Research Article Computer Based Melanocytic and Nevus ...downloads.hindawi.com/journals/bmri/2016/2082589.pdfsegmentation technique based on double thresholding for segmentation. Edge-based

Research ArticleComputer Based Melanocytic and Nevus ImageEnhancement and Segmentation

Uzma Jamil12 M Usman Akram3 Shehzad Khalid1 Sarmad Abbas3 and Kashif Saleem1

1Department of Computer Engineering Bahria University Islamabad Pakistan2Government College University Faisalabad Pakistan3National University of Sciences amp Technology Islamabad Pakistan

Correspondence should be addressed to M Usman Akram usmakramgmailcom

Received 10 November 2015 Accepted 18 July 2016

Academic Editor Hesham H Ali

Copyright copy 2016 Uzma Jamil et al This is an open access article distributed under the Creative Commons Attribution Licensewhich permits unrestricted use distribution and reproduction in any medium provided the original work is properly cited

Digital dermoscopy aids dermatologists in monitoring potentially cancerous skin lesions Melanoma is the 5th common form ofskin cancer that is rare but the most dangerous Melanoma is curable if it is detected at an early stage Automated segmentationof cancerous lesion from normal skin is the most critical yet tricky part in computerized lesion detection and classification Theeffectiveness and accuracy of lesion classification are critically dependent on the quality of lesion segmentation In this paperwe have proposed a novel approach that can automatically preprocess the image and then segment the lesion The system filtersunwanted artifacts including hairs gel bubbles and specular reflection A novel approach is presented using the concept of waveletsfor detection and inpainting the hairs present in the cancer images The contrast of lesion with the skin is enhanced using adaptivesigmoidal function that takes care of the localized intensity distribution within a given lesionrsquos images We then present a segmen-tation approach to precisely segment the lesion from the background The proposed approach is tested on the European databaseof dermoscopic images Results are compared with the competitors to demonstrate the superiority of the suggested approach

1 Introduction

Skin cancer early detection is very important because ofits successful and economic treatment Skin cancer is themost common form of cancer which is rapidly increasing inrecent decades [1] Malignant melanoma (MM) is the mostcommon type of skin cancer which is usually found in whiteskin people but has also been rarely seen in the dark skinindividuals Early stage diagnosis of skin cancer is extremelycritical for its treatment Amongst various types of skincancer melanoma is one of the most fatal diseases which hasthe highest death rate Melanoma originates in melanocyteswhich are the cells in the skin that produce pigment ormelanin It mostly occurs on those parts of the human bodythat are exposed to sunlight such as head neck arms trunkand legs Nowadays manual visual diagnosis of melanomasby trained professionals is most commonly used in thedetection and classification of melanomas Melanoma skincancer can further be categorized into malignant melanomaand nonmalignant melanoma Malignant melanoma is the

least common but yet the more aggressive of the two typesof skin cancer

Dermoscopy also called dermatoscopy or skin surfacemicroscopy is the technique of investigating the skin lesionsthat helps professionals in cancer examination Digital der-moscopy is considered to be preferred approach due to itscorrectness and accuracy in the results Digital cancer imagesare used in Computer Aided Diagnostic (CAD) systems forscreening ofmelanoma and its different stagesDifferent signsof disease appear with different properties on the surfaceof lesion and it is the goal of CAD systems to identifythese signs for timely and accurate treatment of cancerAutomated detection of melanoma is comprised of varioussteps including preprocessing extracting region of interestpostprocessing and finally segmentation

Image segmentation is the process of segmenting a digitalimage into multiple partitions The aim of segmentation isto filter irrelevant information from the image that is notthe part of lesion Dermoscopic images are generally affectedby certain artifacts including smooth transition between

Hindawi Publishing CorporationBioMed Research InternationalVolume 2016 Article ID 2082589 13 pageshttpdxdoiorg10115520162082589

2 BioMed Research International

(a) (b)

(c) (d)

Figure 1 Sample dermoscopic images containing malignant melanoma and highlighting the variations in the melanoma lesions (a) Smoothtransition between lesion and the image (b) Specular reflection (c) Presence of hairs (d) Gel and bubble presence

lesion and the skin presence of hairs transition effects ofgel and water bubble multiple colored lesions and specularcolor reflections Segmentation is further problematic due toirregular shapes [2] and sizes of lesions with different texturesand skin types [3 4] Some sample digital dermoscopicimages of malignant melanoma highlighting these artifactsare presented in Figure 1 Preprocessing steps are requiredto handle these artifacts which will otherwise have negativeimpact on the feature computation and in turn skin cancercategorization After removal of problematic artifacts regionof interest from the image is extracted and the lesion is seg-mented from the image After segmenting we try to identifyand compute features which makes different categories ofmelanoma distinct from each otherThese features are furtherused by classifiers such as [5ndash8] to automatically recognizedifferent types of melanoma This paper is primarily focusedon the first two steps of dermoscopic analysis of melanomathat include preprocessing and segmentation [9] Image pro-cessing approaches [10 11] and research in related domainssuch as retinal lesion detection and classification [12ndash14] arealso somewhat relevant to the problem at hand of skin lesionsegmentation feature extraction and classification

In this paper we present our segmentation approachfor extracting lesion from skin whilst taking care of theproblems of gel bubbles hairs vessels contrast variationsand other artifacts The main contribution of the paper isa novel approach for effectively handling the problem ofhairs and vessels enhancing the luminance information andstretching the contrast between skin and lesion pixels for

effective segmentation of lesion In this paper we present aneffective approach to handle the problems of unwanted arti-facts such as hairs and tiny vessels by employing directionalwavelet filters and enhancingdetecting pixels representingthese artifactsThe detected hairs and vessel pixels are filteredand a novel inpainting approach is presented to fill themissing pixels using neighborhood informationTheproblemof uneven luminance is also addressed by estimating nonuni-form illumination and performing equalization in luminanceinformation The problem of contrast stretching betweenskin and lesion is addressed by proposing adaptive sigmoidalfunction that computes and utilizes cut-off value suitable forindividual images The enhanced image is then processed tosegment lesion from skin using combination of thresholdingand morphological operations The proposed approach istested on the European database of dermoscopic images

The remainder of the paper is organized as followsIn Section 2 we present a review of recent lesion seg-ment approaches Section 3 presents the overview of ourproposed lesion segmentation approach In Section 4 wepresent our novel methodology for effective detection ofhairs and later removing them using proposed inpaintingapproach Section 5 presents our proposed contrast enhance-mentstretching and lesion segmentation approach usingenhanced dermoscopic image Experiments are conductedto demonstrate the superiority of the proposed approachas compared with the competitors These experiments arediscussed in Section 6 The last section summarizes theproposed approach and the experimental findings

BioMed Research International 3

2 Related Work

To deal with preprocessing and segmentation problemsmany algorithms have been proposed that can be broadlyclassified as thresholding edge-based and region-basedmethods In [15] a thresholding based segmentation ap-proach is presentedThey used combination of global thresh-olding adaptive thresholding and clustering techniquesTheir approach attains good results when there is goodcontrast between the lesion and the skin However theirapproach will not provide impressive performance wherethere is a poor contrast and smooth transition between skinand lesion which is a norm In assumption of bimodal dis-tribution if a global threshold is applied it will result in poorsegmentation results Abbas et al [16] presented an automaticsegmentation technique based on double thresholding forsegmentation

Edge-based lesion segmentation approaches are pre-sented in [17ndash19] Gonzalez and Woods [17] presented anapproach for lesion segmentation that is based on thezero-crossings of the Laplacian-of-Gaussian A variety ofactive-contour based approaches have been proposed [1819] Argenziano et al [18] presented gradient vector flow(GVF) based active-contour model whereas Celebi et al [19]presented geodesic active-contour model and the geodesicedge tracing for lesion segmentation Edge-based approachesperform poorly when the boundaries of the lesions are notwell defined and the color transition between skin and lesionis smooth Another difficulty is the presence of fake edgepoints caused by the presence of artifacts such as hair colorreflections or irregularities in the skin texture which are notthe part of the lesion boundary [20]

Region-based approaches have also been employedwhichinclude multiscale region growing [21] direction sensi-tive modified fuzzy c-means algorithm [22] morphologicalflooding [15] multiresolution Markov random field algo-rithm [18] and statistical region merging [23] Region-based approaches have difficulties in the presence of vari-ation in color andor the presence of texture in lesion orskin region leading to oversegmentation Gomez et al [24]provide a comparison of different techniques presented tosegment lesions in dermoscopic images including adaptivethresholding [25] fuzzy c-means [26] spherical coordinatetransform (SCT)center split [27] principal componentstransform (PCT)median cut [27] split and merge [28] andmultiresolution segmentation [28] They have not includedany edge-based techniques in their comparative analysis

In the past decades a variety of techniques have beenproposed for skin lesion detection or segmentation [18 1929 30] based on thresholding clustering and region growingin gray scale Unsupervised techniques based on statisticalregion merging (SRM) and color-texture (JSEG) algorithmsare proposed in [23] and [31] respectively Abbas et al [16]proposed a segmentation technique based on region-basedactive contour (RAC) Lissner and Urban [23] introduced animproved dermatologist-like tumor area extraction (DTAE)algorithm In addition the remaining techniques are mostlybased on nonuniform color spaces [32 33]

According to Emre Celebi et al in [34] in automatedskin lesion diagnostic system boundary detection system andimage acquisition process should be described in sufficientdetail Testing should be done from random collection ofimages from a large and diverse image database Test set ofimages should be large enough to ensure statistically validconclusions Test image group should not be used to train theboundary detection method Distribution of the diagnosticimage sets should be specifiedThe algorithmwith reasonablecomputational requirements should be used The evaluationof the results to determine lesion boundaries should becompared with marked boundaries taken from different der-matologistsThe results should be compared with other stud-ies published The implementation of boundary detectionmethod should be public to improve accessibility and reuse

Another skin cancer segmentation approach is variationalmodel for image segmentation presented in [35] Mean shiftbased gradient vector flow algorithm was validated againstcompeting methods including classical GVF and level setand it provided best accuracy and robustness Among themost advanced technologies this method is quite accuratebecause it gets iterative energy minimization process the bestsolution The algorithm incorporates a function of the massdensity with classic GVF term Achieving this final solution isbased on the integration of support functions and mean driftestimate numerical optimization program [27]

3 Overview of ProposedSegmentation Framework

In this section we present an overview of our proposedapproach for the segmentation of melanocytic and nevuslesions from skin whilst handling the problems of hairs gelinconsistent contrast and other artifacts The proposed algo-rithm is composed of three major steps including hair detec-tion and inpainting color space transformation and contraststretchingenhancement and then finally segmentation of thelesion area The flow diagram of the proposed methodologyis presented in Figure 2 The proposed system takes adermoscopic image as an input and removesmajor unwantedartifact of hairs to avoid its effect on subsequent steps of imageenhancement and segmentation Hair removal is achievedby highlighting and detecting hairs and further removingthem by inpainting the hair pixels using their neighborhoodinformation The system then performs desired color spacetransformation and performs image enhancement on theselected component of color space to better differentiatebetween skin and lesion Lesions are then segmented fromthe background skin image in the final step

4 Hair Artifacts Removal

In this section we present our proposed approach forremoval of hair artifacts before segmentation of lesionsThis is a classification-free method which enhances thehair information using Gabor wavelet-based directional fil-tersenhancement and then mitigates its effect using inpaint-ing based segmentation technique employing neighborhood


Hair artifacts removal

Image enhancement andlesion segmentation

Input skin cancer image

Hair enhancement andsegmentation

Hair inpainting

Contrast enhancement andstretching

Lesion segmentation

Output image withsegmented lesion

Figure 2 Phases of automated skin lesion segmentation system

Hairdetection

Enhancement

Hairinpainting

Figure 3 Flow diagram for hair artifacts removal algorithm

estimation The proposed approach for hair artifact removalfocuses on reducing the effect of hair in lesion segmentationwhich appears as false positive and degrades system per-formance Figure 3 shows flow diagram of proposed systemfor hair artifact removal Hair artifact removal method firstdetects all the hairs present in image and creates a binarymask of all hairs The detected hair regions are then filledusing an algorithm of Neighborhood Based Region Filling(NBRF) The key idea of this method is to inpaint the hairsbefore lesion enhancement and segmentation so as to avoidthe enhancement of artifacts upon applying enhancementand segmentation techniques

41 Hair Enhancement and Segmentation Themajor artifactthat creates problem in lesion segmentation is the presenceof thin hairs in the image So it is necessary to improve the

image pattern We have used 2D Gabor wavelet to enhancethe image pattern and highlight the hair definition [25]Thesewavelets are the best option to tune in frequencies due to theirpotential of fine feature detection In order to apply Gaborwavelet we used continuous time wavelet transformation(CWT) [26] The 2D CWT 119879

Φ(b 120579 119886) is defined in terms of

the scalar product of 119891(119909 119910) with the transformed waveletΦb120579119886(x)

119879Φ (b 120579 119886)

= 119882minus12

Φ119886int exp (119895Rb) Φlowast (119886119903

minus120579R) (R) 1198892R

(1)

Here 119895 = radicminus1 and Φ is Fourier transform ofΦ119882ΦΦlowast b 120579

and 119886 denote the normalizing constant complex conjugate ofΦ the displacement vector the rotation angle and the dila-tion parameter respectively 119903

minus120579is two-dimensional rotation

along x The mathematical functions for Gabor wavelet andits Fourier transform are defined as

ΦGabor (x) = exp (119895R0x) exp(minus1

2|Ax|2)

ΦGabor (x) = (det119860minus1)12

exp(minus1

2(119860minus1

(R minus R0)2))

(2)

where x = [119909 119910]119879 andR

0is a vector that defines the frequency

of the complex exponential119860 = [ 120598minus12

0

0 1]with elongation 120598 ge

1 is a 2 times 2 positive definite diagonal matrix which definesthe wavelet anisotropy and elongation of filter in any desireddirection For each pixel position with fixed values of b and119886 the Gabor wavelet transform 119879

Φ(b 120579 119886) is computed for 120579


(a) (b)

(c)

Figure 4 Hair detection (a) Original image (b) Hair enhancement using Gabor wavelet (c) Binary hair mask using adaptive thresholding

Table 1 Parameter values for Gabor wavelet

Parameter ValueDilation (119886) 25Elongation (120598) 3Rotation angle (120579) 15∘

1198960

[0 3]

spanning from 0∘ up to 165∘ at steps of 15∘ and the maximum

is taken

119872Φ (120579) = max 1003816100381610038161003816119879Φ (b 120579 119886)

1003816100381610038161003816 (3)

Table 1 shows the values for Gabor wavelet parameters whichare used for enhancement of hairs

The next step is to make binary mask of hair and it isdone by applying adaptive OTSUrsquos thresholding algorithm[14] Figure 4 shows hair enhancement and binary maskgeneration results

42 Hair Inpainting Hair detection module makes a binarymask for all hairs present in the image The next step is toremove all hair pixels and fill those pixels to have a smoothimageHair inpainting step takes binary hairmask as an inputand fills all the hair regions in a very smooth manner byusing an algorithm of Neighborhood Based Region Filling(NBRF) NBRF algorithm works in radial way towards thecenter of the objects It fills hairs in a very smooth way byestimating the neighborhood and averaging the backgroundIt runs iteratively which eventually blends all hairs withinthe filled region Algorithm 1 shows the algorithm for NBRF

The algorithm runs in a recursive manner and fills all hairpixels using morphological operations until all pixels arefilled Figure 5 shows the results of image with inpainted hairsobtained after applying NBRF algorithm

5 Lesion Segmentation afterContrast Enhancement and Stretching inLuminance (119871) Space

In this section we present our approach for image enhance-ment by performing contrast enhancement and stretchingOur proposed approach operates on the luminance (L) com-ponent of the 119871lowast119886lowast119887lowast color space It has been observed thatthere is awide variation in luminance information of differentimages Similarly the contrast between the skin and the lesionis critically dependent on lighting conditions and tone ofskin and lesion pixels This results in significant variation incontrasts of skin and lesion pixels over different dermoscopicimages Equalization of luminance and enhancement ofcontrast between skin and lesion pixel are therefore criticalfor the success of lesion segmentation algorithm

51 Contrast Enhancement In this section we present anapproach for correction of illumination in the extracted Lchannel of the image based on imaging condition and thegeometry knowledge to provide a visually standard value oflightning for all of the images This approach can be triviallyextended to correct the color components of the dermoscopicimages if required The proposed solution works with thefollowing steps


(a) (b)

(c)Figure 5 (a) Original image (b) Image after removal of hair pixels (c) Image after filling hair pixels using NBRF

Step 1Calculate the hair mask

HairMask larr dilate(HairMask)Step 2

Set the intensity values of original image equals to zero where intensities of the Hair Mask are not equal to zeroTmpInp larr Org Img(HairMask = 0) = 0

Step 3Repeat the process while all Hairs are not inpainted

Do ErosionHairMask larr erode(HairMask)

Perform Hair InpaintingTmpInp larr 119888119886119897119897 HairInp(TmpInpHairMask)

End of the Iterative processStep 4

Final Inpainted image is obtainedImgInp larr TmpInp

119868 = HairInp(IHairMask)PxToFill larr HairMask minus erode(HairMask)forall119901isin PxToFill | PxToFill (119901) = 0

119868(119901) = mean 119868(119902)

Algorithm 1 Stepwise specification of NBRF Algorithm

511 Estimation of Nonuniform Illumination in L ChannelThisphase takes luminance (L) channel as input andperformsthe nonuniform illumination estimation present in L channelof the image using the following steps

(1) Compute initial background mask (potentially repre-senting the skin pixels) as follows

bg (119909 119910) =

1 iff 119871 (119909 119910) gt 12 lowast thresh (119871)

0 otherwise(4)

where thresh(sdot) is a function that automaticallydetects a threshold value to binarize gray scale imageas proposed in [36]


(2) Calculate 119906119871(119909 119910) as follows

119906119871(119909 119910) =

(119871 (119909 119910) lowast bg (119909 119910)) oplus 119866119898(119909 119910)

bg (119909 119910) oplus 119866119898(119909 119910)

(5)

where

119866119898(1198991 1198992) =

minus (1198992

1+ 1198992

2minus 21205902) ℎ119892(1198991 1198992)

21205871205906sum1198991sum1198992ℎ119892(1198991 1198992)

ℎ119892(1198991 1198992) = exp

minus (1198992

1+ 1198992

2)

21205902

(6)

is Gaussian kernel with the standardized parameters(3) Compute the illumination estimation 119864

119871as follows

119864119871(119909 119910) =

119866119891119898

119866 (119909 119910) lowast 119906 (119909 119910) oplus 119866119898(119909 119910)

119866119891119898

(119909 119910) oplus 119866119898(119909 119910)

(7)

where

119866119891119898

(119909 119910) = (119906119871(119909 119910) minus 119871 (119909 119910) le 005) lowast bg (119909 119910) (8)

512 Equalization of L Channel The estimated nonuniformluminance distribution component119864

119871 computed using (7) is

then used to generate the equalizedL channel (119871eq) as follows

119871eq (119909 119910) = 119871 ref lowast119871 (119909 119910)

119864119871(119909 119910)

(9)

where 119871 ref is a constant used to adjust the overall illuminationof the image We assume 119871 ref = 059 based on empiricalevaluation

513 Contrast Stretching In this section we present ourproposed approach to stretch the contrast of the dermoscopicimage to enhance the illumination variation between skinand lesion pixels To achieve this we normalize the equalizedluminance image (119871eq) using

119871eq (119909 119910) =119871eq (119909 119910) minus 119871eq min

119896 (119871eq max minus 119871eq min)

for 119871eq min le 119871eq (119909 119910) le 119871eq max

(10)

where 119871eq min and 119871eq max are the 1st and 99th percentile val-ues respectively within the equalized L component 119871eq(119909 119910)and 119896 is the constant that determines the intensity of skinpixels other than lesions The value of 119896 is determinedempirically and is set to 075

The contrast between the luminance of skin and lesionpixels is enhanced by employing adaptive sigmoidal usingcut-off value computed separately for each dermoscopicimage A good value for cut-off 120575 is approximated by com-puting cumulative histogram of 119871eq imageWe can employ 119887-bit quantization for the computation of cumulative histogramwhich implies that the histogram of 119871eq component is

quantized using 2119887 bins Employing higher number of bins

makes the proposed approach sensitive to local fluctuationsin histogram whereas lower number of bins will result inrough approximation of 120575 We have employed 119887 = 4 basedon empirical evaluation Let CH

119871represent the cumulative

histogram of enhanced luminance component of the image(119871eq) the value for 120575 is approximated as follows

120575 = weierp lowast arg max(119894)

(CH1015840119871(119894)) lowast 2

119887 (11)

where 119894 is the bin index of histogram CH119871 CH1015840119871is first-order

derivative of CH119871 and weierp is the scaling parameter with values

0 le weierp le 1 Lower values of weierp result in enhancing theluminance of even relatively darker pixels (representing partof lesions) to be detected as part of the skin and vice versaThe value ofweierp is determined empirically and is set to 088Wethen update the luminance information by applying adaptivesigmoidal function based on the dynamic cut-off value 120575 aslearned using (11) as follows

119871 (119909 119910) =1

1 + exp (gain lowast (120575 minus 119871 (119909 119910)))forall119909 119910 (12)

where the value of gain is determined empirically and we setgain = 10 Qualitative analysis of adaptive sigmoidal functionas compared to traditional sigmoidal function with varyingcut-off values is presented in Figure 6 It is apparent fromFigure 6 that the static value for cut-off gives good contraststretching between lesion and skin pixels for some imagesbut deteriorates the contrast in other images On the otherhand computation of dynamic cut-off in adaptive sigmoidalfunction enhances the contrast for all the dermoscopicimages

52 Lesion Segmentation After contrast enhancement andstretching of luminance component the target lesion differsgreatly in contrast from the skin part We further applymedian filtering (119871

119872= median filtering (119871eq)) with a disk

shaped structuring element to handle the presence of fine-level noise We then perform a sequence of morphologicaloperations to make the lesion definition more clear to berobustly segmented from the skin We first employ morpho-logical erosion operation on 119871 channel (119871

119874= open (119871

119872))

using a disk shaped structuring element 119871119874is later used

as a marker to perform morphological reconstruction withthe median filtered luminance image (119871

119872) as the mask

Morphological reconstruction can be thought of conceptuallyas repeated dilations of marker image until the contour ofthe marker image fits under mask image Morphologicalreconstruction has been employed using 8-connective neigh-borhood (119871MC = morphological reconstruction (119871

119872)) In

morphological reconstruction the peaks in themarker imagespread out or dilate This process is followed by closingand erosion by a relatively smaller disk shaped structuringelement to merge pixels at the border of lesion whilsteliminating fine-level noise elsewhere This will have a neteffect of removing small blemishes without affecting the


(a) (b) (c) (d) (e) (f)Figure 6 Qualitative analysis of proposed adaptive sigmoidal function (a) Original images (b) Before applying sigmoidal function Contraststretching achieved after applying sigmoidal function with (c) 120575 = 065 (d) 120575 = 075 and (e) 120575 = 08 (f) Proposed adaptive value for cut-off120575

overall shapes of the lesion object The processed image isthen binarized as follows

119871119861(119909 119910) =

1 119871MC (119909 119910) lt thresh (119871119861) minus 120581

0 otherwise

forall119909 119910

(13)

where thresh(sdot) is a function that automatically detects athreshold value to binarize based on the histogram of 119871MCas proposed in [36] and 120581 is a constant which determines thesensitivity of our proposed segmentation algorithm to detectthe boundary pixels of lesions as foreground or backgroundAs there is sometimes a smooth transition from lesion pixels

to skin pixels lower values of 120581 will result in shrinking oflesion regions and result in boundary pixels to be classifiedas skin region and vice versa We assumed 120581 = 008 based onempirical evaluation

We than employ connected component labelling using 8-connective neighborhood to identify distinct object in119871

119861 Let

O be the list of detected objects (inclusive of noisy objects)the filtered list by removing the unwanted objects Ofiltered isthen generated as follows

Ofiltered = Ofiltered isin O | forall119874 isin Ofiltered |119874| lt 20

and Solidity (119874119894) gt 025

(14)


where | sdot | is a pixel count function for a given object andSolidity(sdot) is a function that computes the proportion of thepixels in the convex hull that are also in the object The pixelsbelonging to objectobjects in Ofiltered are the segmentedpixels representing the lesion in the image

6 Results and Discussion

In this section we present results to demonstrate theeffectiveness of proposed approach for the segmentation ofmelanocytic and nevus lesions in the presence of unwantedartifacts of hairs gel illumination variation and so forthVarious experiments are performed to evaluate the validityof the proposed approach and to provide comparison withexisting approaches The lesion segmentation technique istested effectively on dataset of a total of 100 dermoscopyimages This dataset consists of both invasive malignantmelanoma and benign lesions

61 Performance Metrics The performance metrics used forquantitative analysis of proposed approach and its compari-son with the competitors are True Positive Rate (TPR) FalsePositive Rate (FPR) and error probability (EP) measuresThese evaluations metrics are computed as follows

True Detection Rate (TDR) = TPTP + FN

lowast 100

False Positive Rate (FPR) = TNFP + TN

lowast 100

Error rate (ER) = FP + FNTP + TN + FP + FN

lowast 100

(15)

where TP TN FP and FN are true positives true negativesfalse positive and false negatives respectively

Experiment 1 (evaluation of proposed hair artifact removaland image inpainting approach) The purpose of this exper-iment is to perform a qualitative evaluation of our proposedapproach to remove the hair artifacts from the dermoscopicimages to mitigate its effect on lesion segmentation Theproposed approach whilst filtering the noisy hair pixels fillsfiltered pixels using the neighboring skin pixels Figure 7highlights the effectiveness of the proposed hair detectionand inpainting approach as presented in Section 4 on variousdermoscopic images containing hair artifacts The imagescontaining varying density of hairs are selected to highlightthe performance of proposed approach on different possiblescenarios of the presence of hair artifacts As obvious fromFigure 7 the proposed approach successfully detects all hairsin the image and then correctly replaces them with the skininformation from the neighboring skin pixels whilst havingminimal side effects of hair removal The successful removalof hair artifacts contributes significantly to the segmentationof lesions from skin

Experiment 2 (quantitative evaluation of proposed contrastenhancement and adaptive contrast stretching) The pur-pose of this experiment is to analyze the contribution of

Table 2 Performance analysis of proposed dynamic contrastenhancement approach as compared to static alternative

Cut-off value TDR FPR EP055 8808 226 491065 9156 30 452071 9595 454 452075 9728 520 467080 9733 610 547Adaptive value 9726 352 301

proposed contrast enhancement and stretching approach inthe effectiveness of lesion segmentation To perform theanalysis of contrast enhancement component as proposedin Section 51 we performed complete lesion segmentationprocess compared the segmentation results with groundtruth and computed the performancemetrics as discussed inSection 61The experiments are repeated with different staticvalues of cut-off and other parameters employed in sigmoidalfunction typically used in literature and are comparedwith the dynamic values of these parameters computed asproposed in Section 51 Performance metrics of TDR FPRand ER computed for these different settings are presentedin Table 2 Ideally we want higher value of TDR (TrueDetection Rate) and lower values of FPR (False PositiveRate) and ER (error rate) As obvious from the resultsthe proposed dynamic computation of cut-off parameter insigmoidal function significantly enhances the performanceof proposed system as compared to user-specified globalcut-off values for all images This improved performance isexplained by the fact that each image has different luminanceand significant contrast variations between skin and lesionThus a static cut-off value working good for one image maygive poor results for the other imagesThese phenomena havebeen highlighted in Figure 6 through qualitative results Theproposed approach to employ dynamically computed valueof cut-off results in overall good values for TDR FPR andEP Setting cut-off value to 08 resulted in the highest TDRbut results in lots of false positives and hence poor FPR andEP values On the other hand cut-off value of 065 resultedin lowering of FPR and EP values but significantly degradingthe TDR

Experiment 3 (comparison of proposed approach with com-petitors) In this experiment we compare the performanceof proposed lesion segmentation approach with existingapproaches including melanoma border detection (MBD)algorithm color-texture algorithm (JSeg) dermatologisttumor area (DTEA) detection and region-based active-contour (RAC) algorithm [16 18 19 30] The lesion segmen-tation is carried out using 100 dermoscopic images and theperformance metrics of TDR FPR and EP computed forvarious algorithms are presented in Table 3 It is observedthat the proposed approach achieves the higher average valueof TDRof 9726 value of FPRof 352 and error probabilityof 301 The obtained results indicate that the proposedalgorithm obtains better melanoma border detection results


(a)

(b)

Figure 7 Qualitative evaluation of proposed hair artifact detection and image inpainting approach (a) Original image with hair artifacts(b) Corresponding processed images after hair removal

Table 3 Performance comparison of proposed techniquewith othermethods of segmentation

Cut-off value TDR FPR EPJSeg 8420 1246 15DTAE 8750 978 13RAC 8941 834 07MBD 9425 356 04Proposed method withcut-off = 075 9728 520 467

Proposed method withadaptive cut-off 9726 352 301

which are significant comparedwith other three state-of-the-art techniques Better results of the proposed approach can beattributed to the removal of hair artifacts and dynamic natureof contrast enhancement and stretching which results insignificant variation in skin and lesion pixels to be identifiedwith significant degree of accuracy Figure 8 shows somequalitative results to demonstrate the effectiveness of theproposed lesion segmentation approach The red outlineof the lesion depicts the manual segmentation done bydermatologist and blue outline represents the segmentationachieved by the proposed system It is clear from the figurethat the proposed approach gives good segmentation resultswhilst handling the artifacts of hairs veins gel and varyingluminance and contrasts However still there are minor over-and undersegmentation of lesion These are normally causedby variation in color and intensity within the lesion thatmatches with the background skin Adaptive sigmoidal func-tion occasionally suppresses the pixels of multitone lesionsThese pixelsmatch the skin or some artifacts in the skinwhichmight not be removed by preprocessing and falsely appearas part of lesion Overall the proposed approach gives goodperformance as depicted by results presented in Figure 8

Figure 9 shows the pictorial comparison of proposedsystem with MBD as compared to the ground truth Forthe competitorrsquos result the black outline represents groundtruth and the blue outline represents the segmentation results

obtained using MBD The difference between ground truthand MBD technique is clearly highlighted in the figure Theproposed method gives almost the same boundary as givenin the ground truth

In order to compute the computational time of proposedsystem we run our algorithm and existing algorithms onCore i-5 21 Ghz system with 4GB RAM The average pro-cessing time of the proposed approach is 176 sec Many othertechniques are doing their jobs well but complexity of thecomputation has become their drawback The processingtime of MBD is 10 sec Similarly other techniquesrsquo processingtime is as follows classical GVF 12 sec level sets 14 secMGVF 17 sec MSGVF 22 sec and so forth Our proposedtechnique is efficient enough with respect to computationtime

7 Conclusion

In this paper we have proposed a novel method of seg-menting melanocytic and nevus lesions from skin usingdermoscopic images The proposed approach handles theproblems of unwanted artifacts such as hairs and tiny ves-sels by employing directional wavelet filters enhancing anddetecting pixels representing these artifactsThe detected andfiltered hairs and vessel pixels and the missing pixels areinpainted using neighborhood information using our pro-posed NBRF algorithm The problem of uneven luminanceis also addressed by estimating nonuniform illuminationand performing equalization in luminance information Thecontrast stretching between skin and lesion is further doneusing proposed adaptive sigmoidal function that computesand utilizes cut-off value suitable for individual images Theenhanced image is then processed to segment lesion fromskin using combination of thresholding and morphologicaloperations The proposed algorithm was tested on 100 der-moscopy images containing invasive malignant melanomanevus and benign lesions The performance of the proposedapproach is comparedwith existing approaches (JSeg DTEAand RAC) The experimental results as presented in Sec-tion 7 demonstrate the superiority of the proposed approach


Figure 8 Lesion segmentation using proposed approach (blue contour) along with ground truth (red contour) imposed on variousdermoscopic images

as compared with competitors The proposed approach givesthe best values of performancemetrics with TDRof 972618FPR of 362 and error probability of 339 Qualitativeresults of segmentation as presented in Figure 8 depict theability of proposed approach to handle the problems ofhairs and vessels and problems associated with luminance

and contrasts Figure 9 further highlights the superiority ofproposed approach as compared to the closest competitor(MBD) The experimental results give evidence to the claimthat the proposed lesion segmentation approach is veryeffective approach that is robust to the presence of unwantedartifacts of hairs vessels variable luminance and contrasts


(1)

(2)

Figure 9 Segmentation comparison results (a) Results of our proposed method (b) Results of MBD technique

Competing Interests

The authors do not have a direct financial relation that mightlead to a conflict of interests for any of them

References

[1] G Argenziano H P Soyer V D Giorgio et al ldquoDer-moscopyrdquo An Interactive Atlas EDRA Medical Publishing2000 httpwwwdermoscopyorg

[2] S Khalid ldquoRobust shape matching using global feature spacerepresentation of contoursrdquo in Proceedings of the InternationalConference on Computing Networking and Communications(ICNC rsquo12) pp 724ndash728 IEEE Maui Hawaii USA February2012

[3] G S Vennila L P Suresh and K L Shunmuganathan ldquoDer-moscopic image segmentation and classification using machinelearning algorithmsrdquo in Proceedings of the International Con-ference on Computing Electronics and Electrical Technologies(ICCEET rsquo12) pp 1122ndash1127 Kumaracoil India March 2012

[4] M Binder M Schwarz A Winkler et al ldquoEpiluminescencemicroscopy A useful tool for the diagnosis of pigmentedskin lesions for formally trained dermatologistsrdquo Archives ofDermatology vol 131 no 3 pp 286ndash291 1995

[5] S Khalid and S Razzaq ldquoFrameworks for multivariate m-mediods based modeling and classification in Euclidean andgeneral feature spacesrdquo Pattern Recognition vol 45 no 3 pp1092ndash1103 2012

[6] S Khalid and S Arshad ldquoA robust ensemble based approachto combine heterogeneous classifiers in the presence of classlabel noiserdquo inProceedings of the 5th International Conference onComputational Intelligence Modelling and Simulation pp 157ndash162 Seoul South Korea September 2013

[7] A Waheed M U Akram S Khalid Z Waheed M A Khanand A Shaukat ldquoHybrid features and mediods classification

based robust segmentation of blood vesselsrdquo Journal of MedicalSystems vol 39 article 128 2015

[8] M T Khan and S Khalid ldquoSentiment analysis for healthcarerdquo International Journal of Privacy and Health InformationManagement vol 3 no 2 pp 78ndash91 2015

[9] Q AbbasM E Celebi and I F Garcıa ldquoSkin tumor area extrac-tion using an improved dynamic programming approachrdquo SkinResearch and Technology vol 18 no 2 pp 133ndash142 2012

[10] A Paul J Wu J-F Yang and J Jeong ldquoGradient-based edgedetection for motion estimation in H264AVCrdquo IET ImageProcessing vol 5 no 4 pp 323ndash327 2011

[11] J Wu A Paul Y Xing et al ldquoMorphological dilation imagecoding with context weights predictionrdquo Signal ProcessingImage Communication vol 25 no 10 pp 717ndash728 2010

[12] M Usman Akram S Khalid A Tariq S A Khan and F AzamldquoDetection and classification of retinal lesions for grading ofdiabetic retinopathyrdquo Computers in Biology and Medicine vol45 no 1 pp 161ndash171 2014

[13] MUAkramA Tariq S KhalidMY Javed S Abbas andUUYasin ldquoGlaucoma detection using novel optic disc localizationhybrid feature set and classification techniquesrdquo AustralasianPhysical amp Engineering Sciences in Medicine vol 38 no 4 pp643ndash655 2015

[14] M Usman Akram S Khalid A Tariq and M Younus JavedldquoDetection of neovascularization in retinal images using mul-tivariate m-Mediods based classifierrdquo Computerized MedicalImaging and Graphics vol 37 no 5-6 pp 346ndash357 2013

[15] P Soille Morphological Image Analysis Principles and Applica-tions Springer Berline Germany 2nd edition 2003

[16] Q Abbas I Fondon and M Rashid ldquoUnsupervised skinlesions border detection via two-dimensional image analysisrdquoComputer Methods and Programs in Biomedicine vol 104 no 3pp e1ndashe15 2011

[17] R C Gonzalez and R E Woods Digital Image ProcessingPrentice-Hall Englewood Cliffs NJ USA 2002


[18] G Argenziano P H Soyer V G De P Carli and M DelfinoInteractive Atlas of Dermoscopy CD EDRA Medical Publishingand New Media Milan Italy 2002

[19] M E Celebi H A Kingravi H Iyatomi et al ldquoBorder detectionin dermoscopy images using statistical region mergingrdquo SkinResearch and Technology vol 14 no 3 pp 347ndash353 2008

[20] P Schmid ldquoSegmentation of digitized dermatoscopic imagesby two-dimensional color clusteringrdquo IEEE Transactions onMedical Imaging vol 18 no 2 pp 164ndash171 1999

[21] K Hoffmann T Gambichler A Rick et al ldquoDiagnostic andneural analysis of skin cancer (DANAOS) A multicentre studyfor collection and computer-aided analysis of data from pig-mented skin lesions using digital dermoscopyrdquo British Journalof Dermatology vol 149 no 4 pp 801ndash809 2003

[22] S Dinesh ldquoApplication of opening by reconstruction to charac-terize the size distribution of catchments extracted from digitalelevation modelsrdquo Applied Mathematical Sciences vol 1 no 13pp 615ndash628 2007

[23] I Lissner and P Urban ldquoToward a unified color space forperception-based image processingrdquo IEEE Transactions onImage Processing vol 21 no 3 pp 1153ndash1168 2012

[24] D D Gomez C Butakoff B K Ersboslashll andW Stoecker ldquoInde-pendent histogram pursuit for segmentation of skin lesionsrdquoIEEE Transactions on Biomedical Engineering vol 55 no 1 pp157ndash161 2008

[25] M Usman Akram and S A Khan ldquoMultilayered thresholding-based blood vessel segmentation for screening of diabeticretinopathyrdquo Engineering with Computers vol 29 no 2 pp 165ndash173 2013

[26] G Schaefer M I Rajab M E Celebi and H Iyatomi ldquoColourand contrast enhancement for improved skin lesion segmenta-tionrdquo Computerized Medical Imaging and Graphics vol 35 no2 pp 99ndash104 2011

[27] R C Gonzalez R E Woods and S L Eddins Digital ImageProcessingUsingMatlab PrenticeHall NewYorkNYUSA 2ndedition 2007

[28] A Arneodo N Decoster and S G Roux ldquoA wavelet-basedmethod for multifractal image analysis I Methodology andtest applications on isotropic and anisotropic random roughsurfacesrdquo European Physical Journal B vol 15 no 3 pp 567ndash600 2000

[29] L Xu M Jackowski A Goshtasby et al ldquoSegmentation of skincancer imagesrdquo Image and Vision Computing vol 17 no 1 pp65ndash74 1999

[30] M E Celebi Y A Aslandogan W V Stoecker H IyatomiH Oka and X Chen ldquoUnsupervised border detection indermoscopy imagesrdquo Skin Research and Technology vol 13 no4 pp 454ndash462 2007

[31] Q Abbas I F Garcia M Emre Celebi W Ahmad andQ Mushtaq ldquoA perceptually oriented method for contrastenhancement and segmentation of dermoscopy imagesrdquo SkinResearch and Technology vol 19 no 1 pp e490ndashe497 2013

[32] H Iyatomi H Oka M E Celebi et al ldquoComputer-basedclassification of dermoscopy images of melanocytic lesions onacral volar skinrdquo Journal of Investigative Dermatology vol 128no 8 pp 2049ndash2054 2008

[33] H Zhou G Schaefer A H Sadka and M E CelebildquoAnisotropic mean shift based fuzzy C-means segmentation ofdermoscopy imagesrdquo IEEE Journal on Selected Topics in SignalProcessing vol 3 no 1 pp 26ndash34 2009

[34] M Emre Celebi Q Wen H Iyatomi K Shimizu H Zhouand G Schaefer A State-of-the-Art Survey on Lesion BorderDetection in Dermoscopy Images chapter September 2015httpswwwresearchgatenetpublication282124553

[35] H Zhou X Li G Schaefer M Emre Celebi and P MillerldquoMean shift based gradient vector ow for image segmentationrdquoComputer Vision and Image Understanding vol 117 no 9 pp1004ndash1016 2013

[36] N Otsu ldquoA threshold selection method for gray level his-togramsrdquo IEEE Transactions on Systems Man and Cyberneticsvol 9 no 1 pp 62ndash66 1979

Submit your manuscripts athttpwwwhindawicom

Hindawi Publishing Corporationhttpwwwhindawicom Volume 2014

Anatomy Research International

PeptidesInternational Journal of


Hindawi Publishing Corporation httpwwwhindawicom

International Journal of

Volume 2014

Zoology


Molecular Biology International

GenomicsInternational Journal of


The Scientific World JournalHindawi Publishing Corporation httpwwwhindawicom Volume 2014


BioinformaticsAdvances in

Marine BiologyJournal of



Signal TransductionJournal of


BioMed Research International

Evolutionary BiologyInternational Journal of



Biochemistry Research International

ArchaeaHindawi Publishing Corporationhttpwwwhindawicom Volume 2014


Genetics Research International


Advances in

Virolog y

Hindawi Publishing Corporationhttpwwwhindawicom

Nucleic AcidsJournal of

Volume 2014

Stem CellsInternational



Enzyme Research



Microbiology


(a) (b)

(c) (d)

Figure 1 Sample dermoscopic images containing malignant melanoma and highlighting the variations in the melanoma lesions (a) Smoothtransition between lesion and the image (b) Specular reflection (c) Presence of hairs (d) Gel and bubble presence

lesion and the skin presence of hairs transition effects ofgel and water bubble multiple colored lesions and specularcolor reflections Segmentation is further problematic due toirregular shapes [2] and sizes of lesions with different texturesand skin types [3 4] Some sample digital dermoscopicimages of malignant melanoma highlighting these artifactsare presented in Figure 1 Preprocessing steps are requiredto handle these artifacts which will otherwise have negativeimpact on the feature computation and in turn skin cancercategorization After removal of problematic artifacts regionof interest from the image is extracted and the lesion is seg-mented from the image After segmenting we try to identifyand compute features which makes different categories ofmelanoma distinct from each otherThese features are furtherused by classifiers such as [5ndash8] to automatically recognizedifferent types of melanoma This paper is primarily focusedon the first two steps of dermoscopic analysis of melanomathat include preprocessing and segmentation [9] Image pro-cessing approaches [10 11] and research in related domainssuch as retinal lesion detection and classification [12ndash14] arealso somewhat relevant to the problem at hand of skin lesionsegmentation feature extraction and classification

In this paper we present our segmentation approachfor extracting lesion from skin whilst taking care of theproblems of gel bubbles hairs vessels contrast variationsand other artifacts The main contribution of the paper isa novel approach for effectively handling the problem ofhairs and vessels enhancing the luminance information andstretching the contrast between skin and lesion pixels for

effective segmentation of lesion In this paper we present aneffective approach to handle the problems of unwanted arti-facts such as hairs and tiny vessels by employing directionalwavelet filters and enhancingdetecting pixels representingthese artifactsThe detected hairs and vessel pixels are filteredand a novel inpainting approach is presented to fill themissing pixels using neighborhood informationTheproblemof uneven luminance is also addressed by estimating nonuni-form illumination and performing equalization in luminanceinformation The problem of contrast stretching betweenskin and lesion is addressed by proposing adaptive sigmoidalfunction that computes and utilizes cut-off value suitable forindividual images The enhanced image is then processed tosegment lesion from skin using combination of thresholdingand morphological operations The proposed approach istested on the European database of dermoscopic images

The remainder of the paper is organized as followsIn Section 2 we present a review of recent lesion seg-ment approaches Section 3 presents the overview of ourproposed lesion segmentation approach In Section 4 wepresent our novel methodology for effective detection ofhairs and later removing them using proposed inpaintingapproach Section 5 presents our proposed contrast enhance-mentstretching and lesion segmentation approach usingenhanced dermoscopic image Experiments are conductedto demonstrate the superiority of the proposed approachas compared with the competitors These experiments arediscussed in Section 6 The last section summarizes theproposed approach and the experimental findings


2 Related Work
















Hair inpainting


Lesion segmentation



Hairdetection

Enhancement

Hairinpainting







119879Φ (b 120579 119886)

= 119882minus12


minus120579R) (R) 1198892R

(1)






2|Ax|2)


exp(minus1

2(119860minus1

(R minus R0)2))

(2)

where x = [119909 119910]119879 andR



0





(a) (b)

(c)




1198960

[0 3]


is taken

119872Φ (120579) = max 1003816100381610038161003816119879Φ (b 120579 119886)

1003816100381610038161003816 (3)









(a) (b)











119868(119901) = mean 119868(119902)




bg (119909 119910) =


0 otherwise(4)




119906119871(119909 119910) =

(119871 (119909 119910) lowast bg (119909 119910)) oplus 119866119898(119909 119910)

bg (119909 119910) oplus 119866119898(119909 119910)

(5)

where

119866119898(1198991 1198992) =

minus (1198992

1+ 1198992

2minus 21205902) ℎ119892(1198991 1198992)

21205871205906sum1198991sum1198992ℎ119892(1198991 1198992)

ℎ119892(1198991 1198992) = exp

minus (1198992

1+ 1198992

2)

21205902

(6)


119871as follows

119864119871(119909 119910) =

119866119891119898

119866 (119909 119910) lowast 119906 (119909 119910) oplus 119866119898(119909 119910)

119866119891119898

(119909 119910) oplus 119866119898(119909 119910)

(7)

where

119866119891119898





119871eq (119909 119910) = 119871 ref lowast119871 (119909 119910)

119864119871(119909 119910)

(9)



119871eq (119909 119910) =119871eq (119909 119910) minus 119871eq min



(10)








(CH1015840119871(119894)) lowast 2

119887 (11)




119871 (119909 119910) =1






119874= open (119871

119872))



119872) as the mask


119872)) In





119871119861(119909 119910) =


0 otherwise

forall119909 119910

(13)




119861 Let



and Solidity (119874119894) gt 025

(14)







lowast 100


lowast 100


lowast 100

(15)









(a)

(b)









7 Conclusion







(1)

(2)


Competing Interests


References














































Volume 2014

Zoology






















Advances in

Virolog y



Volume 2014




Enzyme Research



Microbiology


2 Related Work
















Hair inpainting


Lesion segmentation



Hairdetection

Enhancement

Hairinpainting







119879Φ (b 120579 119886)

= 119882minus12


minus120579R) (R) 1198892R

(1)






2|Ax|2)


exp(minus1

2(119860minus1

(R minus R0)2))

(2)

where x = [119909 119910]119879 andR



0





(a) (b)

(c)




1198960

[0 3]


is taken

119872Φ (120579) = max 1003816100381610038161003816119879Φ (b 120579 119886)

1003816100381610038161003816 (3)









(a) (b)











119868(119901) = mean 119868(119902)




bg (119909 119910) =


0 otherwise(4)




119906119871(119909 119910) =

(119871 (119909 119910) lowast bg (119909 119910)) oplus 119866119898(119909 119910)

bg (119909 119910) oplus 119866119898(119909 119910)

(5)

where

119866119898(1198991 1198992) =

minus (1198992

1+ 1198992

2minus 21205902) ℎ119892(1198991 1198992)

21205871205906sum1198991sum1198992ℎ119892(1198991 1198992)

ℎ119892(1198991 1198992) = exp

minus (1198992

1+ 1198992

2)

21205902

(6)


119871as follows

119864119871(119909 119910) =

119866119891119898

119866 (119909 119910) lowast 119906 (119909 119910) oplus 119866119898(119909 119910)

119866119891119898

(119909 119910) oplus 119866119898(119909 119910)

(7)

where

119866119891119898





119871eq (119909 119910) = 119871 ref lowast119871 (119909 119910)

119864119871(119909 119910)

(9)



119871eq (119909 119910) =119871eq (119909 119910) minus 119871eq min



(10)








(CH1015840119871(119894)) lowast 2

119887 (11)




119871 (119909 119910) =1






119874= open (119871

119872))



119872) as the mask


119872)) In





119871119861(119909 119910) =


0 otherwise

forall119909 119910

(13)




119861 Let



and Solidity (119874119894) gt 025

(14)







lowast 100


lowast 100


lowast 100

(15)









(a)

(b)









7 Conclusion







(1)

(2)


Competing Interests


References














































Volume 2014

Zoology






















Advances in

Virolog y



Volume 2014




Enzyme Research



Microbiology






Hair inpainting


Lesion segmentation



Hairdetection

Enhancement

Hairinpainting







119879Φ (b 120579 119886)

= 119882minus12


minus120579R) (R) 1198892R

(1)






2|Ax|2)


exp(minus1

2(119860minus1

(R minus R0)2))

(2)

where x = [119909 119910]119879 andR



0





(a) (b)

(c)




1198960

[0 3]


is taken

119872Φ (120579) = max 1003816100381610038161003816119879Φ (b 120579 119886)

1003816100381610038161003816 (3)









(a) (b)











119868(119901) = mean 119868(119902)




bg (119909 119910) =


0 otherwise(4)




119906119871(119909 119910) =

(119871 (119909 119910) lowast bg (119909 119910)) oplus 119866119898(119909 119910)

bg (119909 119910) oplus 119866119898(119909 119910)

(5)

where

119866119898(1198991 1198992) =

minus (1198992

1+ 1198992

2minus 21205902) ℎ119892(1198991 1198992)

21205871205906sum1198991sum1198992ℎ119892(1198991 1198992)

ℎ119892(1198991 1198992) = exp

minus (1198992

1+ 1198992

2)

21205902

(6)


119871as follows

119864119871(119909 119910) =

119866119891119898

119866 (119909 119910) lowast 119906 (119909 119910) oplus 119866119898(119909 119910)

119866119891119898

(119909 119910) oplus 119866119898(119909 119910)

(7)

where

119866119891119898





119871eq (119909 119910) = 119871 ref lowast119871 (119909 119910)

119864119871(119909 119910)

(9)



119871eq (119909 119910) =119871eq (119909 119910) minus 119871eq min



(10)








(CH1015840119871(119894)) lowast 2

119887 (11)




119871 (119909 119910) =1






119874= open (119871

119872))



119872) as the mask


119872)) In





119871119861(119909 119910) =


0 otherwise

forall119909 119910

(13)




119861 Let



and Solidity (119874119894) gt 025

(14)







lowast 100


lowast 100


lowast 100

(15)









(a)

(b)









7 Conclusion







(1)

(2)


Competing Interests


References














































Volume 2014

Zoology






















Advances in

Virolog y



Volume 2014




Enzyme Research



Microbiology


(a) (b)

(c)




1198960

[0 3]


is taken

119872Φ (120579) = max 1003816100381610038161003816119879Φ (b 120579 119886)

1003816100381610038161003816 (3)









(a) (b)











119868(119901) = mean 119868(119902)




bg (119909 119910) =


0 otherwise(4)




119906119871(119909 119910) =

(119871 (119909 119910) lowast bg (119909 119910)) oplus 119866119898(119909 119910)

bg (119909 119910) oplus 119866119898(119909 119910)

(5)

where

119866119898(1198991 1198992) =

minus (1198992

1+ 1198992

2minus 21205902) ℎ119892(1198991 1198992)

21205871205906sum1198991sum1198992ℎ119892(1198991 1198992)

ℎ119892(1198991 1198992) = exp

minus (1198992

1+ 1198992

2)

21205902

(6)


119871as follows

119864119871(119909 119910) =

119866119891119898

119866 (119909 119910) lowast 119906 (119909 119910) oplus 119866119898(119909 119910)

119866119891119898

(119909 119910) oplus 119866119898(119909 119910)

(7)

where

119866119891119898





119871eq (119909 119910) = 119871 ref lowast119871 (119909 119910)

119864119871(119909 119910)

(9)



119871eq (119909 119910) =119871eq (119909 119910) minus 119871eq min



(10)








(CH1015840119871(119894)) lowast 2

119887 (11)




119871 (119909 119910) =1






119874= open (119871

119872))



119872) as the mask


119872)) In





119871119861(119909 119910) =


0 otherwise

forall119909 119910

(13)




119861 Let



and Solidity (119874119894) gt 025

(14)







lowast 100


lowast 100


lowast 100

(15)









(a)

(b)









7 Conclusion







(1)

(2)


Competing Interests


References














































Volume 2014

Zoology






















Advances in

Virolog y



Volume 2014




Enzyme Research



Microbiology


(a) (b)











119868(119901) = mean 119868(119902)




bg (119909 119910) =


0 otherwise(4)




119906119871(119909 119910) =

(119871 (119909 119910) lowast bg (119909 119910)) oplus 119866119898(119909 119910)

bg (119909 119910) oplus 119866119898(119909 119910)

(5)

where

119866119898(1198991 1198992) =

minus (1198992

1+ 1198992

2minus 21205902) ℎ119892(1198991 1198992)

21205871205906sum1198991sum1198992ℎ119892(1198991 1198992)

ℎ119892(1198991 1198992) = exp

minus (1198992

1+ 1198992

2)

21205902

(6)


119871as follows

119864119871(119909 119910) =

119866119891119898

119866 (119909 119910) lowast 119906 (119909 119910) oplus 119866119898(119909 119910)

119866119891119898

(119909 119910) oplus 119866119898(119909 119910)

(7)

where

119866119891119898





119871eq (119909 119910) = 119871 ref lowast119871 (119909 119910)

119864119871(119909 119910)

(9)



119871eq (119909 119910) =119871eq (119909 119910) minus 119871eq min



(10)








(CH1015840119871(119894)) lowast 2

119887 (11)




119871 (119909 119910) =1






119874= open (119871

119872))



119872) as the mask


119872)) In





119871119861(119909 119910) =


0 otherwise

forall119909 119910

(13)




119861 Let



and Solidity (119874119894) gt 025

(14)







lowast 100


lowast 100


lowast 100

(15)









(a)

(b)









7 Conclusion







(1)

(2)


Competing Interests


References














































Volume 2014

Zoology






















Advances in

Virolog y



Volume 2014




Enzyme Research



Microbiology



119906119871(119909 119910) =

(119871 (119909 119910) lowast bg (119909 119910)) oplus 119866119898(119909 119910)

bg (119909 119910) oplus 119866119898(119909 119910)

(5)

where

119866119898(1198991 1198992) =

minus (1198992

1+ 1198992

2minus 21205902) ℎ119892(1198991 1198992)

21205871205906sum1198991sum1198992ℎ119892(1198991 1198992)

ℎ119892(1198991 1198992) = exp

minus (1198992

1+ 1198992

2)

21205902

(6)


119871as follows

119864119871(119909 119910) =

119866119891119898

119866 (119909 119910) lowast 119906 (119909 119910) oplus 119866119898(119909 119910)

119866119891119898

(119909 119910) oplus 119866119898(119909 119910)

(7)

where

119866119891119898





119871eq (119909 119910) = 119871 ref lowast119871 (119909 119910)

119864119871(119909 119910)

(9)



119871eq (119909 119910) =119871eq (119909 119910) minus 119871eq min



(10)








(CH1015840119871(119894)) lowast 2

119887 (11)




119871 (119909 119910) =1






119874= open (119871

119872))



119872) as the mask


119872)) In





119871119861(119909 119910) =


0 otherwise

forall119909 119910

(13)




119861 Let



and Solidity (119874119894) gt 025

(14)







lowast 100


lowast 100


lowast 100

(15)









(a)

(b)









7 Conclusion







(1)

(2)


Competing Interests


References














































Volume 2014

Zoology






















Advances in

Virolog y



Volume 2014




Enzyme Research



Microbiology




119871119861(119909 119910) =


0 otherwise

forall119909 119910

(13)




119861 Let



and Solidity (119874119894) gt 025

(14)







lowast 100


lowast 100


lowast 100

(15)









(a)

(b)









7 Conclusion







(1)

(2)


Competing Interests


References














































Volume 2014

Zoology






















Advances in

Virolog y



Volume 2014




Enzyme Research



Microbiology







lowast 100


lowast 100


lowast 100

(15)









(a)

(b)









7 Conclusion







(1)

(2)


Competing Interests


References














































Volume 2014

Zoology






















Advances in

Virolog y



Volume 2014




Enzyme Research



Microbiology


(a)

(b)









7 Conclusion







(1)

(2)


Competing Interests


References














































Volume 2014

Zoology






















Advances in

Virolog y



Volume 2014




Enzyme Research



Microbiology






(1)

(2)


Competing Interests


References














































Volume 2014

Zoology






















Advances in

Virolog y



Volume 2014




Enzyme Research



Microbiology


(1)

(2)


Competing Interests


References














































Volume 2014

Zoology






















Advances in

Virolog y



Volume 2014




Enzyme Research



Microbiology




























Volume 2014

Zoology






















Advances in

Virolog y



Volume 2014




Enzyme Research



Microbiology








Volume 2014

Zoology






















Advances in

Virolog y



Volume 2014




Enzyme Research



Microbiology

Documents

Research Article Computer Based Melanocytic and Nevus ...downloads.hindawi.com/journals/bmri/2016/2082589.pdfsegmentation technique based on double thresholding for segmentation. Edge-based