceweuse a pair of statis-tics parameters including sensitivity and specificity whichare not affected by the disease prevalence

The sensitivity or True Positive rate is the probability thatthe lesion is said to be malignant given that it is This can beestimated by relative frequency [15]

sensitivity =

TPTP + FN

(12)

where TP is the True Positive answer and FP is the FalsePositive answer

On the other hand we calculate specificity (False Positiverate) which is defined as the proportion of the True Negativesagainst all negative results Specificity is defined by thefollowing equation

specificity =

TNTN + FN

(13)

where TN is the True Negative answer and FN is the FalseNegative answer

Table 1 shows the classification results obtained with theimplemented algorithm in a confusion matrix

Another dominant technique for assessing computer clas-sification performance is the receiver operation characteristic(ROC) analysis A ROC curve is a plot of sensitivity versus 1 minusspecificity (Figure 6) Area under the ROC curve is the mostcommonly used accuracy index The area under a perfect-accuracy ROC curve when sensitivity and specificity reach10 is also 10 Another interesting fact is that for randomguessing is the positive diagonal line and the area under thecurve is 05 [15]

1009080706050403020100

10

09

08

07

06

05

04

03

02

01

00

Sens

itivi

tyRandom

ROC curve

1 minus specificity

Figure 6 ROC plot for the SVM classifier (AUC = 9324)

The conducted results (sensitivity = 90 specificity =96 and AUC = 9324) achieved by the implementedmicro-melanoma classification system are much better thanthe results described in [8 9] Ourmethod allowed classifica-tion of micro-melanomas very precisely and we can confirmthat the classification ofmicro skin lesions has to be done sep-arately from the classification of developed skin moles Over-all our experiments clearly show that the classification ofmicro skin moles which causes problems can be supportedby a powerful system based on proposed artificial neuralnetwork

4 Conclusions and Future Plans

Due to the difficulty and subjectivity of human interpretationthe computerized image analysis techniques have become animportant tool in the interpretation of dermoscopic imagesIn this research we propose a new approach for the classifica-tion of melanocytic skin moles The results obtained withinthis study indicate that the proposed algorithm can be usedfor the diagnosis of very small skin moles These conclusionswere reached after comparing computer classification withbiopsy outcome using statistical measurements and ROCcurve The automatic procedures have been tested and itseems to be a very promising software tool supporting thephysician The proposed diagnostic system can be used notonly by the young inexperienced dermatologist but also firstand foremost by family physiciansThis can be a great oppor-tunity for people that live in remote and rural areas outsidethe regional centre and frequently find it difficult to make anappointment with a dermatologist The importance of diag-nosing amelanoma in the early stage and the reduction of the

8 BioMed Research International

melanoma-related mortality rate can be achieved by precisecomputer-aided diagnostic systems

However computer classification of micro-melanomasmust also be further developed and tested on a larger datasetbefore it can be used clinically and before its potential inreducing the mortality rate of malignant melanoma can berealized in clinical medicine

Competing Interests

The author declares that there are no competing interestsregarding the publication of this paper

Acknowledgments

The scientific work was partly supported by the NationalScience Centre based on Decision no 201101NST706783and by AGH Research Grant 1511120408

References

[1] G Argenziano H P Soyer V De Giorgio et al Interactive Atlasof Dermoscopy (Book and CD-ROM) Edra Medical Publishingand New Media Milano Italy 2000

[2] Malignant melanoma 2015 httpwwwdermatologycaskin-hair-nailsskinskin-cancermalignant-melanoma

[3] S Seidenari C Ferrari S Borsari et al ldquoDermoscopy of smallmelanomas just miniaturized dermoscopyrdquo British Journal ofDermatology vol 171 no 5 pp 1006ndash1013 2014

[4] A Bono C Bartoli D Moglia et al ldquoSmall melanomas aclinical study on 270 consecutive cases of cutaneousmelanomardquoMelanoma Research vol 9 no 6 pp 583ndash586 1999

[5] C Schmoeckel ldquoSmall malignant melanomas clinicopatho-logic correlation and DNA ploidy analysisrdquo Journal of theAmerican Academy of Dermatology vol 24 no 6 part 1 pp1036ndash1037 1991

[6] S Seidenari S Bassoli S Borsari et al ldquoVariegated dermoscopyof in situ melanomardquoDermatology vol 224 no 3 pp 262ndash2702012

[7] J Scharcanski and M E Celebi Eds Computer Vision Tech-niques for the Diagnosis of Skin Cancer Springer New York NYUSA 2013

[8] A Bono C BartoliM Baldi et al ldquoMicro-melanomadetectionA clinical study on 22 cases of melanoma with a diameter equalto or less than 3mmrdquo Tumori vol 90 no 1 pp 128ndash131 2004

[9] A Bono E Tolomio S Trincone et al ldquoMicro-melanomadetection a clinical study on 206 consecutive cases of pig-mented skin lesions with a diameter le 3mmrdquo British Journalof Dermatology vol 155 no 3 pp 570ndash573 2006

[10] R Tadeusiewicz ldquoPlace and role of intelligent systems incomputer sciencerdquo Computer Methods in Materials Science vol10 no 4 pp 193ndash206 2010

[11] Disease overview 2015 httpsolaranrxcomunmet-medical-needdisease-overview

[12] J Jaworek-Korjakowska and R Tadeusiewicz ldquoHair removalfrom dermoscopic colour imagesrdquo Bio-Algorithms andMed Sys-tems vol 9 no 2 pp 53ndash58 2013

[13] J Jaworek-Korjakowska ldquoNovel method for border irregularityassessment in dermoscopic color imagesrdquo Computational and

MathematicalMethods inMedicine vol 2015 Article ID 49620211 pages 2015

[14] J Jaworek-Korjakowska and R Tadeusiewicz ldquoAssessment ofdots and globules in dermoscopic color images as one of the7-point check list criteriardquo in Proceedings of the 20th IEEE Inter-national Conference on Image Processing (ICIP rsquo13) pp 1456ndash1460 Melbourne Australia September 2013

[15] I N BankmanHandbook of Medical Imaging Academic PressNew York NY USA 2000

[16] V Ng and D Cheung ldquoMeasuring asymmetries of skin lesionsrdquoin Proceedings of the IEEE International Conference on SystemsMan and Cybernetics pp 4211ndash4216 Orlando Fla USA Octo-ber 1997

[17] V T YNg B YM Fung andT K Lee ldquoDetermining the asym-metry of skin lesion with fuzzy bordersrdquo Computers in Biologyand Medicine vol 35 no 2 pp 103ndash120 2005

[18] J Premaladha and K S Ravichandran ldquoAsymmetry analysis ofmalignant melanoma using image processing a surveyrdquo Journalof Artificial Intelligence vol 7 no 2 pp 45ndash53 2014

[19] M E Celebi H A Kingravi B Uddin et al ldquoA methodologicalapproach to the classification of dermoscopy imagesrdquo Comput-erized Medical Imaging and Graphics vol 31 no 6 pp 362ndash3732007

[20] R M Haralick K Shanmugam and I Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973

[21] J Smietanski R Tadeusiewicz and E Łuczynska ldquoTextureanalysis in perfusion images of prostate cancermdasha case studyrdquoInternational Journal of Applied Mathematics and ComputerScience vol 20 no 1 pp 149ndash156 2010

[22] M Nixon Feature Extraction amp Image Processing for ComputerVision Academic Press New York NY USA 3rd edition 2012

[23] httpwwwcsientuedutwsimcjlinpapersguideguidepdf

Submit your manuscripts athttpwwwhindawicom

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Anatomy Research International

PeptidesInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

International Journal of

Volume 2014

Zoology

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Molecular Biology International

GenomicsInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioinformaticsAdvances in

Marine BiologyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Signal TransductionJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioMed Research International

Evolutionary BiologyInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Biochemistry Research International

ArchaeaHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Genetics Research International

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Advances in

Virolog y

Hindawi Publishing Corporationhttpwwwhindawicom

Nucleic AcidsJournal of

Volume 2014

Stem CellsInternational

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Enzyme Research

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

Microbiology

2 BioMed Research International

Four stages of melanoma

Epidermis

Dermis

Subcutaneoustissue

Stage IDisease only inskin very thin

Stage IIISpread to lymph nodes

Stage 0Melanoma confined toepidermal region of skin

Stage IILocalized diseasethicker than stage I

Stage IVDisease spread toother organs

Figure 1 Melanoma is staged in four segments depending upon the spread of the disease Stage 0 is also called melanoma in situ themelanoma cells are only in the outer layer of the skin (the epidermis) Stages I and II mean there are melanoma cells in the layer directlyunder the epidermis or touching on the next layer down but there is no sign that it has spread to lymph nodes or other parts of the body [11]

(1) Clinical Importance and Motivation The most importantparameter which predicts the stage of melanoma is the thick-ness of the examined lesion (Figure 1) Skin moles with thethickness less than 1mm are near 100 curable [1] Seidenariand coworkers reported the direct correlation between diam-eter and thickness and as the diameter of in situ melanomasis smaller than that of invasive lesions it is reasonable tobelieve that small melanomas are usually in an initial growthphase [3] Therefore the aim of each clinician is to detectmalignant melanomas when they are still small and thin

Micro-melanomas represent a minority of diagnosedlesions their frequency ranging from 1 to 17 The meandiameter of in situ melanomas is around 1 cm and invasivemelanomas are usually gt6mm [3ndash6] In our research we con-centrate on micro-melanomas with the diameter lower than5mm

In the light of the above data prevention and earlydiagnosis of melanoma become extremely important issuesMicro-melanomas due to the small size and low advancementof development create enormous difficulties during the diag-nosis even for experts Such lesions are so little that they areclose to the limit of clinical relevance

Nowadays themedical checkup of smallmoles is possiblethrough a dermoscope (also called digital epiluminescencemicroscopy or ELM) which consists of a magnifier (typicallytimes10) a nonpolarised or polarised light source a transparentplate and a liquidmedium (for nonpolarised light) It enablesperforming a noninvasive in vivo examination making sub-surface structures of the skin visible (Figure 2)

The dermoscopic diagnosis of pigmented skin lesionis based on the assessment of the presence or absence ofdifferent global and local features Various analytic methodsand algorithms have been set forth in the last 25 years Themost important and widely used are pattern analysis ABCDrule of dermoscopy 7-point checklist and Menzies method[1 7]The classification of the obtained parameters is the finalstep of the diagnosis process where the extracted features are

Figure 2 Dermoscopic images of micro-melanoma lesions (basedon [1])

utilized to ascertain whether the lesion is malignant suspi-cious or benign

(2) Related Works Different groups of researchers havefocused on the extraction and classification of significantfeatures to provide distinction between benign andmalignantskin lesions however when dealing with micro-melanomasonly few works have been published but only in the scientificliterature ofmedicine In 2004Bono and coworkers describeda clinical study of 22 cases of melanoma with the diameterlower than 3mmThey extracted geometric and dermoscopicfeatures A lesion was considered suspicious for melanomawhen at least one parameter has been positive [8] In theirwork they notice that the ABCD diagnostic algorithmbecomes redundant while evaluating micro-melanomas andthat traditional dermoscopic diagnostic methods often failIn 2006 Bono et al published another work about micro-melanoma lesions based on more than 200 medical cases [9]The clinical evaluation gave a sensitivity of 43 and speci-ficity of 91 However using a dermoscope the sensitivityincreased to 83 while the specificity dropped to 69 Thedescribed above papers approach the subject only from theclinical point of view

During patient examinations researchers observe thatyoung inexperienced dermatologists and family physicianshave huge difficulties in the correct visual assessment of skin

BioMed Research International 3

Micro-melanomaimage

Preprocessing Segmentation Feature extractionFeature selection

Classification

Figure 3 Flowchart of the proposed automated image analysis system

lesions Due to this obstacle researchers try to implementand build computer-aided diagnosis (CAD) systems forautomated diagnosis of melanoma to increase the specificityand sensitivity and make the assessment of the skin molemore simple [10] Although such computer programs aredeveloped for different diagnostic algorithms to the best ofour knowledge a CAD system to classify micro-melanomalesions has not been proposed yet In our approach we pro-pose to divide skinmoles into twomain groups in terms of thediameter In this research we concentrate on the evaluationof significant and discriminant features as well as an accurateclassification process of only micro-melanoma skin lesions

2 Materials and Methods

After the introduction to the aim of this research we wouldlike to introduce the technical aswell as algorithmic solutionsThe first part presents the implemented diagnostic systemwith its particular steps In the second part we concentrateon the feature extraction and selection as well as the SVMclassification method

The implemented application is divided into four stagespreprocessing (image enhancement) skin lesion segmen-tation feature extraction and selection and classification(Figure 3) Since themain aimof this research is to present thefeature extraction and classification step the preprocessingand segmentation stages will be described shortly A detaileddescription of these steps can be found in our works [12ndash14]

21 Image Enhancement Dermoscopy images are often dete-riorated by noise due to various sources of interference andother phenomena that affect the image acquisition processImage enhancement techniques are used to refine the ana-lyzed image so that desired image features become easier toperceive and to be detected by the automated image analysissystem [15] Dermoscopic images are inhomogeneous andcomplex and furthermore they contain extraneous artifactssuch as skin lines air bubbles and hairs which appear inalmost every image Therefore the preprocessing step isessential for dermoscopic images to improve the quality andit is responsible for reducing the amount of artifacts andnoise The preprocessing step includes black frame removalsmoothing of the air bubbles and black hair inpainting

The black frame is detected on the basis of the light-ness component value in the HSL color space In order to

determine the darkness of a pixel with (119877 119866 119861) coordinatesthe lightness component of the HSL color space is calculated[12]

119871 =

max (119877 119866 119861) +min (119877 119866 119861)

2

(1)

A pixel is considered to be black when the lightness value isless than 15 (range of the lightness value [0 255]) For thesmoothing of air bubbles and light hairs we use the Gaussianfilter

The removing of thick black hair is the most importantstep in dermoscopy image preprocessing since it influencesthe segmentation and feature extraction step Firstly the der-moscopic RGB image is being converted into grayscale withthe NTSC 1953 standard Secondly the black top-hat trans-form which is a morphological image processing techniqueis used to detect thick dark hairsThe result of this step is thedifference between the closing operation and the input image

119879119908(119868) = 119868 ∘ 119887 minus 119868 (2)

where ∘ denotes the closing operation 119868 is the grayscale inputimage and 119887 is a grayscale structuring element [15] The laststep is the hair inpainting Hair line pixels are replaced withvalues calculated on the basis of the neighborhood pixels(Figure 4)

22 Segmentation Algorithm The principal goal of the seg-mentation process is to partition an image into regions thatare homogeneous in terms of pixel intensity or specific fea-tures [15] Inmedical imaging segmentation is important andcrucial for feature extraction and image measurements andclassification Precision of classification of a tumor dependsgreatly on the accuracy of the extracted tumor area as mostof the features are calculated on this basis During our workwe have compared different segmentation methods [13 14]Based on the previous research the applied segmentationalgorithm for the skin lesion extraction is based on seededregion-growing algorithm The region-growing algorithmalso called region merging looks for groups of pixels thathave similar intensities whereas thresholding methods focuson the difference of pixel intensities [15]Themain advantageof the region-growing algorithm is that it is capable ofcorrectly segmenting regions that have the same propertiesand are spatially separated Because the healthy skin in der-moscopy images is mostly homogenous the region-growing

4 BioMed Research International

(a) (b)

Figure 4 Outcome of the preprocessing step (a) dermoscopic input image and (b) image after black frame removal and hair inpainting

(a) (b) (c)

Figure 5 Dermoscopic image segmentation using region-growing method (a) dermoscopic input image (b) image after enhancement and(c) segmented skin mole

algorithm can be applied to detect it For the skin lesion seg-mentation we select one seed which is located in the left cor-ner on the healthy skin It gives us the certainty that we willseparate the homogeneous background from the skin moleThe region is iteratively grown by comparing all unallocatedneighboring pixels to the regionThe region-growing processconsists of picking a seed from the set investigating all 4-connectedneighbors of this seed andmerging suitable neigh-bors to the seed The seed is then removed and all mergedneighbors are added to the seed set The region-growingprocess continues until the seed set is empty Figure 5presents the results of preprocessing and segmentation stepsfor micro-melanoma skin lesion

23 Dermoscopic ImageAnalysis After the preprocessing andsegmentation steps the dermoscopy image is prepared for adetailed analysisNow it is necessary to analyze the image andrecognize critical parameters that can indicate melanomaThese properties are basically the same ones that a physician

looks for In this research we concentrate on two of the mostinformative visual cues in dermoscopy image interpretationincluding shape and texture Shape parameters correspondwith the diagnostic algorithm called the ABCD rule of der-moscopy while the texture parameters correspond with thediagnostic algorithm 7-point checklist

231 Feature Extraction Many different image features havebeen proposed for the analysis of medical images includ-ing intensity geometrical morphological fractal dimensionbased and texture featuresThe feature extraction step can beconsidered as data compression that removes additional andirrelevant information while preserving relevant informationin a data vector Features selected for the calculation ofdetected skin moles are based on the knowledge and expe-rience about the properties and differences in appearancebetween the benign and malignant skin lesions However tofind out the most redundant and effective features we extracta large number of initial features The proposed features are

BioMed Research International 5

identifiable translatable rotation and scale invariant noiseresistant and reliable

Shape Features (Geometrical Features) Shape-based imageretrieval consists of measuring of similarity between shapesrepresented by their features Shapes that have large differ-ences can be identifiedwith the geometrical features howevereven small changes can affect the final value of the featureWe calculated parameters including major and minor axeslength eccentricity (measure of aspect ratio it is the ratioof the major axis to the minor axis) elongation (based onthe bounding box) compactness (normalized version byHaralick) ellipse variance rectangularity (ratio area of shapeand area of the minimum bounding box) convexity (ratio ofperimeter of convex hull and contour) and solidity (ratio ofarea of convex hull and shape area)

Asymmetry Asymmetry is the major parameter in the diag-nostic algorithm ABCD rule of dermoscopy We quantifyasymmetry calculating the symmetry distance (SD) proposedby NG in 1997 [16 17] It is the calculated result of typical dis-placement among the number of verticeswhen the skin lesionshape is transformed into its symmetric shape [18] The SDvalue is measured as

SD =

1

119899

119899minus1

sum

119894=0

10038171003817100381710038171003817119875119894 minus 119894

10038171003817100381710038171003817 (3)

Color Variegation A skin lesion may have a variety of colorsincluding tan dark brown black blue red and occasionallylight gray Melanomas are usually characterized by three ormore colors and in about 40 of melanomas even five or sixcolors are present We analyze the color variation throughfour different channels including red green blue and inten-sity The intensity channel for each pixel 119868(119909 119910) is defined

119868 (119909 119910) = radic1198772(119909 119910) + 119866

2(119909 119910) + 119861

2(119909 119910) (4)

The following parameters are calculated for assessing thecolor variegation [7]

cv1= log 120590 (119877)

120583 (119877)

cv2= log 120590 (119866)

120583 (119866)

cv3= log 120590 (119861)

120583 (119861)

cv4= log 120590 (119868)

120583 (119868)

(5)

where 120583 is the mean value and 120590 denotes standard deviation

Texture Features Examination of dermoscopy images oftenrequires the interpretation of global pattern appearancewhich can be generally described with terms such as regular-ity homogeneity smoothness or grain These attributes are

related to the local intensity variations and can be quantifiedby using texture metrics For the analysis of texture parame-ters we use cooccurrence matrix measures that are computedfrom a 2D histogram which preserves spatial information

A cooccurrence matrix is a matrix that represents thedistribution of cooccurring values at a given offset A cooc-currence matrix 119862 is defined over an 119899 times 119898 image 119868 param-eterized by an offset (Δ119909 Δ119910) as [7]

119862Δ119909Δ119910

(119894 119895)

=

119899

sum

119901=1

119898

sum

119902=1

1 if 119868 (119901 119902) = 119894 119868 (119901 + Δ119909 119902 + Δ119910) = 119895

0 otherwise

(6)

The values 119894 and 119895 and size of 119862 are determined by the valuesin 119868

The texture analysis for dermoscopic images is performedfor 8-bit grayscale images quantized to 256 values and isreferred to as gray-level cooccurrence matrices (GLCM)

Textural features are frequently used in advanced melan-ocytic lesion analysis [7] Celebi et al suggested uniformlyquantizing the dermoscopy image to 64 gray levels and thenextracting 8 Haralick features for each one of the fourorientations using single pixel offsets [7 19ndash21]The followingparameters have been calculated

maximum probability = max119894119895

119888 (119894 119895)

energy = sum

119894

sum

119895

119888 (119894 119895)2

entropy = minussum

119894

sum

119895

119901 (119894 119895) log (119888 (119894 119895))

dissimilarity =1003816100381610038161003816119894 minus 119895

1003816100381610038161003816119888 (119894 119895)

contrast =119873minus1

sum

119899=0

sum

119894

sum

119895

119888 (119894 119895) |1003816100381610038161003816119894 minus 119895

1003816100381610038161003816= 119899

inverse difference = sum

119888 (119894 119895)

1 +1003816100381610038161003816119894 minus 119895

1003816100381610038161003816

inverse difference moment = sum

119888 (119894 119895)

1 + (119894 minus 119895)2

correlation =

(119894119895) 119901 (119894 119895) minus 120583119909120583119910

120590119909120590119910

(7)

where 119888(119894 119895) is the value of the normalized cooccurrencematrix at indexes (119894 119895)119873 is the number of gray levels 120583

119909and

120583119910are the mean values of the rows and columns of 119888 120590

119909and

120590119910are the respective standard deviations and the symbol |

indicates a condition that must be validSince the range of values of raw data varies widely the

feature scaling step has to be done before feature selection andclassification Feature scaling is a method used to standardizethe range of independent variables or features of data In this

6 BioMed Research International

work we rescale the range of features to scale the range in[minus1 1] The formula is given as

1199091015840

= 2 sdot

119909 minusmin (119909)

max (119909) minusmin (119909)

minus 1 (8)

232 Feature Selection In the next step we select the mostsuitable features from the set of all features described in theprevious sectionThere are many criteria to describe effectivefeatures but they should possess the following advantageslarge interclass mean distance small interclass varianceinsensitivity to extraneous variables and low correlationwithother features [15]

In this work we propose to use the correlation-basedfeature selection (CFS) algorithm CFS relies on a heuristicfor evaluating the worth or merit of a subset of featuresThis heuristic takes into account the usefulness of individualfeatures for predicting the class label along with the level ofintercorrelation among them

If the correlation between each of the components in atest and the outside variable is known and the intercorre-lation between each pair of components is given then thecorrelation between a composite consisting of the summedcomponents and the outside variable can be predicted from[22]

119903119911119888

=

119896119903119911119894

radic119896 + 119896 minus (119896 minus 1) 119903119894119894

(9)

where 119903119911119888

is correlation between the summed componentsand the outside variable 119896 is number of components (fea-tures) 119903

119911119894is average of the correlations between the com-

ponents and the outside variable and 119903119894119894is average intercor-

relation between componentsEquation (9) shows that the correlation between a com-

posite and an outside variable is a function of the numberof component variables in the composite and the magnitudeof the intercorrelations among them together with the mag-nitude of the correlations between the components and theoutside variable [22]

The following features have been chosen for the classifica-tion process compactness solidity symmetry distance colorvariegation features variance of entropy contrast dissimilar-ity and ellipse variance

24 Classification Using Support Vector Machine (SVM)After the set of discriminative features has been selected aclassifier can be designed for the evaluation of micro skinmoles The Support Vector Machine (SVM) is a supervisedlearning model used for data analysis pattern recognitionclassification and regression analysis An SVM trainingalgorithmbuilds amodel that assigns new samples to a relatedclass An SVMmodel is a representation of samples as pointsin space a linear function is used so that the examples of theseparate classes are divided by a clear gap

Given a training set of instance-label pairs (x119894 y119894) 119894 = 1

119897 where x119894isin 119877119899 and y isin 1 minus1

119897 the SVM requires thesolution of the following optimization problem [23]

minw119887120585

1

2

w119879w + 119862

119897

sum

119894=1

120585119894

subject to y119894(w119879120601 (x

119894) + 119887) ge 1 minus 120585

119894

120585119894ge 0

(10)

where training vectors x119894are mapped into a higher dimen-

sional space by the function 120601 The SVM finds a linear sep-arating hyperplane with the maximal margin in this higherdimensional space119862 gt 0 is the penalty parameter of the errorterm

Moreover 119870(x119894 x119895) equiv 120601(x

119894)119879

120601(x119895) is called the kernel

function [23] We use the commonly implemented RadialBasis Function (RBF) defined as

119870(x119894 x119895) = exp (minus120574

1003817100381710038171003817x119894minus x119895

10038171003817100381710038171003817

2

) 120574 gt 0 (11)

25 Model Evaluation Method For the evaluation of thedesigned algorithm we use cross-validation which is a tech-nique for assessing how the results of a statistical analysis willgeneralize to an independent dataset Cross-validation (CV)is primarily a way ofmeasuring the predictive performance ofa statisticalmodel In this approach the so called 119896-fold cross-validation method has been used In 119896-fold cross-validationthe original sample is randomly split into 119896 smaller setsThe following steps are undertaken a model is trained using119896 minus 1 of the folds as training data and the resulting model isvalidated on the remaining part of the data

The performance measure reported by 119896-fold cross-validation is then the average of the values computed in theloop

3 Results and Discussion

31 Database Specification The described algorithm for theautomatic classification of micro-melanocytic lesions hasbeen tested on dermoscopic images from a widely usedInteractive Atlas of Dermoscopy [1] and a private databaseImages for this atlas have been provided by two universityhospitals (University of Naples Italy and University of GrazAustria) and stored on aCD-ROMin a JPEG formatThedoc-umentation of each dermoscopic image was performed usinga Dermaphot apparatus (Heine Optotechnik HerrschingGermany) and a photo camera (Nikon F3) mounted on astereomicroscope (Wild M650 Heerbrugg AG Switzerland)in order to produce digitized ELM images of skin lesions Allof the images have been assessed manually by a dermoscopicexpert with extensive clinical experience

Furthermore all of the descriptions of skin cases werebased on histopathological examination of the biopsy mate-rial In order to develop and test the automatic procedurefor the classification of micro-melanocytic skin lesions 200images with different resolutions ranging from 0033 to

BioMed Research International 7

Table 1 List of parameters evaluated by the proposed method

Desired result Output result TotalBenign Malignant

Benign 65 5 70Malignant 7 123 130

Sensitivity 90 Specificity 96 200

05mmpixel were chosen The database included 70 malig-nant melanomas and 130 benign cases

The preprocessing step (black frame removal and hairremoval) as well as the segmentation step (border error lessthan 6) did not affect the further research [13 14]

32 Statistical Analysis Diagnostic exactness is often used tomean the fraction of cases on which the automated imageanalysis system gives a correct answer where correctness isdetermined by comparing the diagnostic decision to somedefinition of ldquotruthrdquo [15] There are different ways to definethe ldquotruthrdquo In our case we depend on the separate goldstandard which is determined by the results of biopsy for thedermoscopic images

The way to assess the accuracy of the designed classifica-tion system is to compare its results against the gold standardThe analyzed task is a binary problem where the answer canbe ldquo1rdquo when the lesion is malignant or ldquo0rdquo when it is benignTomeasure the diagnostic performanceweuse a pair of statis-tics parameters including sensitivity and specificity whichare not affected by the disease prevalence

The sensitivity or True Positive rate is the probability thatthe lesion is said to be malignant given that it is This can beestimated by relative frequency [15]

sensitivity =

TPTP + FN

(12)

where TP is the True Positive answer and FP is the FalsePositive answer

On the other hand we calculate specificity (False Positiverate) which is defined as the proportion of the True Negativesagainst all negative results Specificity is defined by thefollowing equation

specificity =

TNTN + FN

(13)

where TN is the True Negative answer and FN is the FalseNegative answer

Table 1 shows the classification results obtained with theimplemented algorithm in a confusion matrix

Another dominant technique for assessing computer clas-sification performance is the receiver operation characteristic(ROC) analysis A ROC curve is a plot of sensitivity versus 1 minusspecificity (Figure 6) Area under the ROC curve is the mostcommonly used accuracy index The area under a perfect-accuracy ROC curve when sensitivity and specificity reach10 is also 10 Another interesting fact is that for randomguessing is the positive diagonal line and the area under thecurve is 05 [15]

1009080706050403020100

10

09

08

07

06

05

04

03

02

01

00

Sens

itivi

tyRandom

ROC curve

1 minus specificity

Figure 6 ROC plot for the SVM classifier (AUC = 9324)

The conducted results (sensitivity = 90 specificity =96 and AUC = 9324) achieved by the implementedmicro-melanoma classification system are much better thanthe results described in [8 9] Ourmethod allowed classifica-tion of micro-melanomas very precisely and we can confirmthat the classification ofmicro skin lesions has to be done sep-arately from the classification of developed skin moles Over-all our experiments clearly show that the classification ofmicro skin moles which causes problems can be supportedby a powerful system based on proposed artificial neuralnetwork

4 Conclusions and Future Plans

Due to the difficulty and subjectivity of human interpretationthe computerized image analysis techniques have become animportant tool in the interpretation of dermoscopic imagesIn this research we propose a new approach for the classifica-tion of melanocytic skin moles The results obtained withinthis study indicate that the proposed algorithm can be usedfor the diagnosis of very small skin moles These conclusionswere reached after comparing computer classification withbiopsy outcome using statistical measurements and ROCcurve The automatic procedures have been tested and itseems to be a very promising software tool supporting thephysician The proposed diagnostic system can be used notonly by the young inexperienced dermatologist but also firstand foremost by family physiciansThis can be a great oppor-tunity for people that live in remote and rural areas outsidethe regional centre and frequently find it difficult to make anappointment with a dermatologist The importance of diag-nosing amelanoma in the early stage and the reduction of the

8 BioMed Research International

melanoma-related mortality rate can be achieved by precisecomputer-aided diagnostic systems

However computer classification of micro-melanomasmust also be further developed and tested on a larger datasetbefore it can be used clinically and before its potential inreducing the mortality rate of malignant melanoma can berealized in clinical medicine

Competing Interests

The author declares that there are no competing interestsregarding the publication of this paper

Acknowledgments

The scientific work was partly supported by the NationalScience Centre based on Decision no 201101NST706783and by AGH Research Grant 1511120408

References

[1] G Argenziano H P Soyer V De Giorgio et al Interactive Atlasof Dermoscopy (Book and CD-ROM) Edra Medical Publishingand New Media Milano Italy 2000

[2] Malignant melanoma 2015 httpwwwdermatologycaskin-hair-nailsskinskin-cancermalignant-melanoma

[3] S Seidenari C Ferrari S Borsari et al ldquoDermoscopy of smallmelanomas just miniaturized dermoscopyrdquo British Journal ofDermatology vol 171 no 5 pp 1006ndash1013 2014

[4] A Bono C Bartoli D Moglia et al ldquoSmall melanomas aclinical study on 270 consecutive cases of cutaneousmelanomardquoMelanoma Research vol 9 no 6 pp 583ndash586 1999

[5] C Schmoeckel ldquoSmall malignant melanomas clinicopatho-logic correlation and DNA ploidy analysisrdquo Journal of theAmerican Academy of Dermatology vol 24 no 6 part 1 pp1036ndash1037 1991

[6] S Seidenari S Bassoli S Borsari et al ldquoVariegated dermoscopyof in situ melanomardquoDermatology vol 224 no 3 pp 262ndash2702012

[7] J Scharcanski and M E Celebi Eds Computer Vision Tech-niques for the Diagnosis of Skin Cancer Springer New York NYUSA 2013

[8] A Bono C BartoliM Baldi et al ldquoMicro-melanomadetectionA clinical study on 22 cases of melanoma with a diameter equalto or less than 3mmrdquo Tumori vol 90 no 1 pp 128ndash131 2004

[9] A Bono E Tolomio S Trincone et al ldquoMicro-melanomadetection a clinical study on 206 consecutive cases of pig-mented skin lesions with a diameter le 3mmrdquo British Journalof Dermatology vol 155 no 3 pp 570ndash573 2006

[10] R Tadeusiewicz ldquoPlace and role of intelligent systems incomputer sciencerdquo Computer Methods in Materials Science vol10 no 4 pp 193ndash206 2010

[11] Disease overview 2015 httpsolaranrxcomunmet-medical-needdisease-overview

[12] J Jaworek-Korjakowska and R Tadeusiewicz ldquoHair removalfrom dermoscopic colour imagesrdquo Bio-Algorithms andMed Sys-tems vol 9 no 2 pp 53ndash58 2013

[13] J Jaworek-Korjakowska ldquoNovel method for border irregularityassessment in dermoscopic color imagesrdquo Computational and

MathematicalMethods inMedicine vol 2015 Article ID 49620211 pages 2015

[14] J Jaworek-Korjakowska and R Tadeusiewicz ldquoAssessment ofdots and globules in dermoscopic color images as one of the7-point check list criteriardquo in Proceedings of the 20th IEEE Inter-national Conference on Image Processing (ICIP rsquo13) pp 1456ndash1460 Melbourne Australia September 2013

[15] I N BankmanHandbook of Medical Imaging Academic PressNew York NY USA 2000

[16] V Ng and D Cheung ldquoMeasuring asymmetries of skin lesionsrdquoin Proceedings of the IEEE International Conference on SystemsMan and Cybernetics pp 4211ndash4216 Orlando Fla USA Octo-ber 1997

[17] V T YNg B YM Fung andT K Lee ldquoDetermining the asym-metry of skin lesion with fuzzy bordersrdquo Computers in Biologyand Medicine vol 35 no 2 pp 103ndash120 2005

[18] J Premaladha and K S Ravichandran ldquoAsymmetry analysis ofmalignant melanoma using image processing a surveyrdquo Journalof Artificial Intelligence vol 7 no 2 pp 45ndash53 2014

[19] M E Celebi H A Kingravi B Uddin et al ldquoA methodologicalapproach to the classification of dermoscopy imagesrdquo Comput-erized Medical Imaging and Graphics vol 31 no 6 pp 362ndash3732007

[20] R M Haralick K Shanmugam and I Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973

[21] J Smietanski R Tadeusiewicz and E Łuczynska ldquoTextureanalysis in perfusion images of prostate cancermdasha case studyrdquoInternational Journal of Applied Mathematics and ComputerScience vol 20 no 1 pp 149ndash156 2010

[22] M Nixon Feature Extraction amp Image Processing for ComputerVision Academic Press New York NY USA 3rd edition 2012

[23] httpwwwcsientuedutwsimcjlinpapersguideguidepdf

Submit your manuscripts athttpwwwhindawicom

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Anatomy Research International

PeptidesInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

International Journal of

Volume 2014

Zoology

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Molecular Biology International

GenomicsInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioinformaticsAdvances in

Marine BiologyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Signal TransductionJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioMed Research International

Evolutionary BiologyInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Biochemistry Research International

ArchaeaHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Genetics Research International

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Advances in

Virolog y

Hindawi Publishing Corporationhttpwwwhindawicom

Nucleic AcidsJournal of

Volume 2014

Stem CellsInternational

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Enzyme Research

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

Microbiology

BioMed Research International 3

Micro-melanomaimage

Preprocessing Segmentation Feature extractionFeature selection

Classification

Figure 3 Flowchart of the proposed automated image analysis system

lesions Due to this obstacle researchers try to implementand build computer-aided diagnosis (CAD) systems forautomated diagnosis of melanoma to increase the specificityand sensitivity and make the assessment of the skin molemore simple [10] Although such computer programs aredeveloped for different diagnostic algorithms to the best ofour knowledge a CAD system to classify micro-melanomalesions has not been proposed yet In our approach we pro-pose to divide skinmoles into twomain groups in terms of thediameter In this research we concentrate on the evaluationof significant and discriminant features as well as an accurateclassification process of only micro-melanoma skin lesions

2 Materials and Methods

After the introduction to the aim of this research we wouldlike to introduce the technical aswell as algorithmic solutionsThe first part presents the implemented diagnostic systemwith its particular steps In the second part we concentrateon the feature extraction and selection as well as the SVMclassification method

The implemented application is divided into four stagespreprocessing (image enhancement) skin lesion segmen-tation feature extraction and selection and classification(Figure 3) Since themain aimof this research is to present thefeature extraction and classification step the preprocessingand segmentation stages will be described shortly A detaileddescription of these steps can be found in our works [12ndash14]

21 Image Enhancement Dermoscopy images are often dete-riorated by noise due to various sources of interference andother phenomena that affect the image acquisition processImage enhancement techniques are used to refine the ana-lyzed image so that desired image features become easier toperceive and to be detected by the automated image analysissystem [15] Dermoscopic images are inhomogeneous andcomplex and furthermore they contain extraneous artifactssuch as skin lines air bubbles and hairs which appear inalmost every image Therefore the preprocessing step isessential for dermoscopic images to improve the quality andit is responsible for reducing the amount of artifacts andnoise The preprocessing step includes black frame removalsmoothing of the air bubbles and black hair inpainting

The black frame is detected on the basis of the light-ness component value in the HSL color space In order to

determine the darkness of a pixel with (119877 119866 119861) coordinatesthe lightness component of the HSL color space is calculated[12]

119871 =

max (119877 119866 119861) +min (119877 119866 119861)

2

(1)

A pixel is considered to be black when the lightness value isless than 15 (range of the lightness value [0 255]) For thesmoothing of air bubbles and light hairs we use the Gaussianfilter

The removing of thick black hair is the most importantstep in dermoscopy image preprocessing since it influencesthe segmentation and feature extraction step Firstly the der-moscopic RGB image is being converted into grayscale withthe NTSC 1953 standard Secondly the black top-hat trans-form which is a morphological image processing techniqueis used to detect thick dark hairsThe result of this step is thedifference between the closing operation and the input image

119879119908(119868) = 119868 ∘ 119887 minus 119868 (2)

where ∘ denotes the closing operation 119868 is the grayscale inputimage and 119887 is a grayscale structuring element [15] The laststep is the hair inpainting Hair line pixels are replaced withvalues calculated on the basis of the neighborhood pixels(Figure 4)

22 Segmentation Algorithm The principal goal of the seg-mentation process is to partition an image into regions thatare homogeneous in terms of pixel intensity or specific fea-tures [15] Inmedical imaging segmentation is important andcrucial for feature extraction and image measurements andclassification Precision of classification of a tumor dependsgreatly on the accuracy of the extracted tumor area as mostof the features are calculated on this basis During our workwe have compared different segmentation methods [13 14]Based on the previous research the applied segmentationalgorithm for the skin lesion extraction is based on seededregion-growing algorithm The region-growing algorithmalso called region merging looks for groups of pixels thathave similar intensities whereas thresholding methods focuson the difference of pixel intensities [15]Themain advantageof the region-growing algorithm is that it is capable ofcorrectly segmenting regions that have the same propertiesand are spatially separated Because the healthy skin in der-moscopy images is mostly homogenous the region-growing

4 BioMed Research International

(a) (b)

Figure 4 Outcome of the preprocessing step (a) dermoscopic input image and (b) image after black frame removal and hair inpainting

(a) (b) (c)

Figure 5 Dermoscopic image segmentation using region-growing method (a) dermoscopic input image (b) image after enhancement and(c) segmented skin mole

algorithm can be applied to detect it For the skin lesion seg-mentation we select one seed which is located in the left cor-ner on the healthy skin It gives us the certainty that we willseparate the homogeneous background from the skin moleThe region is iteratively grown by comparing all unallocatedneighboring pixels to the regionThe region-growing processconsists of picking a seed from the set investigating all 4-connectedneighbors of this seed andmerging suitable neigh-bors to the seed The seed is then removed and all mergedneighbors are added to the seed set The region-growingprocess continues until the seed set is empty Figure 5presents the results of preprocessing and segmentation stepsfor micro-melanoma skin lesion

23 Dermoscopic ImageAnalysis After the preprocessing andsegmentation steps the dermoscopy image is prepared for adetailed analysisNow it is necessary to analyze the image andrecognize critical parameters that can indicate melanomaThese properties are basically the same ones that a physician

looks for In this research we concentrate on two of the mostinformative visual cues in dermoscopy image interpretationincluding shape and texture Shape parameters correspondwith the diagnostic algorithm called the ABCD rule of der-moscopy while the texture parameters correspond with thediagnostic algorithm 7-point checklist

231 Feature Extraction Many different image features havebeen proposed for the analysis of medical images includ-ing intensity geometrical morphological fractal dimensionbased and texture featuresThe feature extraction step can beconsidered as data compression that removes additional andirrelevant information while preserving relevant informationin a data vector Features selected for the calculation ofdetected skin moles are based on the knowledge and expe-rience about the properties and differences in appearancebetween the benign and malignant skin lesions However tofind out the most redundant and effective features we extracta large number of initial features The proposed features are

BioMed Research International 5

identifiable translatable rotation and scale invariant noiseresistant and reliable

Shape Features (Geometrical Features) Shape-based imageretrieval consists of measuring of similarity between shapesrepresented by their features Shapes that have large differ-ences can be identifiedwith the geometrical features howevereven small changes can affect the final value of the featureWe calculated parameters including major and minor axeslength eccentricity (measure of aspect ratio it is the ratioof the major axis to the minor axis) elongation (based onthe bounding box) compactness (normalized version byHaralick) ellipse variance rectangularity (ratio area of shapeand area of the minimum bounding box) convexity (ratio ofperimeter of convex hull and contour) and solidity (ratio ofarea of convex hull and shape area)

Asymmetry Asymmetry is the major parameter in the diag-nostic algorithm ABCD rule of dermoscopy We quantifyasymmetry calculating the symmetry distance (SD) proposedby NG in 1997 [16 17] It is the calculated result of typical dis-placement among the number of verticeswhen the skin lesionshape is transformed into its symmetric shape [18] The SDvalue is measured as

SD =

1

119899

119899minus1

sum

119894=0

10038171003817100381710038171003817119875119894 minus 119894

10038171003817100381710038171003817 (3)

Color Variegation A skin lesion may have a variety of colorsincluding tan dark brown black blue red and occasionallylight gray Melanomas are usually characterized by three ormore colors and in about 40 of melanomas even five or sixcolors are present We analyze the color variation throughfour different channels including red green blue and inten-sity The intensity channel for each pixel 119868(119909 119910) is defined

119868 (119909 119910) = radic1198772(119909 119910) + 119866

2(119909 119910) + 119861

2(119909 119910) (4)

The following parameters are calculated for assessing thecolor variegation [7]

cv1= log 120590 (119877)

120583 (119877)

cv2= log 120590 (119866)

120583 (119866)

cv3= log 120590 (119861)

120583 (119861)

cv4= log 120590 (119868)

120583 (119868)

(5)

where 120583 is the mean value and 120590 denotes standard deviation

Texture Features Examination of dermoscopy images oftenrequires the interpretation of global pattern appearancewhich can be generally described with terms such as regular-ity homogeneity smoothness or grain These attributes are

related to the local intensity variations and can be quantifiedby using texture metrics For the analysis of texture parame-ters we use cooccurrence matrix measures that are computedfrom a 2D histogram which preserves spatial information

A cooccurrence matrix is a matrix that represents thedistribution of cooccurring values at a given offset A cooc-currence matrix 119862 is defined over an 119899 times 119898 image 119868 param-eterized by an offset (Δ119909 Δ119910) as [7]

119862Δ119909Δ119910

(119894 119895)

=

119899

sum

119901=1

119898

sum

119902=1

1 if 119868 (119901 119902) = 119894 119868 (119901 + Δ119909 119902 + Δ119910) = 119895

0 otherwise

(6)

The values 119894 and 119895 and size of 119862 are determined by the valuesin 119868

The texture analysis for dermoscopic images is performedfor 8-bit grayscale images quantized to 256 values and isreferred to as gray-level cooccurrence matrices (GLCM)

Textural features are frequently used in advanced melan-ocytic lesion analysis [7] Celebi et al suggested uniformlyquantizing the dermoscopy image to 64 gray levels and thenextracting 8 Haralick features for each one of the fourorientations using single pixel offsets [7 19ndash21]The followingparameters have been calculated

maximum probability = max119894119895

119888 (119894 119895)

energy = sum

119894

sum

119895

119888 (119894 119895)2

entropy = minussum

119894

sum

119895

119901 (119894 119895) log (119888 (119894 119895))

dissimilarity =1003816100381610038161003816119894 minus 119895

1003816100381610038161003816119888 (119894 119895)

contrast =119873minus1

sum

119899=0

sum

119894

sum

119895

119888 (119894 119895) |1003816100381610038161003816119894 minus 119895

1003816100381610038161003816= 119899

inverse difference = sum

119888 (119894 119895)

1 +1003816100381610038161003816119894 minus 119895

1003816100381610038161003816

inverse difference moment = sum

119888 (119894 119895)

1 + (119894 minus 119895)2

correlation =

(119894119895) 119901 (119894 119895) minus 120583119909120583119910

120590119909120590119910

(7)

where 119888(119894 119895) is the value of the normalized cooccurrencematrix at indexes (119894 119895)119873 is the number of gray levels 120583

119909and

120583119910are the mean values of the rows and columns of 119888 120590

119909and

120590119910are the respective standard deviations and the symbol |

indicates a condition that must be validSince the range of values of raw data varies widely the

feature scaling step has to be done before feature selection andclassification Feature scaling is a method used to standardizethe range of independent variables or features of data In this

6 BioMed Research International

work we rescale the range of features to scale the range in[minus1 1] The formula is given as

1199091015840

= 2 sdot

119909 minusmin (119909)

max (119909) minusmin (119909)

minus 1 (8)

232 Feature Selection In the next step we select the mostsuitable features from the set of all features described in theprevious sectionThere are many criteria to describe effectivefeatures but they should possess the following advantageslarge interclass mean distance small interclass varianceinsensitivity to extraneous variables and low correlationwithother features [15]

In this work we propose to use the correlation-basedfeature selection (CFS) algorithm CFS relies on a heuristicfor evaluating the worth or merit of a subset of featuresThis heuristic takes into account the usefulness of individualfeatures for predicting the class label along with the level ofintercorrelation among them

If the correlation between each of the components in atest and the outside variable is known and the intercorre-lation between each pair of components is given then thecorrelation between a composite consisting of the summedcomponents and the outside variable can be predicted from[22]

119903119911119888

=

119896119903119911119894

radic119896 + 119896 minus (119896 minus 1) 119903119894119894

(9)

where 119903119911119888

is correlation between the summed componentsand the outside variable 119896 is number of components (fea-tures) 119903

119911119894is average of the correlations between the com-

ponents and the outside variable and 119903119894119894is average intercor-

relation between componentsEquation (9) shows that the correlation between a com-

posite and an outside variable is a function of the numberof component variables in the composite and the magnitudeof the intercorrelations among them together with the mag-nitude of the correlations between the components and theoutside variable [22]

The following features have been chosen for the classifica-tion process compactness solidity symmetry distance colorvariegation features variance of entropy contrast dissimilar-ity and ellipse variance

24 Classification Using Support Vector Machine (SVM)After the set of discriminative features has been selected aclassifier can be designed for the evaluation of micro skinmoles The Support Vector Machine (SVM) is a supervisedlearning model used for data analysis pattern recognitionclassification and regression analysis An SVM trainingalgorithmbuilds amodel that assigns new samples to a relatedclass An SVMmodel is a representation of samples as pointsin space a linear function is used so that the examples of theseparate classes are divided by a clear gap

Given a training set of instance-label pairs (x119894 y119894) 119894 = 1

119897 where x119894isin 119877119899 and y isin 1 minus1

119897 the SVM requires thesolution of the following optimization problem [23]

minw119887120585

1

2

w119879w + 119862

119897

sum

119894=1

120585119894

subject to y119894(w119879120601 (x

119894) + 119887) ge 1 minus 120585

119894

120585119894ge 0

(10)

where training vectors x119894are mapped into a higher dimen-

sional space by the function 120601 The SVM finds a linear sep-arating hyperplane with the maximal margin in this higherdimensional space119862 gt 0 is the penalty parameter of the errorterm

Moreover 119870(x119894 x119895) equiv 120601(x

119894)119879

120601(x119895) is called the kernel

function [23] We use the commonly implemented RadialBasis Function (RBF) defined as

119870(x119894 x119895) = exp (minus120574

1003817100381710038171003817x119894minus x119895

10038171003817100381710038171003817

2

) 120574 gt 0 (11)

25 Model Evaluation Method For the evaluation of thedesigned algorithm we use cross-validation which is a tech-nique for assessing how the results of a statistical analysis willgeneralize to an independent dataset Cross-validation (CV)is primarily a way ofmeasuring the predictive performance ofa statisticalmodel In this approach the so called 119896-fold cross-validation method has been used In 119896-fold cross-validationthe original sample is randomly split into 119896 smaller setsThe following steps are undertaken a model is trained using119896 minus 1 of the folds as training data and the resulting model isvalidated on the remaining part of the data

The performance measure reported by 119896-fold cross-validation is then the average of the values computed in theloop

3 Results and Discussion

31 Database Specification The described algorithm for theautomatic classification of micro-melanocytic lesions hasbeen tested on dermoscopic images from a widely usedInteractive Atlas of Dermoscopy [1] and a private databaseImages for this atlas have been provided by two universityhospitals (University of Naples Italy and University of GrazAustria) and stored on aCD-ROMin a JPEG formatThedoc-umentation of each dermoscopic image was performed usinga Dermaphot apparatus (Heine Optotechnik HerrschingGermany) and a photo camera (Nikon F3) mounted on astereomicroscope (Wild M650 Heerbrugg AG Switzerland)in order to produce digitized ELM images of skin lesions Allof the images have been assessed manually by a dermoscopicexpert with extensive clinical experience

Furthermore all of the descriptions of skin cases werebased on histopathological examination of the biopsy mate-rial In order to develop and test the automatic procedurefor the classification of micro-melanocytic skin lesions 200images with different resolutions ranging from 0033 to

BioMed Research International 7

Table 1 List of parameters evaluated by the proposed method

Desired result Output result TotalBenign Malignant

Benign 65 5 70Malignant 7 123 130

Sensitivity 90 Specificity 96 200

05mmpixel were chosen The database included 70 malig-nant melanomas and 130 benign cases

The preprocessing step (black frame removal and hairremoval) as well as the segmentation step (border error lessthan 6) did not affect the further research [13 14]

32 Statistical Analysis Diagnostic exactness is often used tomean the fraction of cases on which the automated imageanalysis system gives a correct answer where correctness isdetermined by comparing the diagnostic decision to somedefinition of ldquotruthrdquo [15] There are different ways to definethe ldquotruthrdquo In our case we depend on the separate goldstandard which is determined by the results of biopsy for thedermoscopic images

The way to assess the accuracy of the designed classifica-tion system is to compare its results against the gold standardThe analyzed task is a binary problem where the answer canbe ldquo1rdquo when the lesion is malignant or ldquo0rdquo when it is benignTomeasure the diagnostic performanceweuse a pair of statis-tics parameters including sensitivity and specificity whichare not affected by the disease prevalence

The sensitivity or True Positive rate is the probability thatthe lesion is said to be malignant given that it is This can beestimated by relative frequency [15]

sensitivity =

TPTP + FN

(12)

where TP is the True Positive answer and FP is the FalsePositive answer

On the other hand we calculate specificity (False Positiverate) which is defined as the proportion of the True Negativesagainst all negative results Specificity is defined by thefollowing equation

specificity =

TNTN + FN

(13)

where TN is the True Negative answer and FN is the FalseNegative answer

Table 1 shows the classification results obtained with theimplemented algorithm in a confusion matrix

Another dominant technique for assessing computer clas-sification performance is the receiver operation characteristic(ROC) analysis A ROC curve is a plot of sensitivity versus 1 minusspecificity (Figure 6) Area under the ROC curve is the mostcommonly used accuracy index The area under a perfect-accuracy ROC curve when sensitivity and specificity reach10 is also 10 Another interesting fact is that for randomguessing is the positive diagonal line and the area under thecurve is 05 [15]

1009080706050403020100

10

09

08

07

06

05

04

03

02

01

00

Sens

itivi

tyRandom

ROC curve

1 minus specificity

Figure 6 ROC plot for the SVM classifier (AUC = 9324)

The conducted results (sensitivity = 90 specificity =96 and AUC = 9324) achieved by the implementedmicro-melanoma classification system are much better thanthe results described in [8 9] Ourmethod allowed classifica-tion of micro-melanomas very precisely and we can confirmthat the classification ofmicro skin lesions has to be done sep-arately from the classification of developed skin moles Over-all our experiments clearly show that the classification ofmicro skin moles which causes problems can be supportedby a powerful system based on proposed artificial neuralnetwork

4 Conclusions and Future Plans

Due to the difficulty and subjectivity of human interpretationthe computerized image analysis techniques have become animportant tool in the interpretation of dermoscopic imagesIn this research we propose a new approach for the classifica-tion of melanocytic skin moles The results obtained withinthis study indicate that the proposed algorithm can be usedfor the diagnosis of very small skin moles These conclusionswere reached after comparing computer classification withbiopsy outcome using statistical measurements and ROCcurve The automatic procedures have been tested and itseems to be a very promising software tool supporting thephysician The proposed diagnostic system can be used notonly by the young inexperienced dermatologist but also firstand foremost by family physiciansThis can be a great oppor-tunity for people that live in remote and rural areas outsidethe regional centre and frequently find it difficult to make anappointment with a dermatologist The importance of diag-nosing amelanoma in the early stage and the reduction of the

8 BioMed Research International

melanoma-related mortality rate can be achieved by precisecomputer-aided diagnostic systems

However computer classification of micro-melanomasmust also be further developed and tested on a larger datasetbefore it can be used clinically and before its potential inreducing the mortality rate of malignant melanoma can berealized in clinical medicine

Competing Interests

The author declares that there are no competing interestsregarding the publication of this paper

Acknowledgments

The scientific work was partly supported by the NationalScience Centre based on Decision no 201101NST706783and by AGH Research Grant 1511120408

References

[1] G Argenziano H P Soyer V De Giorgio et al Interactive Atlasof Dermoscopy (Book and CD-ROM) Edra Medical Publishingand New Media Milano Italy 2000

[2] Malignant melanoma 2015 httpwwwdermatologycaskin-hair-nailsskinskin-cancermalignant-melanoma

[3] S Seidenari C Ferrari S Borsari et al ldquoDermoscopy of smallmelanomas just miniaturized dermoscopyrdquo British Journal ofDermatology vol 171 no 5 pp 1006ndash1013 2014

[4] A Bono C Bartoli D Moglia et al ldquoSmall melanomas aclinical study on 270 consecutive cases of cutaneousmelanomardquoMelanoma Research vol 9 no 6 pp 583ndash586 1999

[5] C Schmoeckel ldquoSmall malignant melanomas clinicopatho-logic correlation and DNA ploidy analysisrdquo Journal of theAmerican Academy of Dermatology vol 24 no 6 part 1 pp1036ndash1037 1991

[6] S Seidenari S Bassoli S Borsari et al ldquoVariegated dermoscopyof in situ melanomardquoDermatology vol 224 no 3 pp 262ndash2702012

[7] J Scharcanski and M E Celebi Eds Computer Vision Tech-niques for the Diagnosis of Skin Cancer Springer New York NYUSA 2013

[8] A Bono C BartoliM Baldi et al ldquoMicro-melanomadetectionA clinical study on 22 cases of melanoma with a diameter equalto or less than 3mmrdquo Tumori vol 90 no 1 pp 128ndash131 2004

[9] A Bono E Tolomio S Trincone et al ldquoMicro-melanomadetection a clinical study on 206 consecutive cases of pig-mented skin lesions with a diameter le 3mmrdquo British Journalof Dermatology vol 155 no 3 pp 570ndash573 2006

[10] R Tadeusiewicz ldquoPlace and role of intelligent systems incomputer sciencerdquo Computer Methods in Materials Science vol10 no 4 pp 193ndash206 2010

[11] Disease overview 2015 httpsolaranrxcomunmet-medical-needdisease-overview

[12] J Jaworek-Korjakowska and R Tadeusiewicz ldquoHair removalfrom dermoscopic colour imagesrdquo Bio-Algorithms andMed Sys-tems vol 9 no 2 pp 53ndash58 2013

[13] J Jaworek-Korjakowska ldquoNovel method for border irregularityassessment in dermoscopic color imagesrdquo Computational and

MathematicalMethods inMedicine vol 2015 Article ID 49620211 pages 2015

[14] J Jaworek-Korjakowska and R Tadeusiewicz ldquoAssessment ofdots and globules in dermoscopic color images as one of the7-point check list criteriardquo in Proceedings of the 20th IEEE Inter-national Conference on Image Processing (ICIP rsquo13) pp 1456ndash1460 Melbourne Australia September 2013

[15] I N BankmanHandbook of Medical Imaging Academic PressNew York NY USA 2000

[16] V Ng and D Cheung ldquoMeasuring asymmetries of skin lesionsrdquoin Proceedings of the IEEE International Conference on SystemsMan and Cybernetics pp 4211ndash4216 Orlando Fla USA Octo-ber 1997

[17] V T YNg B YM Fung andT K Lee ldquoDetermining the asym-metry of skin lesion with fuzzy bordersrdquo Computers in Biologyand Medicine vol 35 no 2 pp 103ndash120 2005

[18] J Premaladha and K S Ravichandran ldquoAsymmetry analysis ofmalignant melanoma using image processing a surveyrdquo Journalof Artificial Intelligence vol 7 no 2 pp 45ndash53 2014

[19] M E Celebi H A Kingravi B Uddin et al ldquoA methodologicalapproach to the classification of dermoscopy imagesrdquo Comput-erized Medical Imaging and Graphics vol 31 no 6 pp 362ndash3732007

[20] R M Haralick K Shanmugam and I Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973

[21] J Smietanski R Tadeusiewicz and E Łuczynska ldquoTextureanalysis in perfusion images of prostate cancermdasha case studyrdquoInternational Journal of Applied Mathematics and ComputerScience vol 20 no 1 pp 149ndash156 2010

[22] M Nixon Feature Extraction amp Image Processing for ComputerVision Academic Press New York NY USA 3rd edition 2012

[23] httpwwwcsientuedutwsimcjlinpapersguideguidepdf

Submit your manuscripts athttpwwwhindawicom

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Anatomy Research International

PeptidesInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

International Journal of

Volume 2014

Zoology

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Molecular Biology International

GenomicsInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioinformaticsAdvances in

Marine BiologyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Signal TransductionJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioMed Research International

Evolutionary BiologyInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Biochemistry Research International

ArchaeaHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Genetics Research International

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Advances in

Virolog y

Hindawi Publishing Corporationhttpwwwhindawicom

Nucleic AcidsJournal of

Volume 2014

Stem CellsInternational

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Enzyme Research

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

Microbiology

4 BioMed Research International

(a) (b)

Figure 4 Outcome of the preprocessing step (a) dermoscopic input image and (b) image after black frame removal and hair inpainting

(a) (b) (c)

Figure 5 Dermoscopic image segmentation using region-growing method (a) dermoscopic input image (b) image after enhancement and(c) segmented skin mole

algorithm can be applied to detect it For the skin lesion seg-mentation we select one seed which is located in the left cor-ner on the healthy skin It gives us the certainty that we willseparate the homogeneous background from the skin moleThe region is iteratively grown by comparing all unallocatedneighboring pixels to the regionThe region-growing processconsists of picking a seed from the set investigating all 4-connectedneighbors of this seed andmerging suitable neigh-bors to the seed The seed is then removed and all mergedneighbors are added to the seed set The region-growingprocess continues until the seed set is empty Figure 5presents the results of preprocessing and segmentation stepsfor micro-melanoma skin lesion

23 Dermoscopic ImageAnalysis After the preprocessing andsegmentation steps the dermoscopy image is prepared for adetailed analysisNow it is necessary to analyze the image andrecognize critical parameters that can indicate melanomaThese properties are basically the same ones that a physician

looks for In this research we concentrate on two of the mostinformative visual cues in dermoscopy image interpretationincluding shape and texture Shape parameters correspondwith the diagnostic algorithm called the ABCD rule of der-moscopy while the texture parameters correspond with thediagnostic algorithm 7-point checklist

231 Feature Extraction Many different image features havebeen proposed for the analysis of medical images includ-ing intensity geometrical morphological fractal dimensionbased and texture featuresThe feature extraction step can beconsidered as data compression that removes additional andirrelevant information while preserving relevant informationin a data vector Features selected for the calculation ofdetected skin moles are based on the knowledge and expe-rience about the properties and differences in appearancebetween the benign and malignant skin lesions However tofind out the most redundant and effective features we extracta large number of initial features The proposed features are

BioMed Research International 5

identifiable translatable rotation and scale invariant noiseresistant and reliable

Shape Features (Geometrical Features) Shape-based imageretrieval consists of measuring of similarity between shapesrepresented by their features Shapes that have large differ-ences can be identifiedwith the geometrical features howevereven small changes can affect the final value of the featureWe calculated parameters including major and minor axeslength eccentricity (measure of aspect ratio it is the ratioof the major axis to the minor axis) elongation (based onthe bounding box) compactness (normalized version byHaralick) ellipse variance rectangularity (ratio area of shapeand area of the minimum bounding box) convexity (ratio ofperimeter of convex hull and contour) and solidity (ratio ofarea of convex hull and shape area)

Asymmetry Asymmetry is the major parameter in the diag-nostic algorithm ABCD rule of dermoscopy We quantifyasymmetry calculating the symmetry distance (SD) proposedby NG in 1997 [16 17] It is the calculated result of typical dis-placement among the number of verticeswhen the skin lesionshape is transformed into its symmetric shape [18] The SDvalue is measured as

SD =

1

119899

119899minus1

sum

119894=0

10038171003817100381710038171003817119875119894 minus 119894

10038171003817100381710038171003817 (3)

Color Variegation A skin lesion may have a variety of colorsincluding tan dark brown black blue red and occasionallylight gray Melanomas are usually characterized by three ormore colors and in about 40 of melanomas even five or sixcolors are present We analyze the color variation throughfour different channels including red green blue and inten-sity The intensity channel for each pixel 119868(119909 119910) is defined

119868 (119909 119910) = radic1198772(119909 119910) + 119866

2(119909 119910) + 119861

2(119909 119910) (4)

The following parameters are calculated for assessing thecolor variegation [7]

cv1= log 120590 (119877)

120583 (119877)

cv2= log 120590 (119866)

120583 (119866)

cv3= log 120590 (119861)

120583 (119861)

cv4= log 120590 (119868)

120583 (119868)

(5)

where 120583 is the mean value and 120590 denotes standard deviation

Texture Features Examination of dermoscopy images oftenrequires the interpretation of global pattern appearancewhich can be generally described with terms such as regular-ity homogeneity smoothness or grain These attributes are

related to the local intensity variations and can be quantifiedby using texture metrics For the analysis of texture parame-ters we use cooccurrence matrix measures that are computedfrom a 2D histogram which preserves spatial information

A cooccurrence matrix is a matrix that represents thedistribution of cooccurring values at a given offset A cooc-currence matrix 119862 is defined over an 119899 times 119898 image 119868 param-eterized by an offset (Δ119909 Δ119910) as [7]

119862Δ119909Δ119910

(119894 119895)

=

119899

sum

119901=1

119898

sum

119902=1

1 if 119868 (119901 119902) = 119894 119868 (119901 + Δ119909 119902 + Δ119910) = 119895

0 otherwise

(6)

The values 119894 and 119895 and size of 119862 are determined by the valuesin 119868

The texture analysis for dermoscopic images is performedfor 8-bit grayscale images quantized to 256 values and isreferred to as gray-level cooccurrence matrices (GLCM)

Textural features are frequently used in advanced melan-ocytic lesion analysis [7] Celebi et al suggested uniformlyquantizing the dermoscopy image to 64 gray levels and thenextracting 8 Haralick features for each one of the fourorientations using single pixel offsets [7 19ndash21]The followingparameters have been calculated

maximum probability = max119894119895

119888 (119894 119895)

energy = sum

119894

sum

119895

119888 (119894 119895)2

entropy = minussum

119894

sum

119895

119901 (119894 119895) log (119888 (119894 119895))

dissimilarity =1003816100381610038161003816119894 minus 119895

1003816100381610038161003816119888 (119894 119895)

contrast =119873minus1

sum

119899=0

sum

119894

sum

119895

119888 (119894 119895) |1003816100381610038161003816119894 minus 119895

1003816100381610038161003816= 119899

inverse difference = sum

119888 (119894 119895)

1 +1003816100381610038161003816119894 minus 119895

1003816100381610038161003816

inverse difference moment = sum

119888 (119894 119895)

1 + (119894 minus 119895)2

correlation =

(119894119895) 119901 (119894 119895) minus 120583119909120583119910

120590119909120590119910

(7)

where 119888(119894 119895) is the value of the normalized cooccurrencematrix at indexes (119894 119895)119873 is the number of gray levels 120583

119909and

120583119910are the mean values of the rows and columns of 119888 120590

119909and

120590119910are the respective standard deviations and the symbol |

indicates a condition that must be validSince the range of values of raw data varies widely the

feature scaling step has to be done before feature selection andclassification Feature scaling is a method used to standardizethe range of independent variables or features of data In this

6 BioMed Research International

work we rescale the range of features to scale the range in[minus1 1] The formula is given as

1199091015840

= 2 sdot

119909 minusmin (119909)

max (119909) minusmin (119909)

minus 1 (8)

232 Feature Selection In the next step we select the mostsuitable features from the set of all features described in theprevious sectionThere are many criteria to describe effectivefeatures but they should possess the following advantageslarge interclass mean distance small interclass varianceinsensitivity to extraneous variables and low correlationwithother features [15]

In this work we propose to use the correlation-basedfeature selection (CFS) algorithm CFS relies on a heuristicfor evaluating the worth or merit of a subset of featuresThis heuristic takes into account the usefulness of individualfeatures for predicting the class label along with the level ofintercorrelation among them

If the correlation between each of the components in atest and the outside variable is known and the intercorre-lation between each pair of components is given then thecorrelation between a composite consisting of the summedcomponents and the outside variable can be predicted from[22]

119903119911119888

=

119896119903119911119894

radic119896 + 119896 minus (119896 minus 1) 119903119894119894

(9)

where 119903119911119888

is correlation between the summed componentsand the outside variable 119896 is number of components (fea-tures) 119903

119911119894is average of the correlations between the com-

ponents and the outside variable and 119903119894119894is average intercor-

relation between componentsEquation (9) shows that the correlation between a com-

posite and an outside variable is a function of the numberof component variables in the composite and the magnitudeof the intercorrelations among them together with the mag-nitude of the correlations between the components and theoutside variable [22]

The following features have been chosen for the classifica-tion process compactness solidity symmetry distance colorvariegation features variance of entropy contrast dissimilar-ity and ellipse variance

24 Classification Using Support Vector Machine (SVM)After the set of discriminative features has been selected aclassifier can be designed for the evaluation of micro skinmoles The Support Vector Machine (SVM) is a supervisedlearning model used for data analysis pattern recognitionclassification and regression analysis An SVM trainingalgorithmbuilds amodel that assigns new samples to a relatedclass An SVMmodel is a representation of samples as pointsin space a linear function is used so that the examples of theseparate classes are divided by a clear gap

Given a training set of instance-label pairs (x119894 y119894) 119894 = 1

119897 where x119894isin 119877119899 and y isin 1 minus1

119897 the SVM requires thesolution of the following optimization problem [23]

minw119887120585

1

2

w119879w + 119862

119897

sum

119894=1

120585119894

subject to y119894(w119879120601 (x

119894) + 119887) ge 1 minus 120585

119894

120585119894ge 0

(10)

where training vectors x119894are mapped into a higher dimen-

sional space by the function 120601 The SVM finds a linear sep-arating hyperplane with the maximal margin in this higherdimensional space119862 gt 0 is the penalty parameter of the errorterm

Moreover 119870(x119894 x119895) equiv 120601(x

119894)119879

120601(x119895) is called the kernel

function [23] We use the commonly implemented RadialBasis Function (RBF) defined as

119870(x119894 x119895) = exp (minus120574

1003817100381710038171003817x119894minus x119895

10038171003817100381710038171003817

2

) 120574 gt 0 (11)

25 Model Evaluation Method For the evaluation of thedesigned algorithm we use cross-validation which is a tech-nique for assessing how the results of a statistical analysis willgeneralize to an independent dataset Cross-validation (CV)is primarily a way ofmeasuring the predictive performance ofa statisticalmodel In this approach the so called 119896-fold cross-validation method has been used In 119896-fold cross-validationthe original sample is randomly split into 119896 smaller setsThe following steps are undertaken a model is trained using119896 minus 1 of the folds as training data and the resulting model isvalidated on the remaining part of the data

The performance measure reported by 119896-fold cross-validation is then the average of the values computed in theloop

3 Results and Discussion

31 Database Specification The described algorithm for theautomatic classification of micro-melanocytic lesions hasbeen tested on dermoscopic images from a widely usedInteractive Atlas of Dermoscopy [1] and a private databaseImages for this atlas have been provided by two universityhospitals (University of Naples Italy and University of GrazAustria) and stored on aCD-ROMin a JPEG formatThedoc-umentation of each dermoscopic image was performed usinga Dermaphot apparatus (Heine Optotechnik HerrschingGermany) and a photo camera (Nikon F3) mounted on astereomicroscope (Wild M650 Heerbrugg AG Switzerland)in order to produce digitized ELM images of skin lesions Allof the images have been assessed manually by a dermoscopicexpert with extensive clinical experience

Furthermore all of the descriptions of skin cases werebased on histopathological examination of the biopsy mate-rial In order to develop and test the automatic procedurefor the classification of micro-melanocytic skin lesions 200images with different resolutions ranging from 0033 to

BioMed Research International 7

Table 1 List of parameters evaluated by the proposed method

Desired result Output result TotalBenign Malignant

Benign 65 5 70Malignant 7 123 130

Sensitivity 90 Specificity 96 200

05mmpixel were chosen The database included 70 malig-nant melanomas and 130 benign cases

The preprocessing step (black frame removal and hairremoval) as well as the segmentation step (border error lessthan 6) did not affect the further research [13 14]

32 Statistical Analysis Diagnostic exactness is often used tomean the fraction of cases on which the automated imageanalysis system gives a correct answer where correctness isdetermined by comparing the diagnostic decision to somedefinition of ldquotruthrdquo [15] There are different ways to definethe ldquotruthrdquo In our case we depend on the separate goldstandard which is determined by the results of biopsy for thedermoscopic images

The way to assess the accuracy of the designed classifica-tion system is to compare its results against the gold standardThe analyzed task is a binary problem where the answer canbe ldquo1rdquo when the lesion is malignant or ldquo0rdquo when it is benignTomeasure the diagnostic performanceweuse a pair of statis-tics parameters including sensitivity and specificity whichare not affected by the disease prevalence

The sensitivity or True Positive rate is the probability thatthe lesion is said to be malignant given that it is This can beestimated by relative frequency [15]

sensitivity =

TPTP + FN

(12)

where TP is the True Positive answer and FP is the FalsePositive answer

On the other hand we calculate specificity (False Positiverate) which is defined as the proportion of the True Negativesagainst all negative results Specificity is defined by thefollowing equation

specificity =

TNTN + FN

(13)

where TN is the True Negative answer and FN is the FalseNegative answer

Table 1 shows the classification results obtained with theimplemented algorithm in a confusion matrix

Another dominant technique for assessing computer clas-sification performance is the receiver operation characteristic(ROC) analysis A ROC curve is a plot of sensitivity versus 1 minusspecificity (Figure 6) Area under the ROC curve is the mostcommonly used accuracy index The area under a perfect-accuracy ROC curve when sensitivity and specificity reach10 is also 10 Another interesting fact is that for randomguessing is the positive diagonal line and the area under thecurve is 05 [15]

1009080706050403020100

10

09

08

07

06

05

04

03

02

01

00

Sens

itivi

tyRandom

ROC curve

1 minus specificity

Figure 6 ROC plot for the SVM classifier (AUC = 9324)

The conducted results (sensitivity = 90 specificity =96 and AUC = 9324) achieved by the implementedmicro-melanoma classification system are much better thanthe results described in [8 9] Ourmethod allowed classifica-tion of micro-melanomas very precisely and we can confirmthat the classification ofmicro skin lesions has to be done sep-arately from the classification of developed skin moles Over-all our experiments clearly show that the classification ofmicro skin moles which causes problems can be supportedby a powerful system based on proposed artificial neuralnetwork

4 Conclusions and Future Plans

Due to the difficulty and subjectivity of human interpretationthe computerized image analysis techniques have become animportant tool in the interpretation of dermoscopic imagesIn this research we propose a new approach for the classifica-tion of melanocytic skin moles The results obtained withinthis study indicate that the proposed algorithm can be usedfor the diagnosis of very small skin moles These conclusionswere reached after comparing computer classification withbiopsy outcome using statistical measurements and ROCcurve The automatic procedures have been tested and itseems to be a very promising software tool supporting thephysician The proposed diagnostic system can be used notonly by the young inexperienced dermatologist but also firstand foremost by family physiciansThis can be a great oppor-tunity for people that live in remote and rural areas outsidethe regional centre and frequently find it difficult to make anappointment with a dermatologist The importance of diag-nosing amelanoma in the early stage and the reduction of the

8 BioMed Research International

melanoma-related mortality rate can be achieved by precisecomputer-aided diagnostic systems

However computer classification of micro-melanomasmust also be further developed and tested on a larger datasetbefore it can be used clinically and before its potential inreducing the mortality rate of malignant melanoma can berealized in clinical medicine

Competing Interests

The author declares that there are no competing interestsregarding the publication of this paper

Acknowledgments

The scientific work was partly supported by the NationalScience Centre based on Decision no 201101NST706783and by AGH Research Grant 1511120408

References

[1] G Argenziano H P Soyer V De Giorgio et al Interactive Atlasof Dermoscopy (Book and CD-ROM) Edra Medical Publishingand New Media Milano Italy 2000

[2] Malignant melanoma 2015 httpwwwdermatologycaskin-hair-nailsskinskin-cancermalignant-melanoma

[3] S Seidenari C Ferrari S Borsari et al ldquoDermoscopy of smallmelanomas just miniaturized dermoscopyrdquo British Journal ofDermatology vol 171 no 5 pp 1006ndash1013 2014

[4] A Bono C Bartoli D Moglia et al ldquoSmall melanomas aclinical study on 270 consecutive cases of cutaneousmelanomardquoMelanoma Research vol 9 no 6 pp 583ndash586 1999

[5] C Schmoeckel ldquoSmall malignant melanomas clinicopatho-logic correlation and DNA ploidy analysisrdquo Journal of theAmerican Academy of Dermatology vol 24 no 6 part 1 pp1036ndash1037 1991

[6] S Seidenari S Bassoli S Borsari et al ldquoVariegated dermoscopyof in situ melanomardquoDermatology vol 224 no 3 pp 262ndash2702012

[7] J Scharcanski and M E Celebi Eds Computer Vision Tech-niques for the Diagnosis of Skin Cancer Springer New York NYUSA 2013

[8] A Bono C BartoliM Baldi et al ldquoMicro-melanomadetectionA clinical study on 22 cases of melanoma with a diameter equalto or less than 3mmrdquo Tumori vol 90 no 1 pp 128ndash131 2004

[9] A Bono E Tolomio S Trincone et al ldquoMicro-melanomadetection a clinical study on 206 consecutive cases of pig-mented skin lesions with a diameter le 3mmrdquo British Journalof Dermatology vol 155 no 3 pp 570ndash573 2006

[10] R Tadeusiewicz ldquoPlace and role of intelligent systems incomputer sciencerdquo Computer Methods in Materials Science vol10 no 4 pp 193ndash206 2010

[11] Disease overview 2015 httpsolaranrxcomunmet-medical-needdisease-overview

[12] J Jaworek-Korjakowska and R Tadeusiewicz ldquoHair removalfrom dermoscopic colour imagesrdquo Bio-Algorithms andMed Sys-tems vol 9 no 2 pp 53ndash58 2013

[13] J Jaworek-Korjakowska ldquoNovel method for border irregularityassessment in dermoscopic color imagesrdquo Computational and

MathematicalMethods inMedicine vol 2015 Article ID 49620211 pages 2015

[14] J Jaworek-Korjakowska and R Tadeusiewicz ldquoAssessment ofdots and globules in dermoscopic color images as one of the7-point check list criteriardquo in Proceedings of the 20th IEEE Inter-national Conference on Image Processing (ICIP rsquo13) pp 1456ndash1460 Melbourne Australia September 2013

[15] I N BankmanHandbook of Medical Imaging Academic PressNew York NY USA 2000

[16] V Ng and D Cheung ldquoMeasuring asymmetries of skin lesionsrdquoin Proceedings of the IEEE International Conference on SystemsMan and Cybernetics pp 4211ndash4216 Orlando Fla USA Octo-ber 1997

[17] V T YNg B YM Fung andT K Lee ldquoDetermining the asym-metry of skin lesion with fuzzy bordersrdquo Computers in Biologyand Medicine vol 35 no 2 pp 103ndash120 2005

[18] J Premaladha and K S Ravichandran ldquoAsymmetry analysis ofmalignant melanoma using image processing a surveyrdquo Journalof Artificial Intelligence vol 7 no 2 pp 45ndash53 2014

[19] M E Celebi H A Kingravi B Uddin et al ldquoA methodologicalapproach to the classification of dermoscopy imagesrdquo Comput-erized Medical Imaging and Graphics vol 31 no 6 pp 362ndash3732007

[20] R M Haralick K Shanmugam and I Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973

[21] J Smietanski R Tadeusiewicz and E Łuczynska ldquoTextureanalysis in perfusion images of prostate cancermdasha case studyrdquoInternational Journal of Applied Mathematics and ComputerScience vol 20 no 1 pp 149ndash156 2010

[22] M Nixon Feature Extraction amp Image Processing for ComputerVision Academic Press New York NY USA 3rd edition 2012

[23] httpwwwcsientuedutwsimcjlinpapersguideguidepdf

Submit your manuscripts athttpwwwhindawicom

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Anatomy Research International

PeptidesInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

International Journal of

Volume 2014

Zoology

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Molecular Biology International

GenomicsInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioinformaticsAdvances in

Marine BiologyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Signal TransductionJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioMed Research International

Evolutionary BiologyInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Biochemistry Research International

ArchaeaHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Genetics Research International

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Advances in

Virolog y

Hindawi Publishing Corporationhttpwwwhindawicom

Nucleic AcidsJournal of

Volume 2014

Stem CellsInternational

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Enzyme Research

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

Microbiology

BioMed Research International 5

identifiable translatable rotation and scale invariant noiseresistant and reliable

Shape Features (Geometrical Features) Shape-based imageretrieval consists of measuring of similarity between shapesrepresented by their features Shapes that have large differ-ences can be identifiedwith the geometrical features howevereven small changes can affect the final value of the featureWe calculated parameters including major and minor axeslength eccentricity (measure of aspect ratio it is the ratioof the major axis to the minor axis) elongation (based onthe bounding box) compactness (normalized version byHaralick) ellipse variance rectangularity (ratio area of shapeand area of the minimum bounding box) convexity (ratio ofperimeter of convex hull and contour) and solidity (ratio ofarea of convex hull and shape area)

Asymmetry Asymmetry is the major parameter in the diag-nostic algorithm ABCD rule of dermoscopy We quantifyasymmetry calculating the symmetry distance (SD) proposedby NG in 1997 [16 17] It is the calculated result of typical dis-placement among the number of verticeswhen the skin lesionshape is transformed into its symmetric shape [18] The SDvalue is measured as

SD =

1

119899

119899minus1

sum

119894=0

10038171003817100381710038171003817119875119894 minus 119894

10038171003817100381710038171003817 (3)

Color Variegation A skin lesion may have a variety of colorsincluding tan dark brown black blue red and occasionallylight gray Melanomas are usually characterized by three ormore colors and in about 40 of melanomas even five or sixcolors are present We analyze the color variation throughfour different channels including red green blue and inten-sity The intensity channel for each pixel 119868(119909 119910) is defined

119868 (119909 119910) = radic1198772(119909 119910) + 119866

2(119909 119910) + 119861

2(119909 119910) (4)

The following parameters are calculated for assessing thecolor variegation [7]

cv1= log 120590 (119877)

120583 (119877)

cv2= log 120590 (119866)

120583 (119866)

cv3= log 120590 (119861)

120583 (119861)

cv4= log 120590 (119868)

120583 (119868)

(5)

where 120583 is the mean value and 120590 denotes standard deviation

Texture Features Examination of dermoscopy images oftenrequires the interpretation of global pattern appearancewhich can be generally described with terms such as regular-ity homogeneity smoothness or grain These attributes are

related to the local intensity variations and can be quantifiedby using texture metrics For the analysis of texture parame-ters we use cooccurrence matrix measures that are computedfrom a 2D histogram which preserves spatial information

A cooccurrence matrix is a matrix that represents thedistribution of cooccurring values at a given offset A cooc-currence matrix 119862 is defined over an 119899 times 119898 image 119868 param-eterized by an offset (Δ119909 Δ119910) as [7]

119862Δ119909Δ119910

(119894 119895)

=

119899

sum

119901=1

119898

sum

119902=1

1 if 119868 (119901 119902) = 119894 119868 (119901 + Δ119909 119902 + Δ119910) = 119895

0 otherwise

(6)

The values 119894 and 119895 and size of 119862 are determined by the valuesin 119868

The texture analysis for dermoscopic images is performedfor 8-bit grayscale images quantized to 256 values and isreferred to as gray-level cooccurrence matrices (GLCM)

Textural features are frequently used in advanced melan-ocytic lesion analysis [7] Celebi et al suggested uniformlyquantizing the dermoscopy image to 64 gray levels and thenextracting 8 Haralick features for each one of the fourorientations using single pixel offsets [7 19ndash21]The followingparameters have been calculated

maximum probability = max119894119895

119888 (119894 119895)

energy = sum

119894

sum

119895

119888 (119894 119895)2

entropy = minussum

119894

sum

119895

119901 (119894 119895) log (119888 (119894 119895))

dissimilarity =1003816100381610038161003816119894 minus 119895

1003816100381610038161003816119888 (119894 119895)

contrast =119873minus1

sum

119899=0

sum

119894

sum

119895

119888 (119894 119895) |1003816100381610038161003816119894 minus 119895

1003816100381610038161003816= 119899

inverse difference = sum

119888 (119894 119895)

1 +1003816100381610038161003816119894 minus 119895

1003816100381610038161003816

inverse difference moment = sum

119888 (119894 119895)

1 + (119894 minus 119895)2

correlation =

(119894119895) 119901 (119894 119895) minus 120583119909120583119910

120590119909120590119910

(7)

where 119888(119894 119895) is the value of the normalized cooccurrencematrix at indexes (119894 119895)119873 is the number of gray levels 120583

119909and

120583119910are the mean values of the rows and columns of 119888 120590

119909and

120590119910are the respective standard deviations and the symbol |

indicates a condition that must be validSince the range of values of raw data varies widely the

feature scaling step has to be done before feature selection andclassification Feature scaling is a method used to standardizethe range of independent variables or features of data In this

6 BioMed Research International

work we rescale the range of features to scale the range in[minus1 1] The formula is given as

1199091015840

= 2 sdot

119909 minusmin (119909)

max (119909) minusmin (119909)

minus 1 (8)

232 Feature Selection In the next step we select the mostsuitable features from the set of all features described in theprevious sectionThere are many criteria to describe effectivefeatures but they should possess the following advantageslarge interclass mean distance small interclass varianceinsensitivity to extraneous variables and low correlationwithother features [15]

In this work we propose to use the correlation-basedfeature selection (CFS) algorithm CFS relies on a heuristicfor evaluating the worth or merit of a subset of featuresThis heuristic takes into account the usefulness of individualfeatures for predicting the class label along with the level ofintercorrelation among them

If the correlation between each of the components in atest and the outside variable is known and the intercorre-lation between each pair of components is given then thecorrelation between a composite consisting of the summedcomponents and the outside variable can be predicted from[22]

119903119911119888

=

119896119903119911119894

radic119896 + 119896 minus (119896 minus 1) 119903119894119894

(9)

where 119903119911119888

is correlation between the summed componentsand the outside variable 119896 is number of components (fea-tures) 119903

119911119894is average of the correlations between the com-

ponents and the outside variable and 119903119894119894is average intercor-

relation between componentsEquation (9) shows that the correlation between a com-

posite and an outside variable is a function of the numberof component variables in the composite and the magnitudeof the intercorrelations among them together with the mag-nitude of the correlations between the components and theoutside variable [22]

The following features have been chosen for the classifica-tion process compactness solidity symmetry distance colorvariegation features variance of entropy contrast dissimilar-ity and ellipse variance

24 Classification Using Support Vector Machine (SVM)After the set of discriminative features has been selected aclassifier can be designed for the evaluation of micro skinmoles The Support Vector Machine (SVM) is a supervisedlearning model used for data analysis pattern recognitionclassification and regression analysis An SVM trainingalgorithmbuilds amodel that assigns new samples to a relatedclass An SVMmodel is a representation of samples as pointsin space a linear function is used so that the examples of theseparate classes are divided by a clear gap

Given a training set of instance-label pairs (x119894 y119894) 119894 = 1

119897 where x119894isin 119877119899 and y isin 1 minus1

119897 the SVM requires thesolution of the following optimization problem [23]

minw119887120585

1

2

w119879w + 119862

119897

sum

119894=1

120585119894

subject to y119894(w119879120601 (x

119894) + 119887) ge 1 minus 120585

119894

120585119894ge 0

(10)

where training vectors x119894are mapped into a higher dimen-

sional space by the function 120601 The SVM finds a linear sep-arating hyperplane with the maximal margin in this higherdimensional space119862 gt 0 is the penalty parameter of the errorterm

Moreover 119870(x119894 x119895) equiv 120601(x

119894)119879

120601(x119895) is called the kernel

function [23] We use the commonly implemented RadialBasis Function (RBF) defined as

119870(x119894 x119895) = exp (minus120574

1003817100381710038171003817x119894minus x119895

10038171003817100381710038171003817

2

) 120574 gt 0 (11)

25 Model Evaluation Method For the evaluation of thedesigned algorithm we use cross-validation which is a tech-nique for assessing how the results of a statistical analysis willgeneralize to an independent dataset Cross-validation (CV)is primarily a way ofmeasuring the predictive performance ofa statisticalmodel In this approach the so called 119896-fold cross-validation method has been used In 119896-fold cross-validationthe original sample is randomly split into 119896 smaller setsThe following steps are undertaken a model is trained using119896 minus 1 of the folds as training data and the resulting model isvalidated on the remaining part of the data

The performance measure reported by 119896-fold cross-validation is then the average of the values computed in theloop

3 Results and Discussion

31 Database Specification The described algorithm for theautomatic classification of micro-melanocytic lesions hasbeen tested on dermoscopic images from a widely usedInteractive Atlas of Dermoscopy [1] and a private databaseImages for this atlas have been provided by two universityhospitals (University of Naples Italy and University of GrazAustria) and stored on aCD-ROMin a JPEG formatThedoc-umentation of each dermoscopic image was performed usinga Dermaphot apparatus (Heine Optotechnik HerrschingGermany) and a photo camera (Nikon F3) mounted on astereomicroscope (Wild M650 Heerbrugg AG Switzerland)in order to produce digitized ELM images of skin lesions Allof the images have been assessed manually by a dermoscopicexpert with extensive clinical experience

Furthermore all of the descriptions of skin cases werebased on histopathological examination of the biopsy mate-rial In order to develop and test the automatic procedurefor the classification of micro-melanocytic skin lesions 200images with different resolutions ranging from 0033 to

BioMed Research International 7

Table 1 List of parameters evaluated by the proposed method

Desired result Output result TotalBenign Malignant

Benign 65 5 70Malignant 7 123 130

Sensitivity 90 Specificity 96 200

05mmpixel were chosen The database included 70 malig-nant melanomas and 130 benign cases

The preprocessing step (black frame removal and hairremoval) as well as the segmentation step (border error lessthan 6) did not affect the further research [13 14]

32 Statistical Analysis Diagnostic exactness is often used tomean the fraction of cases on which the automated imageanalysis system gives a correct answer where correctness isdetermined by comparing the diagnostic decision to somedefinition of ldquotruthrdquo [15] There are different ways to definethe ldquotruthrdquo In our case we depend on the separate goldstandard which is determined by the results of biopsy for thedermoscopic images

The way to assess the accuracy of the designed classifica-tion system is to compare its results against the gold standardThe analyzed task is a binary problem where the answer canbe ldquo1rdquo when the lesion is malignant or ldquo0rdquo when it is benignTomeasure the diagnostic performanceweuse a pair of statis-tics parameters including sensitivity and specificity whichare not affected by the disease prevalence

The sensitivity or True Positive rate is the probability thatthe lesion is said to be malignant given that it is This can beestimated by relative frequency [15]

sensitivity =

TPTP + FN

(12)

where TP is the True Positive answer and FP is the FalsePositive answer

On the other hand we calculate specificity (False Positiverate) which is defined as the proportion of the True Negativesagainst all negative results Specificity is defined by thefollowing equation

specificity =

TNTN + FN

(13)

where TN is the True Negative answer and FN is the FalseNegative answer

Table 1 shows the classification results obtained with theimplemented algorithm in a confusion matrix

Another dominant technique for assessing computer clas-sification performance is the receiver operation characteristic(ROC) analysis A ROC curve is a plot of sensitivity versus 1 minusspecificity (Figure 6) Area under the ROC curve is the mostcommonly used accuracy index The area under a perfect-accuracy ROC curve when sensitivity and specificity reach10 is also 10 Another interesting fact is that for randomguessing is the positive diagonal line and the area under thecurve is 05 [15]

1009080706050403020100

10

09

08

07

06

05

04

03

02

01

00

Sens

itivi

tyRandom

ROC curve

1 minus specificity

Figure 6 ROC plot for the SVM classifier (AUC = 9324)

The conducted results (sensitivity = 90 specificity =96 and AUC = 9324) achieved by the implementedmicro-melanoma classification system are much better thanthe results described in [8 9] Ourmethod allowed classifica-tion of micro-melanomas very precisely and we can confirmthat the classification ofmicro skin lesions has to be done sep-arately from the classification of developed skin moles Over-all our experiments clearly show that the classification ofmicro skin moles which causes problems can be supportedby a powerful system based on proposed artificial neuralnetwork

4 Conclusions and Future Plans

Due to the difficulty and subjectivity of human interpretationthe computerized image analysis techniques have become animportant tool in the interpretation of dermoscopic imagesIn this research we propose a new approach for the classifica-tion of melanocytic skin moles The results obtained withinthis study indicate that the proposed algorithm can be usedfor the diagnosis of very small skin moles These conclusionswere reached after comparing computer classification withbiopsy outcome using statistical measurements and ROCcurve The automatic procedures have been tested and itseems to be a very promising software tool supporting thephysician The proposed diagnostic system can be used notonly by the young inexperienced dermatologist but also firstand foremost by family physiciansThis can be a great oppor-tunity for people that live in remote and rural areas outsidethe regional centre and frequently find it difficult to make anappointment with a dermatologist The importance of diag-nosing amelanoma in the early stage and the reduction of the

8 BioMed Research International

melanoma-related mortality rate can be achieved by precisecomputer-aided diagnostic systems

However computer classification of micro-melanomasmust also be further developed and tested on a larger datasetbefore it can be used clinically and before its potential inreducing the mortality rate of malignant melanoma can berealized in clinical medicine

Competing Interests

The author declares that there are no competing interestsregarding the publication of this paper

Acknowledgments

The scientific work was partly supported by the NationalScience Centre based on Decision no 201101NST706783and by AGH Research Grant 1511120408

References

[1] G Argenziano H P Soyer V De Giorgio et al Interactive Atlasof Dermoscopy (Book and CD-ROM) Edra Medical Publishingand New Media Milano Italy 2000

[2] Malignant melanoma 2015 httpwwwdermatologycaskin-hair-nailsskinskin-cancermalignant-melanoma

[3] S Seidenari C Ferrari S Borsari et al ldquoDermoscopy of smallmelanomas just miniaturized dermoscopyrdquo British Journal ofDermatology vol 171 no 5 pp 1006ndash1013 2014

[4] A Bono C Bartoli D Moglia et al ldquoSmall melanomas aclinical study on 270 consecutive cases of cutaneousmelanomardquoMelanoma Research vol 9 no 6 pp 583ndash586 1999

[5] C Schmoeckel ldquoSmall malignant melanomas clinicopatho-logic correlation and DNA ploidy analysisrdquo Journal of theAmerican Academy of Dermatology vol 24 no 6 part 1 pp1036ndash1037 1991

[6] S Seidenari S Bassoli S Borsari et al ldquoVariegated dermoscopyof in situ melanomardquoDermatology vol 224 no 3 pp 262ndash2702012

[7] J Scharcanski and M E Celebi Eds Computer Vision Tech-niques for the Diagnosis of Skin Cancer Springer New York NYUSA 2013

[8] A Bono C BartoliM Baldi et al ldquoMicro-melanomadetectionA clinical study on 22 cases of melanoma with a diameter equalto or less than 3mmrdquo Tumori vol 90 no 1 pp 128ndash131 2004

[9] A Bono E Tolomio S Trincone et al ldquoMicro-melanomadetection a clinical study on 206 consecutive cases of pig-mented skin lesions with a diameter le 3mmrdquo British Journalof Dermatology vol 155 no 3 pp 570ndash573 2006

[10] R Tadeusiewicz ldquoPlace and role of intelligent systems incomputer sciencerdquo Computer Methods in Materials Science vol10 no 4 pp 193ndash206 2010

[11] Disease overview 2015 httpsolaranrxcomunmet-medical-needdisease-overview

[12] J Jaworek-Korjakowska and R Tadeusiewicz ldquoHair removalfrom dermoscopic colour imagesrdquo Bio-Algorithms andMed Sys-tems vol 9 no 2 pp 53ndash58 2013

[13] J Jaworek-Korjakowska ldquoNovel method for border irregularityassessment in dermoscopic color imagesrdquo Computational and

MathematicalMethods inMedicine vol 2015 Article ID 49620211 pages 2015

[14] J Jaworek-Korjakowska and R Tadeusiewicz ldquoAssessment ofdots and globules in dermoscopic color images as one of the7-point check list criteriardquo in Proceedings of the 20th IEEE Inter-national Conference on Image Processing (ICIP rsquo13) pp 1456ndash1460 Melbourne Australia September 2013

[15] I N BankmanHandbook of Medical Imaging Academic PressNew York NY USA 2000

[16] V Ng and D Cheung ldquoMeasuring asymmetries of skin lesionsrdquoin Proceedings of the IEEE International Conference on SystemsMan and Cybernetics pp 4211ndash4216 Orlando Fla USA Octo-ber 1997

[17] V T YNg B YM Fung andT K Lee ldquoDetermining the asym-metry of skin lesion with fuzzy bordersrdquo Computers in Biologyand Medicine vol 35 no 2 pp 103ndash120 2005

[18] J Premaladha and K S Ravichandran ldquoAsymmetry analysis ofmalignant melanoma using image processing a surveyrdquo Journalof Artificial Intelligence vol 7 no 2 pp 45ndash53 2014

[19] M E Celebi H A Kingravi B Uddin et al ldquoA methodologicalapproach to the classification of dermoscopy imagesrdquo Comput-erized Medical Imaging and Graphics vol 31 no 6 pp 362ndash3732007

[20] R M Haralick K Shanmugam and I Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973

[21] J Smietanski R Tadeusiewicz and E Łuczynska ldquoTextureanalysis in perfusion images of prostate cancermdasha case studyrdquoInternational Journal of Applied Mathematics and ComputerScience vol 20 no 1 pp 149ndash156 2010

[22] M Nixon Feature Extraction amp Image Processing for ComputerVision Academic Press New York NY USA 3rd edition 2012

[23] httpwwwcsientuedutwsimcjlinpapersguideguidepdf

Submit your manuscripts athttpwwwhindawicom

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Anatomy Research International

PeptidesInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

International Journal of

Volume 2014

Zoology

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Molecular Biology International

GenomicsInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioinformaticsAdvances in

Marine BiologyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Signal TransductionJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioMed Research International

Evolutionary BiologyInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Biochemistry Research International

ArchaeaHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Genetics Research International

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Advances in

Virolog y

Hindawi Publishing Corporationhttpwwwhindawicom

Nucleic AcidsJournal of

Volume 2014

Stem CellsInternational

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Enzyme Research

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

Microbiology

6 BioMed Research International

work we rescale the range of features to scale the range in[minus1 1] The formula is given as

1199091015840

= 2 sdot

119909 minusmin (119909)

max (119909) minusmin (119909)

minus 1 (8)

232 Feature Selection In the next step we select the mostsuitable features from the set of all features described in theprevious sectionThere are many criteria to describe effectivefeatures but they should possess the following advantageslarge interclass mean distance small interclass varianceinsensitivity to extraneous variables and low correlationwithother features [15]

In this work we propose to use the correlation-basedfeature selection (CFS) algorithm CFS relies on a heuristicfor evaluating the worth or merit of a subset of featuresThis heuristic takes into account the usefulness of individualfeatures for predicting the class label along with the level ofintercorrelation among them

If the correlation between each of the components in atest and the outside variable is known and the intercorre-lation between each pair of components is given then thecorrelation between a composite consisting of the summedcomponents and the outside variable can be predicted from[22]

119903119911119888

=

119896119903119911119894

radic119896 + 119896 minus (119896 minus 1) 119903119894119894

(9)

where 119903119911119888

is correlation between the summed componentsand the outside variable 119896 is number of components (fea-tures) 119903

119911119894is average of the correlations between the com-

ponents and the outside variable and 119903119894119894is average intercor-

relation between componentsEquation (9) shows that the correlation between a com-

posite and an outside variable is a function of the numberof component variables in the composite and the magnitudeof the intercorrelations among them together with the mag-nitude of the correlations between the components and theoutside variable [22]

The following features have been chosen for the classifica-tion process compactness solidity symmetry distance colorvariegation features variance of entropy contrast dissimilar-ity and ellipse variance

24 Classification Using Support Vector Machine (SVM)After the set of discriminative features has been selected aclassifier can be designed for the evaluation of micro skinmoles The Support Vector Machine (SVM) is a supervisedlearning model used for data analysis pattern recognitionclassification and regression analysis An SVM trainingalgorithmbuilds amodel that assigns new samples to a relatedclass An SVMmodel is a representation of samples as pointsin space a linear function is used so that the examples of theseparate classes are divided by a clear gap

Given a training set of instance-label pairs (x119894 y119894) 119894 = 1

119897 where x119894isin 119877119899 and y isin 1 minus1

119897 the SVM requires thesolution of the following optimization problem [23]

minw119887120585

1

2

w119879w + 119862

119897

sum

119894=1

120585119894

subject to y119894(w119879120601 (x

119894) + 119887) ge 1 minus 120585

119894

120585119894ge 0

(10)

where training vectors x119894are mapped into a higher dimen-

sional space by the function 120601 The SVM finds a linear sep-arating hyperplane with the maximal margin in this higherdimensional space119862 gt 0 is the penalty parameter of the errorterm

Moreover 119870(x119894 x119895) equiv 120601(x

119894)119879

120601(x119895) is called the kernel

function [23] We use the commonly implemented RadialBasis Function (RBF) defined as

119870(x119894 x119895) = exp (minus120574

1003817100381710038171003817x119894minus x119895

10038171003817100381710038171003817

2

) 120574 gt 0 (11)

25 Model Evaluation Method For the evaluation of thedesigned algorithm we use cross-validation which is a tech-nique for assessing how the results of a statistical analysis willgeneralize to an independent dataset Cross-validation (CV)is primarily a way ofmeasuring the predictive performance ofa statisticalmodel In this approach the so called 119896-fold cross-validation method has been used In 119896-fold cross-validationthe original sample is randomly split into 119896 smaller setsThe following steps are undertaken a model is trained using119896 minus 1 of the folds as training data and the resulting model isvalidated on the remaining part of the data

The performance measure reported by 119896-fold cross-validation is then the average of the values computed in theloop

3 Results and Discussion

31 Database Specification The described algorithm for theautomatic classification of micro-melanocytic lesions hasbeen tested on dermoscopic images from a widely usedInteractive Atlas of Dermoscopy [1] and a private databaseImages for this atlas have been provided by two universityhospitals (University of Naples Italy and University of GrazAustria) and stored on aCD-ROMin a JPEG formatThedoc-umentation of each dermoscopic image was performed usinga Dermaphot apparatus (Heine Optotechnik HerrschingGermany) and a photo camera (Nikon F3) mounted on astereomicroscope (Wild M650 Heerbrugg AG Switzerland)in order to produce digitized ELM images of skin lesions Allof the images have been assessed manually by a dermoscopicexpert with extensive clinical experience

Furthermore all of the descriptions of skin cases werebased on histopathological examination of the biopsy mate-rial In order to develop and test the automatic procedurefor the classification of micro-melanocytic skin lesions 200images with different resolutions ranging from 0033 to

BioMed Research International 7

Table 1 List of parameters evaluated by the proposed method

Desired result Output result TotalBenign Malignant

Benign 65 5 70Malignant 7 123 130

Sensitivity 90 Specificity 96 200

05mmpixel were chosen The database included 70 malig-nant melanomas and 130 benign cases

The preprocessing step (black frame removal and hairremoval) as well as the segmentation step (border error lessthan 6) did not affect the further research [13 14]

32 Statistical Analysis Diagnostic exactness is often used tomean the fraction of cases on which the automated imageanalysis system gives a correct answer where correctness isdetermined by comparing the diagnostic decision to somedefinition of ldquotruthrdquo [15] There are different ways to definethe ldquotruthrdquo In our case we depend on the separate goldstandard which is determined by the results of biopsy for thedermoscopic images

The way to assess the accuracy of the designed classifica-tion system is to compare its results against the gold standardThe analyzed task is a binary problem where the answer canbe ldquo1rdquo when the lesion is malignant or ldquo0rdquo when it is benignTomeasure the diagnostic performanceweuse a pair of statis-tics parameters including sensitivity and specificity whichare not affected by the disease prevalence

The sensitivity or True Positive rate is the probability thatthe lesion is said to be malignant given that it is This can beestimated by relative frequency [15]

sensitivity =

TPTP + FN

(12)

where TP is the True Positive answer and FP is the FalsePositive answer

On the other hand we calculate specificity (False Positiverate) which is defined as the proportion of the True Negativesagainst all negative results Specificity is defined by thefollowing equation

specificity =

TNTN + FN

(13)

where TN is the True Negative answer and FN is the FalseNegative answer

Table 1 shows the classification results obtained with theimplemented algorithm in a confusion matrix

Another dominant technique for assessing computer clas-sification performance is the receiver operation characteristic(ROC) analysis A ROC curve is a plot of sensitivity versus 1 minusspecificity (Figure 6) Area under the ROC curve is the mostcommonly used accuracy index The area under a perfect-accuracy ROC curve when sensitivity and specificity reach10 is also 10 Another interesting fact is that for randomguessing is the positive diagonal line and the area under thecurve is 05 [15]

1009080706050403020100

10

09

08

07

06

05

04

03

02

01

00

Sens

itivi

tyRandom

ROC curve

1 minus specificity

Figure 6 ROC plot for the SVM classifier (AUC = 9324)

The conducted results (sensitivity = 90 specificity =96 and AUC = 9324) achieved by the implementedmicro-melanoma classification system are much better thanthe results described in [8 9] Ourmethod allowed classifica-tion of micro-melanomas very precisely and we can confirmthat the classification ofmicro skin lesions has to be done sep-arately from the classification of developed skin moles Over-all our experiments clearly show that the classification ofmicro skin moles which causes problems can be supportedby a powerful system based on proposed artificial neuralnetwork

4 Conclusions and Future Plans

Due to the difficulty and subjectivity of human interpretationthe computerized image analysis techniques have become animportant tool in the interpretation of dermoscopic imagesIn this research we propose a new approach for the classifica-tion of melanocytic skin moles The results obtained withinthis study indicate that the proposed algorithm can be usedfor the diagnosis of very small skin moles These conclusionswere reached after comparing computer classification withbiopsy outcome using statistical measurements and ROCcurve The automatic procedures have been tested and itseems to be a very promising software tool supporting thephysician The proposed diagnostic system can be used notonly by the young inexperienced dermatologist but also firstand foremost by family physiciansThis can be a great oppor-tunity for people that live in remote and rural areas outsidethe regional centre and frequently find it difficult to make anappointment with a dermatologist The importance of diag-nosing amelanoma in the early stage and the reduction of the

8 BioMed Research International

melanoma-related mortality rate can be achieved by precisecomputer-aided diagnostic systems

However computer classification of micro-melanomasmust also be further developed and tested on a larger datasetbefore it can be used clinically and before its potential inreducing the mortality rate of malignant melanoma can berealized in clinical medicine

Competing Interests

The author declares that there are no competing interestsregarding the publication of this paper

Acknowledgments

The scientific work was partly supported by the NationalScience Centre based on Decision no 201101NST706783and by AGH Research Grant 1511120408

References

[1] G Argenziano H P Soyer V De Giorgio et al Interactive Atlasof Dermoscopy (Book and CD-ROM) Edra Medical Publishingand New Media Milano Italy 2000

[2] Malignant melanoma 2015 httpwwwdermatologycaskin-hair-nailsskinskin-cancermalignant-melanoma

[3] S Seidenari C Ferrari S Borsari et al ldquoDermoscopy of smallmelanomas just miniaturized dermoscopyrdquo British Journal ofDermatology vol 171 no 5 pp 1006ndash1013 2014

[4] A Bono C Bartoli D Moglia et al ldquoSmall melanomas aclinical study on 270 consecutive cases of cutaneousmelanomardquoMelanoma Research vol 9 no 6 pp 583ndash586 1999

[5] C Schmoeckel ldquoSmall malignant melanomas clinicopatho-logic correlation and DNA ploidy analysisrdquo Journal of theAmerican Academy of Dermatology vol 24 no 6 part 1 pp1036ndash1037 1991

[6] S Seidenari S Bassoli S Borsari et al ldquoVariegated dermoscopyof in situ melanomardquoDermatology vol 224 no 3 pp 262ndash2702012

[7] J Scharcanski and M E Celebi Eds Computer Vision Tech-niques for the Diagnosis of Skin Cancer Springer New York NYUSA 2013

[8] A Bono C BartoliM Baldi et al ldquoMicro-melanomadetectionA clinical study on 22 cases of melanoma with a diameter equalto or less than 3mmrdquo Tumori vol 90 no 1 pp 128ndash131 2004

[9] A Bono E Tolomio S Trincone et al ldquoMicro-melanomadetection a clinical study on 206 consecutive cases of pig-mented skin lesions with a diameter le 3mmrdquo British Journalof Dermatology vol 155 no 3 pp 570ndash573 2006

[10] R Tadeusiewicz ldquoPlace and role of intelligent systems incomputer sciencerdquo Computer Methods in Materials Science vol10 no 4 pp 193ndash206 2010

[11] Disease overview 2015 httpsolaranrxcomunmet-medical-needdisease-overview

[12] J Jaworek-Korjakowska and R Tadeusiewicz ldquoHair removalfrom dermoscopic colour imagesrdquo Bio-Algorithms andMed Sys-tems vol 9 no 2 pp 53ndash58 2013

[13] J Jaworek-Korjakowska ldquoNovel method for border irregularityassessment in dermoscopic color imagesrdquo Computational and

MathematicalMethods inMedicine vol 2015 Article ID 49620211 pages 2015

[14] J Jaworek-Korjakowska and R Tadeusiewicz ldquoAssessment ofdots and globules in dermoscopic color images as one of the7-point check list criteriardquo in Proceedings of the 20th IEEE Inter-national Conference on Image Processing (ICIP rsquo13) pp 1456ndash1460 Melbourne Australia September 2013

[15] I N BankmanHandbook of Medical Imaging Academic PressNew York NY USA 2000

[16] V Ng and D Cheung ldquoMeasuring asymmetries of skin lesionsrdquoin Proceedings of the IEEE International Conference on SystemsMan and Cybernetics pp 4211ndash4216 Orlando Fla USA Octo-ber 1997

[17] V T YNg B YM Fung andT K Lee ldquoDetermining the asym-metry of skin lesion with fuzzy bordersrdquo Computers in Biologyand Medicine vol 35 no 2 pp 103ndash120 2005

[18] J Premaladha and K S Ravichandran ldquoAsymmetry analysis ofmalignant melanoma using image processing a surveyrdquo Journalof Artificial Intelligence vol 7 no 2 pp 45ndash53 2014

[19] M E Celebi H A Kingravi B Uddin et al ldquoA methodologicalapproach to the classification of dermoscopy imagesrdquo Comput-erized Medical Imaging and Graphics vol 31 no 6 pp 362ndash3732007

[20] R M Haralick K Shanmugam and I Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973

[21] J Smietanski R Tadeusiewicz and E Łuczynska ldquoTextureanalysis in perfusion images of prostate cancermdasha case studyrdquoInternational Journal of Applied Mathematics and ComputerScience vol 20 no 1 pp 149ndash156 2010

[22] M Nixon Feature Extraction amp Image Processing for ComputerVision Academic Press New York NY USA 3rd edition 2012

[23] httpwwwcsientuedutwsimcjlinpapersguideguidepdf

Submit your manuscripts athttpwwwhindawicom

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Anatomy Research International

PeptidesInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

International Journal of

Volume 2014

Zoology

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Molecular Biology International

GenomicsInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioinformaticsAdvances in

Marine BiologyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Signal TransductionJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioMed Research International

Evolutionary BiologyInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Biochemistry Research International

ArchaeaHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Genetics Research International

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Advances in

Virolog y

Hindawi Publishing Corporationhttpwwwhindawicom

Nucleic AcidsJournal of

Volume 2014

Stem CellsInternational

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Enzyme Research

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

Microbiology

BioMed Research International 7

Table 1 List of parameters evaluated by the proposed method

Desired result Output result TotalBenign Malignant

Benign 65 5 70Malignant 7 123 130

Sensitivity 90 Specificity 96 200

05mmpixel were chosen The database included 70 malig-nant melanomas and 130 benign cases

The preprocessing step (black frame removal and hairremoval) as well as the segmentation step (border error lessthan 6) did not affect the further research [13 14]

32 Statistical Analysis Diagnostic exactness is often used tomean the fraction of cases on which the automated imageanalysis system gives a correct answer where correctness isdetermined by comparing the diagnostic decision to somedefinition of ldquotruthrdquo [15] There are different ways to definethe ldquotruthrdquo In our case we depend on the separate goldstandard which is determined by the results of biopsy for thedermoscopic images

The way to assess the accuracy of the designed classifica-tion system is to compare its results against the gold standardThe analyzed task is a binary problem where the answer canbe ldquo1rdquo when the lesion is malignant or ldquo0rdquo when it is benignTomeasure the diagnostic performanceweuse a pair of statis-tics parameters including sensitivity and specificity whichare not affected by the disease prevalence

The sensitivity or True Positive rate is the probability thatthe lesion is said to be malignant given that it is This can beestimated by relative frequency [15]

sensitivity =

TPTP + FN

(12)

where TP is the True Positive answer and FP is the FalsePositive answer

On the other hand we calculate specificity (False Positiverate) which is defined as the proportion of the True Negativesagainst all negative results Specificity is defined by thefollowing equation

specificity =

TNTN + FN

(13)

where TN is the True Negative answer and FN is the FalseNegative answer

Table 1 shows the classification results obtained with theimplemented algorithm in a confusion matrix

Another dominant technique for assessing computer clas-sification performance is the receiver operation characteristic(ROC) analysis A ROC curve is a plot of sensitivity versus 1 minusspecificity (Figure 6) Area under the ROC curve is the mostcommonly used accuracy index The area under a perfect-accuracy ROC curve when sensitivity and specificity reach10 is also 10 Another interesting fact is that for randomguessing is the positive diagonal line and the area under thecurve is 05 [15]

1009080706050403020100

10

09

08

07

06

05

04

03

02

01

00

Sens

itivi

tyRandom

ROC curve

1 minus specificity

Figure 6 ROC plot for the SVM classifier (AUC = 9324)

The conducted results (sensitivity = 90 specificity =96 and AUC = 9324) achieved by the implementedmicro-melanoma classification system are much better thanthe results described in [8 9] Ourmethod allowed classifica-tion of micro-melanomas very precisely and we can confirmthat the classification ofmicro skin lesions has to be done sep-arately from the classification of developed skin moles Over-all our experiments clearly show that the classification ofmicro skin moles which causes problems can be supportedby a powerful system based on proposed artificial neuralnetwork

4 Conclusions and Future Plans

Due to the difficulty and subjectivity of human interpretationthe computerized image analysis techniques have become animportant tool in the interpretation of dermoscopic imagesIn this research we propose a new approach for the classifica-tion of melanocytic skin moles The results obtained withinthis study indicate that the proposed algorithm can be usedfor the diagnosis of very small skin moles These conclusionswere reached after comparing computer classification withbiopsy outcome using statistical measurements and ROCcurve The automatic procedures have been tested and itseems to be a very promising software tool supporting thephysician The proposed diagnostic system can be used notonly by the young inexperienced dermatologist but also firstand foremost by family physiciansThis can be a great oppor-tunity for people that live in remote and rural areas outsidethe regional centre and frequently find it difficult to make anappointment with a dermatologist The importance of diag-nosing amelanoma in the early stage and the reduction of the

8 BioMed Research International

melanoma-related mortality rate can be achieved by precisecomputer-aided diagnostic systems

However computer classification of micro-melanomasmust also be further developed and tested on a larger datasetbefore it can be used clinically and before its potential inreducing the mortality rate of malignant melanoma can berealized in clinical medicine

Competing Interests

The author declares that there are no competing interestsregarding the publication of this paper

Acknowledgments

The scientific work was partly supported by the NationalScience Centre based on Decision no 201101NST706783and by AGH Research Grant 1511120408

References

[1] G Argenziano H P Soyer V De Giorgio et al Interactive Atlasof Dermoscopy (Book and CD-ROM) Edra Medical Publishingand New Media Milano Italy 2000

[2] Malignant melanoma 2015 httpwwwdermatologycaskin-hair-nailsskinskin-cancermalignant-melanoma

[3] S Seidenari C Ferrari S Borsari et al ldquoDermoscopy of smallmelanomas just miniaturized dermoscopyrdquo British Journal ofDermatology vol 171 no 5 pp 1006ndash1013 2014

[4] A Bono C Bartoli D Moglia et al ldquoSmall melanomas aclinical study on 270 consecutive cases of cutaneousmelanomardquoMelanoma Research vol 9 no 6 pp 583ndash586 1999

[5] C Schmoeckel ldquoSmall malignant melanomas clinicopatho-logic correlation and DNA ploidy analysisrdquo Journal of theAmerican Academy of Dermatology vol 24 no 6 part 1 pp1036ndash1037 1991

[6] S Seidenari S Bassoli S Borsari et al ldquoVariegated dermoscopyof in situ melanomardquoDermatology vol 224 no 3 pp 262ndash2702012

[7] J Scharcanski and M E Celebi Eds Computer Vision Tech-niques for the Diagnosis of Skin Cancer Springer New York NYUSA 2013

[8] A Bono C BartoliM Baldi et al ldquoMicro-melanomadetectionA clinical study on 22 cases of melanoma with a diameter equalto or less than 3mmrdquo Tumori vol 90 no 1 pp 128ndash131 2004

[9] A Bono E Tolomio S Trincone et al ldquoMicro-melanomadetection a clinical study on 206 consecutive cases of pig-mented skin lesions with a diameter le 3mmrdquo British Journalof Dermatology vol 155 no 3 pp 570ndash573 2006

[10] R Tadeusiewicz ldquoPlace and role of intelligent systems incomputer sciencerdquo Computer Methods in Materials Science vol10 no 4 pp 193ndash206 2010

[11] Disease overview 2015 httpsolaranrxcomunmet-medical-needdisease-overview

[12] J Jaworek-Korjakowska and R Tadeusiewicz ldquoHair removalfrom dermoscopic colour imagesrdquo Bio-Algorithms andMed Sys-tems vol 9 no 2 pp 53ndash58 2013

[13] J Jaworek-Korjakowska ldquoNovel method for border irregularityassessment in dermoscopic color imagesrdquo Computational and

MathematicalMethods inMedicine vol 2015 Article ID 49620211 pages 2015

[14] J Jaworek-Korjakowska and R Tadeusiewicz ldquoAssessment ofdots and globules in dermoscopic color images as one of the7-point check list criteriardquo in Proceedings of the 20th IEEE Inter-national Conference on Image Processing (ICIP rsquo13) pp 1456ndash1460 Melbourne Australia September 2013

[15] I N BankmanHandbook of Medical Imaging Academic PressNew York NY USA 2000

[16] V Ng and D Cheung ldquoMeasuring asymmetries of skin lesionsrdquoin Proceedings of the IEEE International Conference on SystemsMan and Cybernetics pp 4211ndash4216 Orlando Fla USA Octo-ber 1997

[17] V T YNg B YM Fung andT K Lee ldquoDetermining the asym-metry of skin lesion with fuzzy bordersrdquo Computers in Biologyand Medicine vol 35 no 2 pp 103ndash120 2005

[18] J Premaladha and K S Ravichandran ldquoAsymmetry analysis ofmalignant melanoma using image processing a surveyrdquo Journalof Artificial Intelligence vol 7 no 2 pp 45ndash53 2014

[19] M E Celebi H A Kingravi B Uddin et al ldquoA methodologicalapproach to the classification of dermoscopy imagesrdquo Comput-erized Medical Imaging and Graphics vol 31 no 6 pp 362ndash3732007

[20] R M Haralick K Shanmugam and I Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973

[21] J Smietanski R Tadeusiewicz and E Łuczynska ldquoTextureanalysis in perfusion images of prostate cancermdasha case studyrdquoInternational Journal of Applied Mathematics and ComputerScience vol 20 no 1 pp 149ndash156 2010

[22] M Nixon Feature Extraction amp Image Processing for ComputerVision Academic Press New York NY USA 3rd edition 2012

[23] httpwwwcsientuedutwsimcjlinpapersguideguidepdf

Submit your manuscripts athttpwwwhindawicom

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Anatomy Research International

PeptidesInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

International Journal of

Volume 2014

Zoology

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Molecular Biology International

GenomicsInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioinformaticsAdvances in

Marine BiologyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Signal TransductionJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioMed Research International

Evolutionary BiologyInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Biochemistry Research International

ArchaeaHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Genetics Research International

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Advances in

Virolog y

Hindawi Publishing Corporationhttpwwwhindawicom

Nucleic AcidsJournal of

Volume 2014

Stem CellsInternational

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Enzyme Research

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

Microbiology

8 BioMed Research International

melanoma-related mortality rate can be achieved by precisecomputer-aided diagnostic systems

However computer classification of micro-melanomasmust also be further developed and tested on a larger datasetbefore it can be used clinically and before its potential inreducing the mortality rate of malignant melanoma can berealized in clinical medicine

Competing Interests

The author declares that there are no competing interestsregarding the publication of this paper

Acknowledgments

The scientific work was partly supported by the NationalScience Centre based on Decision no 201101NST706783and by AGH Research Grant 1511120408

References

[1] G Argenziano H P Soyer V De Giorgio et al Interactive Atlasof Dermoscopy (Book and CD-ROM) Edra Medical Publishingand New Media Milano Italy 2000

[2] Malignant melanoma 2015 httpwwwdermatologycaskin-hair-nailsskinskin-cancermalignant-melanoma

[3] S Seidenari C Ferrari S Borsari et al ldquoDermoscopy of smallmelanomas just miniaturized dermoscopyrdquo British Journal ofDermatology vol 171 no 5 pp 1006ndash1013 2014

[4] A Bono C Bartoli D Moglia et al ldquoSmall melanomas aclinical study on 270 consecutive cases of cutaneousmelanomardquoMelanoma Research vol 9 no 6 pp 583ndash586 1999

[5] C Schmoeckel ldquoSmall malignant melanomas clinicopatho-logic correlation and DNA ploidy analysisrdquo Journal of theAmerican Academy of Dermatology vol 24 no 6 part 1 pp1036ndash1037 1991

[6] S Seidenari S Bassoli S Borsari et al ldquoVariegated dermoscopyof in situ melanomardquoDermatology vol 224 no 3 pp 262ndash2702012

[7] J Scharcanski and M E Celebi Eds Computer Vision Tech-niques for the Diagnosis of Skin Cancer Springer New York NYUSA 2013

[8] A Bono C BartoliM Baldi et al ldquoMicro-melanomadetectionA clinical study on 22 cases of melanoma with a diameter equalto or less than 3mmrdquo Tumori vol 90 no 1 pp 128ndash131 2004

[9] A Bono E Tolomio S Trincone et al ldquoMicro-melanomadetection a clinical study on 206 consecutive cases of pig-mented skin lesions with a diameter le 3mmrdquo British Journalof Dermatology vol 155 no 3 pp 570ndash573 2006

[10] R Tadeusiewicz ldquoPlace and role of intelligent systems incomputer sciencerdquo Computer Methods in Materials Science vol10 no 4 pp 193ndash206 2010

[11] Disease overview 2015 httpsolaranrxcomunmet-medical-needdisease-overview

[12] J Jaworek-Korjakowska and R Tadeusiewicz ldquoHair removalfrom dermoscopic colour imagesrdquo Bio-Algorithms andMed Sys-tems vol 9 no 2 pp 53ndash58 2013

[13] J Jaworek-Korjakowska ldquoNovel method for border irregularityassessment in dermoscopic color imagesrdquo Computational and

MathematicalMethods inMedicine vol 2015 Article ID 49620211 pages 2015

[14] J Jaworek-Korjakowska and R Tadeusiewicz ldquoAssessment ofdots and globules in dermoscopic color images as one of the7-point check list criteriardquo in Proceedings of the 20th IEEE Inter-national Conference on Image Processing (ICIP rsquo13) pp 1456ndash1460 Melbourne Australia September 2013

[15] I N BankmanHandbook of Medical Imaging Academic PressNew York NY USA 2000

[16] V Ng and D Cheung ldquoMeasuring asymmetries of skin lesionsrdquoin Proceedings of the IEEE International Conference on SystemsMan and Cybernetics pp 4211ndash4216 Orlando Fla USA Octo-ber 1997

[17] V T YNg B YM Fung andT K Lee ldquoDetermining the asym-metry of skin lesion with fuzzy bordersrdquo Computers in Biologyand Medicine vol 35 no 2 pp 103ndash120 2005

[18] J Premaladha and K S Ravichandran ldquoAsymmetry analysis ofmalignant melanoma using image processing a surveyrdquo Journalof Artificial Intelligence vol 7 no 2 pp 45ndash53 2014

[19] M E Celebi H A Kingravi B Uddin et al ldquoA methodologicalapproach to the classification of dermoscopy imagesrdquo Comput-erized Medical Imaging and Graphics vol 31 no 6 pp 362ndash3732007

[20] R M Haralick K Shanmugam and I Dinstein ldquoTexturalfeatures for image classificationrdquo IEEE Transactions on SystemsMan and Cybernetics vol 3 no 6 pp 610ndash621 1973

[21] J Smietanski R Tadeusiewicz and E Łuczynska ldquoTextureanalysis in perfusion images of prostate cancermdasha case studyrdquoInternational Journal of Applied Mathematics and ComputerScience vol 20 no 1 pp 149ndash156 2010

[22] M Nixon Feature Extraction amp Image Processing for ComputerVision Academic Press New York NY USA 3rd edition 2012

[23] httpwwwcsientuedutwsimcjlinpapersguideguidepdf

Submit your manuscripts athttpwwwhindawicom

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Anatomy Research International

PeptidesInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

International Journal of

Volume 2014

Zoology

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Molecular Biology International

GenomicsInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioinformaticsAdvances in

Marine BiologyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Signal TransductionJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioMed Research International

Evolutionary BiologyInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Biochemistry Research International

ArchaeaHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Genetics Research International

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Advances in

Virolog y

Hindawi Publishing Corporationhttpwwwhindawicom

Nucleic AcidsJournal of

Volume 2014

Stem CellsInternational

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Enzyme Research

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

Microbiology

Submit your manuscripts athttpwwwhindawicom

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Anatomy Research International

PeptidesInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporation httpwwwhindawicom

International Journal of

Volume 2014

Zoology

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Molecular Biology International

GenomicsInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioinformaticsAdvances in

Marine BiologyJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Signal TransductionJournal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

BioMed Research International

Evolutionary BiologyInternational Journal of

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Biochemistry Research International

ArchaeaHindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Genetics Research International

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Advances in

Virolog y

Hindawi Publishing Corporationhttpwwwhindawicom

Nucleic AcidsJournal of

Volume 2014

Stem CellsInternational

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Enzyme Research

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

International Journal of

Microbiology