Detecting in situ melanoma using multi parameter extraction and neural classification mechanisms
Upcoming SlideShare
Loading in...5

Detecting in situ melanoma using multi parameter extraction and neural classification mechanisms






Total Views
Views on SlideShare
Embed Views



0 Embeds 0

No embeds



Upload Details

Uploaded via as Adobe PDF

Usage Rights

© All Rights Reserved

Report content

Flagged as inappropriate Flag as inappropriate
Flag as inappropriate

Select your reason for flagging this presentation as inappropriate.

  • Full Name Full Name Comment goes here.
    Are you sure you want to
    Your message goes here
Post Comment
Edit your comment

Detecting in situ melanoma using multi parameter extraction and neural classification mechanisms Detecting in situ melanoma using multi parameter extraction and neural classification mechanisms Document Transcript

  • INTERNATIONAL JOURNAL OF COMPUTER(IJCET), ISSN 0976 – International Journal of Computer Engineering and Technology ENGINEERING 6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 1,(IJCET) & TECHNOLOGY January- February (2013), © IAEMEISSN 0976 – 6367(Print)ISSN 0976 – 6375(Online)Volume 4, Issue 1, January- February (2013), pp. 16-33 IJCET© IAEME: Impact Factor (2012): 3.9580 (Calculated by GISI) © DETECTING IN-SITU MELANOMA USING MULTI PARAMETER EXTRACTION AND NEURAL CLASSIFICATION MECHANISMS Ruksar Fatima1,Dr.Mohammed Zafar Ali Khan2,Dr. A. Govardhan3 and Kashyap Dhruve4 1 (Head Dept. of Bio-Medical Engg, KBN College of Engg, Gulbarga, India, 2 (Associate Professor,IIIT Hyderabad, Hyderabad, India,zafar 3 (Director of Evaluation of JNTU, Hyderabad, India, 4 (Technical Director, Planet-i Technologies, Bangalore, India,,) ABSTRACT Computer aided diagnostics systems are widely used for diagnosis of skin lesions. aimed to enable early detection of skin cancer melanoma. The ‫ ܵܥܧܲܯ‬adopts a supervised This paper discusses a Multi Parameter Extraction and Classification System (‫ )ܵܥܧܲܯ‬to machine learning algorithm for classification. The dermoscopic images are represented by operation of the ‫ ܵܥܧܲܯ‬in the training and testing phase. The adoption of the Multilayer extensive parameter sets extracted using a six phase approach. The paper discusses the Feed Forward Neural Network (‫ )ܰܰܨܯ‬classifier is justified, its proficiency in accurately diagnosing and classifying early signs of skin cancer melanoma in dermoscopic images of skin lesions is proved through the experimental results discussed in the manuscript. Keywords: Skin Cancer, Melanoma, Early Detection, In Situ Melanoma, MPECS , Multi Parameter Extraction, Classification, Computer Aided Diagnostics, Image Processing, Skin Lesions. Feature Extraction, Multilayer Feed Forward, Neural Network, Back Propagation, SVM Classifier, ROC, Receiver Operating Characteristics 1. INTRODUCTION Computer vision based diagnosis systems which are non-invasive have experienced a wide acceptability ratio in the recent years. Deaths arising from skin cancers have increased tremendously in the past decade [1] [2][3]. Skin Cancer Melanoma is dangerous and accounts for maximum number of deaths arising due to skin cancers [3][4]. A recent survey [5] conducted in 2012 for the United States of America alone, states that more than 75% of the deaths arising from skin cancers are due to melanotic skin lesions. Skin cancer melanomas 16
  • International Journal of Computer Engineering and Technology (IJCET), ISSN 0976 –6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 1, January- February (2013), © IAEMEare generally diagnosed by adopting biopsy procedures recommended by dermatologists.Studies have proved that the mortality rate could be reduced if melanotic lesions are detected The research work presented here puts forth the ‫ ܵܥܧܲܯ‬to aid early detection of inat a very early stage [6].situ melanoma and classification of skin lesions. The dermoscopic images of skin lesions arerepresented as a set of 21 features or parameter extracted phase wise. The skin lesionsrepresented as a set of features need to be classified. Statistical analysis carried out for on theparameter set obtained post extraction were inconclusive in classifying the skin lesions [7]into the ‫ .ܵܥܧܲܯ‬Machine learning algorithms are adopted for accurate classification. Ahence there was an requirement for advanced classification mechanisms to be incorporated‫ ܰܰܨܯ‬classifier trained using the back propagation algorithm is adopted. The training phaseof the ‫ ܵܥܧܲܯ‬enables the system to understand the features that exhibit properties of specificclasses defined during training. Subsequently the ‫ ܵܥܧܲܯ‬can be utilized for classification ofthe testing data set. The proposed methodology is compared with the popular ܸܵ‫ ܯ‬classifier. This research manuscript is organized as follows. Section 2 discusses themethodologies adopted by researchers. The next section discusses the proposed ‫ ܵܥܧܲܯ‬atliteraturesurvey about skin cancer melanoma diagnosis systems and classificationlength along with the ‫ ܰܰܨܯ‬classifier. The experimental evaluation and comparisons arediscussed in the penultimate section of this paper. The conclusion of the research workpresented in this paper is discussed in the fifth section.2. LITERATURE SURVEY The Asymmetry Border Color and Differential Structure (ABCD) methodology is themost commonly used in the diagnosis of skin cancer melanoma [8]. To enhance theefficiency additional optimizations such as fuzzy border extraction [9], “JSEG” basedoptimization [10]; hybrid techniques [11] have been incorporated. A classification accuracyof about 60% was achieved by incorporating wavelet decomposition techniques along withthe ABCD principle. Skin lesion classification can also be achieved using the MenziesMethod [12]. The ABCD based methods and the Menzies Method exhibit lesser accuracy indetection of early stages of skin cancer melanoma. G. Betta at el [13] introduced a sevenpoint checklist method to diagnose skin lesions which proved to exhibit higher classificationaccuracy than the other methodologies described. The computational complexity of the sevenpoint checklist method is considered as a major drawback. The, methods discussed aboveexhibited less of low classification accuracy. The need of machine learning algorithms toaccurately classify and diagnose early symptoms of skin cancer melanoma arose. The use ofadvance clustering techniques [14], k Nearest Neighborhood [6], classification/regressiontrees [15] has been closely studied. The use of the support vector machines (ܸܵ‫ )ܯ‬forof skin cancer melanoma [18]. The ‫ ܵܥܧܲܯ‬adopts the ‫ ܰܰܨܯ‬based classifier [23]. Theclassification [16] [17] has been proposed by most of the researchers even for early detectionabove mentioned methodologies are inaccurate to detect and classify early stages of skincancer melanoma. A few systems proposed to detect early symptoms exhibit lowclassification accuracy as the early stages of skin cancer melanoma described throughfeatures belong to the non-melanoma and melanoma classes cumulatively makingclassification a very difficult or cumbersome task. 17
  • International Journal of Computer Engineering and Technology (IJCET), ISSN 0976 –6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 1, January- February (2013), © IAEME3. MULTI PARAMETER EXTRACTION AND CLASSIFICATION SYSTEM(ࡹࡼࡱ࡯ࡿ) MODELING Appropriate skin lesions diagnosis at early stages enable possible cure to melanomaskin cancers. Skin lesions of the melanotic type in the in situ stages are easily mistaken withother skin related ailments by even experienced physicians. This paper introduces a computeraided skin lesion diagnosis system named Multi Parameter Extraction and Classificationaccounts for the highest number of deaths amongst skin cancers. The ‫ ܵܥܧܲܯ‬relies on theSystem (‫ )ܵܥܧܲܯ‬aiding early detection of skin cancer melanoma. Skin cancer melanomamachine learning techniques incorporated to accurately classify the skin lesions defined by aset of parameters extracted. 3.1. ࡹࡼࡱ࡯ࡿ – PRELIMINARY NOTATION Let‫ܫ‬ௌ௄ represent a set of dermoscopic skin lesion images for analysis defined as‫ܫ‬ௌ௄ ൌ ሼ݅ଵ , ݅ଶ , ݅ଷ , ݅ସ , … … ݅௡ ሽ݅ represents an imageWhere݊represents the total number of images to be analyzed. The MPECS system proposed in this paper is aids the early diagnosis of skin cancermelanoma skin lesions. Let the classification set of the skin lesions be defined asC ൌ ሼcଵ , cଶ , cଷ … cୡ୪ ሽWhere c represents the class and cl represents the total number of classes considered.Based on the above mentioned definitions it could be stated that ‫ ׊‬i ‫ א‬I ‫ !׌ ׷‬c ‫ א‬C | i ‫ א‬c. The ‫ ܵܥܧܲܯ‬aids the classification of skin lesions from dermoscopic images. The skinlesions are defined as a set of parameter P that it exhibits. Let the Parameter set be defined byP ൌ ሼpଵ , pଶ , pଷ , … pୟ ሽWhere a represents the total number of p parameters In the MPECS an image i ‫ א‬I is represented by a set of parameters P is used forclassification. The image I୬ could now defined asI୬ ൌ ൛p୬ଵ , p୬ଶ , p୬ଷ , … p୬୮ ൟAn image I୬ could be represented as a matrix and is defined as 18
  • International Journal of Computer Engineering and Technology (IJCET), ISSN 0976 –6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 1, January- February (2013), © IAEME x଴,଴ ‫ ڮ‬x଴,୛ୢ I୬ ൌ ൥ ‫ڭ‬ ‫ڰ‬ ‫ ڭ‬൩ xୌ୲,଴ ‫ ڮ‬xୌ୲,୛ୢWhere Ht represents the height and Wd represents the width of the image I୬ and x representsthe pixel value at the given location. The MPECS adopts a 6 phase approach to extract the parameters set P ‫ ׊‬i ‫ א‬Iୗ୏ . Thesystem considers 21 parameters for classification i.e. a ൌ 21. Each phase provides towardsthe construction of the parameter set P.The ‫ ܵܥܧܲܯ‬relies on the supervised machine learningwith back propagation [24] training is adopted for classification. The trainings set ܶோ isalgorithm for classification. The use of multi-layer feed forward neural networks (‫)ܰܰܨܯ‬defined asܶோ ൌ ሼ்ܴோଵ , ்ܴோଶ , ்ܴோଷ , … … . ்ܴோ௥ ሽWhere ்ܴோ௥ represents the ‫ ݎ‬௧௛ training record.c from the training image. I.e. ்ܴோ௥ ൌ ሼI୰ ‫ ׫‬c௥ ሽwhere I୰ represents the parameter set that The training record is constructed from the parameter set and the corresponding classdefines the ‫ ݎ‬௧௛ training image and c௥ represents the corresponding classification class.Similarly the test set ்ܶ is defined as்ܶ ൌ ሼ்்ܴଵ , ்்ܴଶ , ்ܴோଷ , … … . ்்ܴ௧ ሽThe objective of the ‫ ܵܥܧܲܯ‬is to identify c௧ | ்்ܴ௧ ‫ א‬c௧ ܽ݊݀ c௧ ‫ܥ א‬ 3.2 OPERATION OF THE ࡹࡼࡱ࡯ࡿ The ‫ ܵܥܧܲܯ‬relies on supervised learning techniques for classification. To impartintelligence the ‫ ܵܥܧܲܯ‬is trained using the pre-classified dermoscopic images. The trainingoperation of the ‫ ܵܥܧܲܯ‬is described by the algorithm given below where‫்ܰܰܨܯ‬ோ ሺܺሻrepresents the back propagation training function based on the inputs set ܺ. Theparameters extracted from the phase ‫ ݌‬is represented by ‫்݌‬ோ௜ .Algorithm Name:Training Phase of the ‫ܵܥܧܲܯ‬ 1. ‫ݎ‬training dermoscopic imagesInput: 2. classification set C ൌ ሼcଵ , cଶ , cଷ … c୰ ሽ 1. Trained ‫ ܰܰܨܯ‬classifier.Output: 19
  • International Journal of Computer Engineering and Technology (IJCET), ISSN 0976 –6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 1, January- February (2013), © IAEMEAlgorithmStep 2: ܶோ ൌ ‫׎‬Step 1: Training Phase BeginStep 3: For all training images ݅ ‫ݎ ݋ݐ 1 ׷‬ ்ܴோ௜ ൌ ‫׎‬Step 5: ࡹࡼࡱ࡯ࡿ Parameter Extraction BeginStep 4: Parameter Set P்ோ௜ ൌ ‫׎‬ For phases ࢖ ‫ ׷‬૚: ૟Step 6: P்ோ௜ ൌ P்ோ௜ ‫்݌ ׫‬ோ௜Step 7:Step 8:Step 9: End phaseStep 11:்ܴோ௜ ൌ P்ோ௜ ‫ ׫‬c௜Step 10: Parameter Extraction EndStep 12: ܶோ ൌ ܶோ ‫்ܴ ׫‬ோ௜Step 14: Perform Training ‫்ܰܰܨܯ‬ோ ሺܶோ ሻStep 13: End ForStep 15: End Training Phase The ‫ ܰܰܨܯ‬post training could be used for classification of the skin lesions represented in thetesting dermoscopic image dataset. The testing phase of the ‫ ܵܥܧܲܯ‬is described in the algorithmgiven below where‫்்݌‬௜ represents the parameters extracted in the ‫݌‬௧௛ extraction phaseand‫ ்்ܰܰܨܯ‬ሺܻሻis the classification function of ܻ defined as‫ ்்ܰܰܨܯ‬ሺܻሻ ൌ ܿ௧Where ܿ௧ ‫ܥ א‬Algorithm Name: Testing Phase of the ‫ܵܥܧܲܯ‬ 1. ‫ݐ‬testing dermoscopic imagesInput: 2. Trained ‫்ܰܰܨܯ‬ோ ሺܶோ ሻ 1. Classification Output ܿ௧ ‫ܥ א‬Output:AlgorithmStep 2:்ܶ ൌ ‫׎‬Step 1: Testing Phase BeginStep 3: For all testing images ݅ ‫ݐ ݋ݐ 1 ׷‬ ்்ܴ௜ ൌ ‫׎‬Step 5: ࡹࡼࡱ࡯ࡿ Parameter Extraction BeginStep 4: Parameter Set P்்௜ ൌ ‫׎‬ For phases ࢖ ‫ ׷‬૚: ૟Step 6: P்்௜ ൌ P்்௜ ‫்்݌ ׫‬௜Step 7:Step 8:Step 9: End phaseStep 11:்்ܴ௜ ൌ P்்௜Step 10: Parameter Extraction EndStep 12: ்ܶ ൌ ்ܶ ‫்்ܴ ׫‬௜Step 14: Perform Classification Query‫ ்்ܰܰܨܯ‬ሺ்ܶ ሻ ൌ ܿ௧Step 13: End ForStep 15: End Testing Phase 20
  • International Journal of Computer Engineering and Technology (IJCET), ISSN 0976 –6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 1, January- February (2013), © IAEME The training and testing phases of the ‫ ܵܥܧܲܯ‬have been discussed above whichexplains the operation. Parameter extraction is an integral part of the ‫ ܵܥܧܲܯ‬which defines‫ ܵܥܧܲܯ‬is acheied using a 6 phase approach described in the subsequent section of this paper.the skin lesions and aids in training and classification. The 21 parameter extraction of the 3.3. PARAMETER EXTRACTION IN THE ࡹࡼࡱ࡯ࡿ The dermoscopic images of skin lesions is represented as a set of 21 parameters extractedin 6 phases described below [7] 3.3.1. ‫܁۱۳۾ۻ‬PARAMETER EXTRACTION – PHASE 1present in an image ݅ ‫ܫ א‬and extracting the Lesion Borders, the symmetry of the lesions and This preliminary phase is primarily designed towards identifying the skin lesionsthe color spreading factor of the skin lesions. To extract the parameters discussed the images are resized using a bicubicpixels in the nearest ሾ 4 ൈ 4 ሿproximity and is defined asinterpolation technique. The bicubic interpolation technique is an weighted average of the ଷ ଷPୖःऊ ሺx, yሻ ൌ ෍ ෍൫wt ୧୨ ൈ x୧ y ୨ ൯ ୧౤ ୧ୀଵ ୨ୀଵWhere ‫ݕ , ݔ‬represent the pixel position and wt ୧୨is the weight at position ݅, ݆ of the imagei୬ ‫ א‬Iୗ୏ Grey scaling is performed on the resized pixels, and the resultant grey scale pixel iscomputed based on the following definition. Pୋ୰ୣ୷ୗୡୟ୪ୣ ୧౤ ሺx, yሻ ൌ ቂ0.2989 ൈ R ቀPୖःऊ ୧౤ ሺx, yሻቁቃ ൅ ቂ0.5870 ൈ G ቀPୖःऊ ୧౤ ሺx, yሻቁቃ ൅ ቂ0.11400 ൈ B ቀPୖःऊ ୧౤ ሺx, yሻቁቃWhere R ቀPୖःऊ ୧౤ ሺx, yሻቁrepresented the red channel value, G ቀPୖःऊ ୧౤ ሺx, yሻቁ represents thegreen channel value and B ቀPୖःऊ ୧౤ ሺx, yሻቁrepresents the blue channel value of the resizedpixelPୖःऊ ୧౤ ሺx, yሻ. Binirazation is performed on the grayscale image utilizing the adaptive thresholdingalgorithm. The image noise elimination property of the adaptive thresholding basedbinirazation algorithm is an additional parameter for adopting this algorithm. The adaptivethresholding based binirazation is an iterative process wherein the thresholds are adaptedconvergence is achieved. Let P୆୧୬ ୧౤ ሺx, yሻ represent the pixels of the binirized image i୬ . Thebased on the regions it is being performed on. The iterative process is terminated whenregions of interest ሺ‫݅݋ݎ‬ሻ is extracted using the connected component labeling algorithm. The‫ ݏ’݅݋ݎ‬obtained hold the information required and are identified by labels Lୖ୓୍ି୪ ୧౤ . The image 21
  • International Journal of Computer Engineering and Technology (IJCET), ISSN 0976 –6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 1, January- February (2013), © IAEMEi୬ ‫ א‬Iୗ୏ based on the ܴܱ‫ ݏ’ܫ‬is defined as i୬ ି୰୭୧ ൌ ሼL୰୭୧ିଵ ୧౤ ‫ ׫‬L୰୭୧ିଶ ୧౤ ‫ ׫‬L୰୭୧ିଷ ୧౤ ‫ ׫ ڮ ׫‬L୰୭୧ି୪ ୧౤ ሽWhere ݈represents the total number of ‫ ݏ’݅݋ݎ‬extracted. The centroids of the ‫ ݏ’݅݋ݎ‬are computed based on the area enclosed by the ‫ .ݏ’݅݋ݎ‬Thecentroids are computed to cluster the similar roi’s together in the same vicinity as it isobserved that the roi’s are similar in nature and can be clustered not effecting theclassification accuracy. This operation adopted enables roi’s exhibiting similar properties tobe clustered into a single cluster together represented asܴܱ‫ .ܫ‬The labels L୰୭୧ି୪ ୧౤ used toaddress the roi’s are clustered based on the energy computed and is defined asࣟgyୖ୓୍ି୪ ൫Lୖ୓୍ି୪ ୧౤ ൯ ൌ ෍ ቎ ෍ ∆൫L୰୭୧ିୟ ୧౤ , L୰୭୧ିୠ ୧౤ ൯቏ ୮୭ୱ ୪ ‫א‬ୗ౨౥౟Where ∆൫L୰୭୧ିୟ ୧౤ , L୰୭୧ିୠ ୧౤ ൯represents a function ∆൫L୰୭୧ିୟ ୧౤ , L୰୭୧ିୠ ୧౤ ൯ ൌ ሼ0,1ሽ i.e. if theroi’s L୰୭୧ିୟ ୧౤ and L୰୭୧ିୠ ୧౤ are similar, 0 is returned and 1 in the case of dissimilarity. S୰୭୧ ൌሼ1,2,3, … . lሽrepresents the set of roi’s obtained and a ‫ א‬S୰୭୧and b ‫ א‬S୰୭୧. pos represents theposition of the l୲୦ clustered ROI represented as Lୖ୓୍ି୪ ୧౤ .The clustered ROI’s obtained define the image i୬ asi୬ ିୖ୓୍ ൌ ሼLୖ୓୍ିଵ ୧౤ ‫ ׫‬Lୖ୓୍ିଶ ୧౤ ‫ ׫‬Lୖ୓୍ିଷ ୧౤ ‫ ׫ ڮ ׫‬Lୖ୓୍ି୪ ୧౤ ሽ Sobel edge detection is performed on the image i୬ ୖ୓୍. Post the edge detection thedistortion on the basis of the coordinate displacement and the frequency of distortionobserved is computed using the Fast Fourier Transforms. The axis based asymmetryparameter of the lesion is obtained based on the principal component decomposition. The skinlesion extracted is rotated such that the principal component coincides with the horizontalp୅୶୧ୱ_ୗ୷୫୫ୣ୲୰୷ ൌ ቂ෍|i୬ ሺxሻ െ ıഥ ሺxሻ|ቃൗቂ෍ i୬ ሺxሻቃaxis. The skin lesion axis based asymmetry parameter is defined as ୬Where i୬ ሺxሻthe original skin lesion and its reflected version is represented as ıഥ ሺxሻ. ୬ The symmetry parameter is computed for all the pixels within the lesion along thebased on the similarity of a pixel in a n ൈ n neighbourhood. The standard deviation of thehorizontal and vertical axis respectively. The color spreading factor of the lesions is computedpixels in the neighborhood based on the Red channel, Green channel and Blue channel iscomputed using 1 ୬మ തതതതതത ൌ ඩ ଶ ෍ሺChnl୧ െ Chnlሻଶ nσେ୦୬୪ ୧ୀଵ തതതതതതWhere Chnl ൌ ሼR େ୦୬୪ , Gେ୦୬୪ , Bେ୦୬୪ ሽ and Chnlrepresents the mean value defined as 22
  • International Journal of Computer Engineering and Technology (IJCET), ISSN 0976 –6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 1, January- February (2013), © IAEME 1 ୬మതതതതതChn ൌ ଶ ෍ Chn୧ n ୧ୀଵσେ୳୫୧୪୧୲୧୴ୣ ൌ σୖి౞౤ౢ ൅ σୋి౞౤ౢ ൅ σ୆ి౞౤ౢThe cumulative standard deviation for all the channels is defined as ‫ۍ‬ ‫ۍ ې‬ ‫ې‬ 1 1 ୬మ ୬మ ൌ ‫ێ‬ඩ ଶ ෍ሺR େ୦୬୪ ୧ െ R େ୦୬୪ ሻଶ ‫ ۑ‬൅ ‫ێ‬ඩ ଶ ෍ሺGେ୦୬୪ ୧ െ Gେ୦୬୪ ሻଶ ‫ۑ‬ തതതതതതത തതതതതതത ‫ ێ‬n ୧ୀଵ ‫ ێ ۑ‬n ୧ୀଵ ‫ۑ‬σେ୳୫୧୪୧୲୧୴ୣ ‫ۏ‬ ‫ۏ ے‬ ‫ے‬ ‫ۍ‬ ‫ې‬ 1 ୬మ ൅ ‫ێ‬ඩ ෍ሺBେ୦୬୪ െ Bେ୦୬୪ ሻଶ ‫ۑ‬ തതതതതതത ‫ ێ‬nଶ ୧ୀଵ ୧ ‫ۑ‬ ‫ۏ‬ ‫ے‬The color spreading factor of the lesion is computed usingpେ୪୰ିୗ୮୰ୣୟୢ ൌ 1 െ σ୒୭୰୫Where σ୒୭୰୫is the normalized value of σେ୳୫୧୪୧୲୧୴ୣ between the range 0 and 1 and is obtainedσ୒୭୰୫ ൌ σେ୳୫୧୪୧୲୧୴ୣ ⁄σେ୳୫୧୪୧୲୧୴ୣି୑ୟ୶as In order to extract the lesion border parameters the image is filtered using theGaussian and laplacian filter. The parameter describing the lesion borders is obtained bycomputing the mean of the resultant pixels and is defined as ୛ୢ ୌ୲p୐ୣୱ୧୭୬ି୆୭୳୬ୢ୰୷ ൌ ቎෍ ෍ P୧౤ ሺx, yሻ቏൘ሾWd ൈ Htሿ ୶ୀଵ ୷ୀଵ In this phase the parameters of the lesions such as the asymmetry along the horizontaland vertical axis, the color spreading factor of the lesion and the boundary parameters isdiscussed. 3.3.2. ‫ ܁۱۳۾ۻ‬PARAMETER EXTRACTION – PHASE 2 The second phase of the ‫ ܵܥܧܲܯ‬enables the extraction of the area, perimeter and theeccentricity. This phase considers the binirized and clustered ܴܱ‫ ܫ‬image as an input and݅௡ ିோைூ ൌ ሼ‫ܮ‬ோைூିଵ ௜೙ ‫ܮ ׫‬ோைூିଶ ௜೙ ‫ܮ ׫‬ோைூିଷ ௜೙ ‫ܮ ׫ ڮ ׫‬ோைூି௟ ௜೙ ሽcomputes the parameters based on the binirized image obtained from phase 1 defined asWhere ݈ represents the total number of ܴܱ‫ ݏ’ܫ‬identified for the image ݅௡ . The ܴܱ‫ ݏ’ܫ‬of image ݅௡ ିோைூ contain binirized pixel points i.e. black or white pixels.The area parameter defines the white pixels that are encapsulated by the skin lesion and thelesion area parameter is computed using 23
  • International Journal of Computer Engineering and Technology (IJCET), ISSN 0976 –6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 1, January- February (2013), © IAEME ௉௡௧௦ିଵ‫݌‬௅௘௦௜௢௡ି஺௥௘௔ ൌ ሺ1⁄2ሻ ቎ | ෍ ൫‫ݔ‬௜ ‫ݕ‬௝ାଵ െ ‫ݔ‬௜ାଵ ‫ݕ‬௝ ൯ | ቏ ௜,௝ୀ଴Where ܲ݊‫ ݏݐ‬is the number of points enclosed by the skin lesion defined by the ݈ ௧௛ ܴܱ‫ . ܫ‬Also݅, ݆ represent the ܲ݊‫ݏݐ‬location and ൫‫ݔ‬௜ , ‫ݕ‬௝ ൯ ‫ א‬൛൫‫ݔ‬௜ , ‫ݕ‬௝ ൯ห ݂൫‫ݔ‬௜ , ‫ݕ‬௝ ൯ ൌ 1 ሽ as ‫ݔ‬௜ , ‫ݕ‬௝ represents a The perimeter defines the boundary ߂‫ ܤ‬length of the skin lesion. The lesion boundarywhite pixel.߂‫ܤ‬ሺ݈ሻ ൌ ሺ‫ܤ‬ሺ2: ݈, : ሻ െ ‫ܤ‬ሺ1: ݈ െ 1, : ሻሻଶis computed usingWhere ݈ represents the number of ܴܱ‫ ݏ’ܫ‬and its corresponding boundary is represented as ‫.ܤ‬The parameter defining boundary of the skin lesion is defined as‫݌‬௅௘௦௜௢௡ି஻௢௨௡ௗ ൌ ෍ ඨ෍ ߂‫ܤ‬ሺ݈ሻ ௟eccentricity is obtained by obtaining the moments represented as ‫ .ܯ‬The major axis is The aspect ratio of the skin lesion is defined as the eccentricity of a skin lesion. The‫ܣ‬ெ௔௝௢௥ ൌ 2√2ሺඥ‫ܯ‬௫௫ ൅ ‫ܯ‬௬௬ ൅ ‫ܯ‬஼௢௠௠ ሻdefined as‫ܣ‬ெ௜௡௢௥ ൌ 2√2ሺඥ‫ܯ‬௫௫ ൅ ‫ܯ‬௬௬ െ ‫ܯ‬஼௢௠௠ ሻThe minor axis is defined asWhere the moments ‫ܯ‬௫௫ , ‫ܯ‬௬௬ , ‫ܯ‬௫௬ and‫ܯ‬஼௢௠௠ are defined as‫ܯ‬௫௫ ൌ ቂቀ෍ ‫ ݔ‬ଶ ቁൗܰ ൅ 1⁄12ቃ‫ܯ‬௬௬ ൌ ቂቀ෍ ‫ ݕ‬ଶ ቁൗܰ ൅ 1⁄12ቃ‫ܯ‬௫௬ ൌ ቂቀ෍ ‫ݕ כ ݔ‬ቁൗܰቃAnd‫ܯ‬஼௢௠௠ ൌ ටሺ‫ܯ‬௫௫ െ ‫ܯ‬௬௬ ሻଶ ൅ 4‫ܯ‬௫௬ ଶThe eccentricity parameter is computed using‫݌‬௅௘௦௜௢௡ିா௖௖௘௡௧ ൌ ቈට൫‫ܣ‬ெ௔௝௢௥ ଶ െ ‫ܣ‬ெ௜௡௢௥ ଶ ൯቉ൗ‫ܣ‬ெ௔௝௢௥ మ 3.3.3. ‫ ܁۱۳۾ۻ‬PARAMETER EXTRACTION – PHASE 3 Researches have well understood the importance to 3‫ ܦ‬depth parameters to enhance theclassification accuracy [6] [19]. To obtain the 3‫ ܦ‬depth of the skin lesions the‫ ܵܥܧܲܯ‬introduced in this paper considers creating a 3 ൈ 3 masked windows represented as ܹଵ ܹଶ ܹଷ݅௡ ଷൈଷିெ௔௦௞ ൌ ൥ܹସ ܹହ ܹ଺ ൩ ܹ଻ ଼ܹ ܹଽWhere ܹ௜ represent the weights of the pixelThe projection filter is defined as 24
  • International Journal of Computer Engineering and Technology (IJCET), ISSN 0976 –6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 1, January- February (2013), © IAEME ଽ݅௡ ଷ஽ି஽௘௣௧௛ ൌ ෍ ܹ௜ ‫ܫ‬௜ ௜ୀଵWhere ‫ܫ‬௜ represents the intensity of the ݅ ௧௛ pixel.The 3D depth projection parameter ‫݌‬௅௘௦௜௢௡ିଷ஽஽௘௣௧ is obtained by considering the geometric‫݌‬௅௘௦௜௢௡ିଷ஽஽௘௣௧mean defined as ൌ ሾ1⁄9ሿ ൈ ൣ‫ܫ‬௫ିଵ,௬ିଵ ൅ ‫ܫ‬௫ିଵ,௬ଵ ൅ ‫ܫ‬௫ିଵ,௬ାଵ ൅ ‫ܫ‬௫,௬ିଵ ൅ ‫ܫ‬௫,௬ ൅ ‫ܫ‬௫,௬ାଵ ൅ ‫ܫ‬௫ାଵ,௬ିଵ ൅ ‫ܫ‬௫ାଵ,௬ଵ ൅ ‫ܫ‬௫ାଵ,௬ାଵ ሿWhere ‫ ݕ ,ݔ‬represent the horizontal and vertical pixel position. 3.3.4. ‫ ܁۱۳۾ۻ‬PARAMETER EXTRACTION – PHASE 4 The fourth phase of the ‫ܵܥܧܲܯ‬is targeted towards obtaining the color components of theskin lesions. The color components are extracted for the red channel blue channel and thegreen channel. The mean and the variance parameters of each channel are considered as theparameters to be utilized for classification. For the red channel the mean parameter is defined ∑ௐௗ ∑ு௧ ܴ஼௛௡௟ ሺ‫ݕ ,ݔ‬ሻas follows ‫݌‬ோ஼௛௡௟ିெ௘௔௡ ൌ ௫ୀଵ ௬ୀଵ ܹ݀ ‫ݐܪ כ‬Where ‫ ݕ ,ݔ‬represent the pixel positions. The variance parameter is computed by obtaining the mean of all the ‫ܤ‬஼௛௡௟ columns , thedifference amongst them and then summed square difference divided by the height ‫ݐܪ‬ ∑ௐௗ ܴሺ݅ሻ‫݊ܽ݁݉_݈݉ܥ‬ோ஼௛௡௟ ሺ݅ሻ ൌ ௜ୀଵ ܹ݀‫݂݂݅ܦ‬ோ஼௛௡௟ ሺ݅ ሻ ൌ ܴ‫݄݊ܥ‬ሺ: , ݅ ሻ െ ‫݊ܽ݁݉_݈݉ܥ‬ோ஼௛௡௟ ሺ݅ሻ ∑௜ ‫݂݂݅ܦ‬ோ஼௛௡௟ ሺ݅ሻ‫݌‬ோ஼௛௡௟ି௏௔௥ ൌ ሺ‫ ݐܪ‬െ 1ሻ ∑ௐௗ ∑ு௧ ‫ܩ‬஼௛௡௟ ሺ‫ݕ ,ݔ‬ሻSimilarly the green channel and blue channel parameters are defined as‫ீ݌‬಴೓೙೗௟ିெ௘௔௡ ൌ ௫ୀଵ ௬ୀଵ ܹ݀ ‫ݐܪ כ‬ ∑௜ ‫ீ݂݂݅ܦ‬಴೓೙೗ ሺ݅ ሻ‫ீ݌‬಴೓೙೗ି௏௔௥ ൌ ሺ‫ ݐܪ‬െ 1ሻ ∑ௐௗ ∑ு௧ ‫ܤ‬஼௛௡௟ ሺ‫ݕ ,ݔ‬ሻ‫݌‬஻಴೓೙೗ିெ௘௔௡ ൌ ௫ୀଵ ௬ୀଵ ܹ݀ ‫ݐܪ כ‬ ∑௜ ‫݂݂݅ܦ‬஻಴೓೙೗ ሺ݅ሻ‫݌‬஻಴೓೙೗ି௏௔௥ ൌ ሺ‫ ݐܪ‬െ 1ሻ 25
  • International Journal of Computer Engineering and Technology (IJCET), ISSN 0976 –6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 1, January- February (2013), © IAEME 3.3.5. ‫ ܁۱۳۾ۻ‬PARAMETER EXTRACTION – PHASE 5 This phase of the ‫ ܵܥܧܲܯ‬discusses the smoothening process of the ܴ‫ ݈݄݊ܥ‬and‫ܩ‬஼௛௡௟ using the 3 ൈ 3 masking procedure discussed in the third phase. The smoothing is notadopted for the ‫ܤ‬஼௛௡௟ as the blue veils of the skin lesions is an important parameter for earlydetection of skin cancer melanoma [20]. The smoothening filter for the ܴ‫ ݈݄݊ܥ‬and ‫ܩ‬஼௛௡௟ is ଽdefined as follows݅௡ ோ஼௛௡௟ିௌ௠௢௢௧௛ ൌ ෍ ܹ௜ ܴ‫݈݄݊ܥ‬௜ ௜ୀଵ ଽ݅௡ ீ ൌ ෍ ܹ௜ ‫ܩ‬஼௛௡௟ ௜ ಴೓೙೗ ିௌ௠௢௢௧௛ ௜ୀଵWhere ܹ௜ represent the weights of the pixel and ܴ‫݈݄݊ܥ‬௜ , ‫ܩ‬஼௛௡௟ ௜ is red channel intensity andgreen channel intensity of the ݅௧௛ pixel.The red channel smoothened parameter ‫݌‬ோ஼௛௡௟ିௌ௠௢௢௧௛ is defined as‫݌‬ோ஼௛௡௟ିௌ௠௢௢௧௛ ൌ ሾ1⁄9ሿ ൈ ൣܴ‫݈݄݊ܥ‬௫ିଵ,௬ିଵ ൅ ܴ‫݈݄݊ܥ‬௫ିଵ,௬ଵ ൅ ܴ‫݈݄݊ܥ‬௫ିଵ,௬ାଵ ൅ ܴ‫݈݄݊ܥ‬௫,௬ିଵ ൅ ܴ‫݈݄݊ܥ‬௫,௬ ൅ ܴ‫݈݄݊ܥ‬௫,௬ାଵ ൅ ܴ‫݈݄݊ܥ‬௫ାଵ,௬ିଵ ൅ ܴ‫݈݄݊ܥ‬௫ାଵ,௬ଵ ൅ ܴ‫݈݄݊ܥ‬௫ାଵ,௬ାଵ ሿThe green channel smoothened parameter ‫ீ݌‬಴೓೙೗ିௌ௠௢௢௧௛‫ீ݌‬಴೓೙೗ିௌ௠௢௢௧௛ ൌ ሾ1⁄9ሿ ൈ ቂ‫ܩ‬஼௛௡௟ ௫ିଵ,௬ିଵ ൅ ‫ܩ‬஼௛௡௟ ௫ିଵ,௬ଵ ൅ ‫ܩ‬஼௛௡௟ ௫ିଵ,௬ାଵ ൅ ‫ܩ‬஼௛௡௟ ௫,௬ିଵ ൅ ‫ܩ‬஼௛௡௟ ௫,௬ ൅ ‫ܩ‬஼௛௡௟ ௫,௬ାଵ ൅ ‫ܩ‬஼௛௡௟ ௫ାଵ,௬ିଵ ൅ ‫ܩ‬஼௛௡௟ ௫ାଵ,௬ଵ ൅ ‫ܩ‬஼௛௡௟ ௫ାଵ,௬ାଵ ൧ 3.3.6. ‫ ܁۱۳۾ۻ‬PARAMETER EXTRACTION – PHASE 6components of the skin lesion to obtain accurate classification results [21][22]. The ‫ܵܥܧܲܯ‬ For early detection of skin cancer melanoma it is essential to extract all the colordiscussed considers the cylindrical coordinate representation of pixels in the fourth and fifth‫ ܵܥܧܲܯ‬considers the Cartesian representation of pixels. Hue, Saturation and Valuephase. In order to extract accurate and elaborate color parameters this phase of therepresentation of the pixels are considered as the Cartesian representations. The mean andfrom the ܴ஼௛௡௟ , ‫ܩ‬஼௛௡௟ ܽ݊݀‫ܤ‬஼௛௡௟ pixel values of the skin lesions. The hue value of a pixel isvariance parameters of the hue channel, variance channel and the value channel are extracted 43 ‫ܩ| כ‬஼௛௡௟ െ ‫ܤ‬஼௛௡௟ |computed using the following definition ‫ۓ‬ቈ0 ൅ ቉ , ‫ ݈ܸܽݔܽܯ‬ൌ ܴ‫݈݄݊ܥ‬ ۖ ‫ ݈ܸܽݔܽܯ‬െ ‫݈ܸܽ݊݅ܯ‬ ۖ 43 ‫ܤ| כ‬஼௛௡௟ െ ܴ‫|݈݄݊ܥ‬‫݁ݑܪ‬௜,௝ ൌ ቈ85 ൅ ቉ , ‫ ݈ܸܽݔܽܯ‬ൌ ‫ܩ‬஼௛௡௟ ‫۔‬ ‫ ݈ܸܽݔܽܯ‬െ ‫݈ܸܽ݊݅ܯ‬ ۖቈ171 ൅ 43 ‫ ݈݄݊ܥܴ| כ‬െ ‫ܩ‬஼௛௡௟ |቉ ۖ , ‫ ݈ܸܽݔܽܯ‬ൌ ‫ܤ‬஼௛௡௟ ‫ە‬ ‫ ݈ܸܽݔܽܯ‬െ ‫݈ܸܽ݊݅ܯ‬Where ‫ ݈ܸܽݔܽܯ‬ൌ ‫ݔܽܯ‬ሺܴ‫ܩ ,݈݄݊ܥ‬஼௛௡௟ , ‫ܤ‬஼௛௡௟ ሻ is the maximum value of the red green andblue channel of the pixel 26
  • International Journal of Computer Engineering and Technology (IJCET), ISSN 0976 –6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 1, January- February (2013), © IAEME‫ ݈ܸܽ݊݅ܯ‬ൌ ‫݊݅ܯ‬ሺܴ‫ܩ ,݈݄݊ܥ‬஼௛௡௟ , ‫ܤ‬஼௛௡௟ ሻis the minimum value of the red green and blue channelof the pixel ሼ‫ ݈ܸܽݔܽܯ‬െ ‫݈ܸܽ݊݅ܯ‬ሽThe Saturation and the value parameter of the pixel is computed usingܵܽ‫݊݋݅ݐܽݎݑݐ‬௜,௝ ൌ 255 ‫כ‬ ‫݈ܸܽݔܽܯ‬ܸ݈ܽ‫݁ݑ‬௜,௝ ൌ ‫݈ܸܽݔܽܯ‬ ∑ௐௗ ∑ு௧ ‫݁ݑܪ‬஼௛௡௟ ሺ‫ݕ ,ݔ‬ሻThe mean of the hue channel and the is defined as ‫݌‬ு௨௘಴೓೙೗ ௟ିெ௘௔௡ ൌ ௫ୀଵ ௬ୀଵ ܹ݀ ‫ݐܪ כ‬The variance is computed in a manner similar to the procedure described in phase 4 and is ∑௜ ‫݂݂݅ܦ‬ு௨௘಴೓೙೗ ሺ݅ ሻdefined as ‫݌‬ு௨௘಴೓೙೗ ି௏௔௥ ൌ ሺ‫ ݐܪ‬െ 1ሻ ∑ௐௗ ∑ு௧ ܵܽ‫݊݋݅ݐܽݎݑݐ‬஼௛௡௟ ሺ‫ݕ ,ݔ‬ሻThe mean and the variance of the saturation channel is defined as ‫݌‬ௌ௔௧௨௥௔௧௜௢௡಴೓೙೗ ିெ௘௔௡ ൌ ௫ୀଵ ௬ୀଵ ܹ݀ ‫ݐܪ כ‬ ∑௜ ‫݂݂݅ܦ‬ௌ௔௧௨௥௔௧௜௢௡಴೓೙೗ ሺ݅ሻ ‫݌‬ௌ௔௧௨௥௔௧௜௢௡಴೓೙೗ ି௏௔௥ ൌ ሺ‫ ݐܪ‬െ 1ሻ ∑ௐௗ ∑ு௧ ܸ݈ܽ‫݁ݑ‬஼௛௡௟ ሺ‫ݕ ,ݔ‬ሻAccordingly the Value channel parameters are obtained using the following equations ‫݌‬௏௔௟௨௘಴೓೙೗ ିெ௘௔௡ ൌ ௫ୀଵ ௬ୀଵ ܹ݀ ‫ݐܪ כ‬ ∑௜ ‫݂݂݅ܦ‬௏௔௟௨௘಴೓೙೗ ሺ݅ ሻ ‫݌‬௏௔௟௨௘಴೓೙೗ ି௏௔௥ ൌ ሺ‫ ݐܪ‬െ 1ሻ The ‫ ܵܥܧܲܯ‬discussed in this paper discusses the extraction 21 vital parameters whichwould enable for classification of skin lesions especially targeted towards early detection ofskin cancer melanoma. The parameter extraction and the procedure involved in extraction areachieved adopting a six phase approach discussed above. 3.4. ࡹࡲࡺࡺ CLASSIFIER Figure 1 : A MFNN Classifier Structure 27
  • International Journal of Computer Engineering and Technology (IJCET), ISSN 0976 –6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 1, January- February (2013), © IAEME The classification of the testing data set of dermoscopic images requires a supervisedtraining procedure to be incorporated. A multilayer feed forward neural network (‫)ܰܰܨܯ‬with the back propagation algorithm for training is adopted for the purpose of classification.of the neural network. A sample ‫ ܰܰܨܯ‬is shown in Fig 1. The ‫ ܰܰܨܯ‬consists of a set ofThe parameters extracted from the testing image set are directly provided to the input neuronsinput neurons ,‫ ܯ‬number of hidden layers and an output layer.The ‫ ܰܰܨܯ‬classifier is amachine learning algorithm. The learning and classification is achieved based on thelayer represented as ܽఊ where ߛ ൌ 1,2,3, … c୰ and c୰ is the trained ‫ ݎ‬classes.weightsࣱ which are changed in the neighboring layers. The input to the neuron in every ࣿࣾషభ The output of the neuron is defined asܽࣼ ൌ ݂൫߮ࣼ ൯ ൌ ݂ ቌ ෍ ࣱࣼࣻ ܽࣻ ሺࣾିଵሻఊ ቍ ࣾఊ ࣾఊ ࣾ ࣻୀ଴݉represents the index of the hidden layer i.e. 1 ൑ ݉ ൑ ‫ܯ‬Where݊௠ represents the number of neurons in݉௧௛ layer݆ represents the index ݆ | 1 ൑ ݆ ൑ ݊௠߮ ࣼ represents the weighted sum of input values for ࣼ ௧௛ neuron in layer ࣱܽࣼ௜ represents the weight between ࣼ ௧௛ neuron in ݉௧௛ layer and ݅ ௧௛ neuron in ሺ݉ െ 1ሻ௧௛ layer γ ௠The classification error ߜࣼ ം observed at the output layer is obtained from the following ࣧdefinitionܽࣼ ൌ ݂ ′ ൫߮ࣼ ൯ߜࣼ ൌ ݂ ′ ൫߮ࣼ ൯൫‫ࣼݕ‬ െ ‫ ࣼݕ‬൯ ࣾఊ ெఊ ఊ ெఊ ࣴఊ ఊThe sum of square errors observed through the ‫ ܰܰܨܯ‬is defined as 1 ௔ ߦఊ ൌ ෍൫ߜ௝ ൯ ఊ ଶ 2 ࣼୀଵ ௔ࣾశభThe error existent at the output layers is propagated back and is defined as ൌ݂ ൫߮ࣼ ൯ ෍ ߜℓ ሺࣾାଵሻఊ ሺࣾାଵሻߜࣼ ࣱℓࣼ ࣾఊ ᇱ ࣾఊ ℓୀଵThe errors propagated is minimized by altering the weights of the neighboring neurons and∆ஓ ࣱ௝௜ ൌ ߟߜࣼ ‫ࣻݑ‬ ௠ ࣾఊ ሺࣾିଵሻఊthe weight update function is defined as ୡ౨ ୡ౨ 1 ௔The classification error that exist through ଶߦ஼௟௔௦௦ ൌ ෍ ߦఓ ൌ ෍ ෍ቀ‫ ࣼݕ‬െ ߮ ࣼ ቁ ୸ஓ ஓ 2 ఓୀଵ ఓୀଵ ࣼୀଵ The ‫ ܰܰܨܯ‬Classifier is designed to minimize the classification error by effectiveweight adjustments achieved due to the back propagation algorithm. 28
  • International Journal of Computer Engineering and Technology (IJCET), ISSN 0976 –6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 1, January- February (2013), © IAEME The ‫ ܵܥܧܲܯ‬adopts machine learning techniques for classification. The ‫ ܰܰܨܯ‬basedclassifier is adopted for classification which classifies based on the supervised learningimparted using the back propagation algorithm. Parameter extraction is achieved using a sixphase approach to define the skin lesion. The training dataset constitutes of a set ofdermoscopic images which are pre-classified with the aid of dermatologists. The imagesset. The ‫ ܰܰܨܯ‬classifier is trained using the training data set. The testing set consists ofdefined in terms of the parameters and the classified type constitute towards the training datadermoscopic images that need to be diagnosed. The diagnostic results are represented asusing the ‫ ܵܥܧܲܯ‬parameter extraction phases. The resulting dataset is termed as the testclasses. To enable diagnosis the test images are represented as a set of parameters extracteddataset. The test dataset is provided to the trained ‫ ܰܰܨܯ‬for classification. The subsequentܸܵ‫ ܯ‬classifiers.section of this paper discusses the experimental verification and comparisons with the4. EXPERIMENTAL STUDY AND COMPARISONS The ‫ ܵܥܧܲܯ‬discussed in this paper was realized on the MATLAB platform. In order toevaluate the ‫ ܵܥܧܲܯ‬training and testing dataset are to be created. To construct the trainingand testing datasets dermoscopic images are obtained from the “The Atlas of Dermoscopy”[25]. The atlas considered is a collection of over 2000 skin lesion images obtained froma reference to evaluate the classification accuracy of the ‫ ܵܥܧܲܯ‬system proposed. The use ofvarious patients. The atlas also provides the classification and diagnosis data which is used asthe atlas is charted as others researchers too have used the same atlas to evaluate their‫ ܵܥܧܲܯ‬proposed is compared with the popular ܸܵ‫ ܯ‬classifiers as proposed by Xiaojingproposed diagnosis mechanisms especially skin cancer melanoma [20] [26]. TheYuan et al.[18]. CLASSIFICATION ERROR 14 13.33 % 12 CLASSIFICATION ERROR IN % 10 8 6 4 2 < 1% 0 MEPCS - CLASSIFICATION ERROR SVM CLASSIFIER - CLASSIFICATION ERROR Figure 2 : Classification Error Observed 29
  • International Journal of Computer Engineering and Technology (IJCET), ISSN 0976 –6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 1, January- February (2013), © IAEME The training dataset consists of pre-classified dermoscopic images. We have considered aset of 45 dermoscopic images from the “Atlas of Dermoscopy” as the training data set. Thetraining dataset consist of skin lesion images of primarily 3 classes namely “AdvancedMelanoma”, “Early Melanoma” and “Non-Melanoma”. The 21 parameters extracted perimage using the 6 phase approach contribute to each record of the training data set along withThe ‫ ܵܥܧܲܯ‬and the ܸܵ‫ ܯ‬Classifier was trained using similar training datasets. Similar testthe classified name. A similar approach is followed to construct the testing sets were evaluated used the ‫ ܵܥܧܲܯ‬and ܸܵ‫ ܯ‬classifier. The classification resultsobtained are provided to the Receiver Operating Characteristic (ܴܱ‫ܥ‬ሻ analysis tool. The errorclassification error of the ܸܵ‫ ܯ‬classifier is greater than the error exhibited by the proposedin classification obtained is graphically shown in Fig 2. From the figure it is evidentthat the‫ ܵܥܧܲܯ‬system.The ‫ ܵܥܧܲܯ‬and the ܸܵ‫ ܯ‬classifier classifiers are compared on the basis ofthe receiver operating characteristic (ܴܱ‫ )ܥ‬analysis. The ܴܱ‫ ܥ‬analysis is carried out using aThe ܴܱ‫ ܥ‬analysis plot obtained is shown in Fig 3.tool developed using C#.Net programming language on the Visual Studio 2010 platform. Receiver Operating Characteristic (ROC) - CURVE 1 0.9 0.8 0.7 SENSITIVITY 0.6 0.5 0.4 0.3 0.2 MPECS SVM CLASSIFIER 0.1 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Figure 3 : ‫ ۱۽܀‬Curve Analysis 1- SPECIFICITY The classification accuracy and the efficiency of the ‫ ܵܥܧܲܯ‬Classifier along with theܸܵ‫ ܯ‬Classifier are diagrammatically shown in Fig 4 and Fig 5. The ‫ ܵܥܧܲܯ‬systemexhibitsan 81% accuracy in classification as compared to the 75% accuracy exhibited by the ܸܵ‫ܯ‬Classifier. The ‫ ܵܥܧܲܯ‬system proposed in this paper improves the efficiency inclassification by about 7.4% when compared to the ܸܵ‫ ܯ‬classifier.The experimental studythe ‫ ܵܥܧܲܯ‬system proposed in this paper is capable of accurately classifying skin lesionsconducted and from the results discussed in this section of the paper, it can be concluded thatexhibiting early melanotic symptoms when compared to the popularly used ܸܵ‫ ܯ‬classifiersystems. 30
  • International Journal of Computer Engineering and Technology (IJCET), ISSN 0976 –6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 1, January- February (2013), © IAEME CLASSIFIER ACCURACY 0.81 0.8 0.79 0.78 ACCURACY 0.77 0.76 0.75 0.74 0.73 0.72 MEPCS ACCURACY SVM CLASSIFIER ACCURACY Figure 4 : Classification Accuracy Comparisons 0.83 CLASSIFIER EFFICIENCY 0.82 0.81 0.8 EFFECIENCY 0.79 0.78 0.77 0.76 0.75 0.74 0.73 MPECS EFFICIENCY SVM CLASSIFIER EFFICIENCY Figure 5: Classification Efficiency Comparisons5. CONCLUSION This paper highlights the benefits and the requirement for early detection of skin cancerpaper relies on a supervised ‫ ܰܰܨܯ‬based classifier to accurately classify skin lesions and enablemelanoma. The Multi Parameter Extraction and Classification System (‫ )ܵܥܧܲܯ‬discussed in thisearly detection of skin cancer melanoma. The ‫ ܵܥܧܲܯ‬comprehensively defines skin lesions in termsof the ‫ ܵܥܧܲܯ‬is discussed in this paper. The ‫ ܵܥܧܲܯ‬system is trained using the Back Propagationof a parameter sets extracted using the six phase approach. The training and testing phase algorithmsAlgorithm. The trained classifier of the ‫ ܵܥܧܲܯ‬is used for analysis and diagnosis of skin lesions onunknown types. The classification preeminence of the ‫ ܵܥܧܲܯ‬is proved over the popular SVMclassifier systems.The results obtained and discussed prove that the ‫ ܵܥܧܲܯ‬system proposed in thispaper aids the detection of skin cancer melanoma in the naïve stages thereby improving computeraided diagnosis mechanisms and early treatment initiatives to be adopted by dermatologists for thebenefit of patients suffering from the life threatening melanotic skin cancer diseases. 31
  • International Journal of Computer Engineering and Technology (IJCET), ISSN 0976 –6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 1, January- February (2013), © IAEMEREFERENCES[1] R. Marks, “Epidemiology of melanoma,” Clin. Exp. Dermatol., vol. 25, pp. 459–463,2000[2] World Health Organization, Ultraviolet Radiation and the INTERSUN Programme.(2007. Dec.). [Online]. Available:[3] Imran Ali, Waseem A. Wani and KishwarSaleem,"Cancer Scenario in India with FuturePerspectives",Cancer Therapy Vol 8, 56-70, 2011[4] SinhaR,Anderson DE, Mc Donald SS and Greenwald P, "Cancer Risk and Diet in India",JPostgrad Med 2003,49: 222-228[5] American Cancer Society,"Cancer Facts and Figures 2012". 2012, American CancerSociety, Inc. pp 1-68[6] Brian D’Alessandro, Atam P. Dhawan,"3D Volume Reconstruction of Skin Lesions forMelanin and Blood Volume Estimation and Lesion Severity Analysis",IEEE Transactions onMedical Imaging, July 2012 DOI : 10.1109/TMI.2012.2209434[7] Ruksar Fatima, Mohammed Zafar Ali Khan, A. Govardhan and Kashyap DDhruve."Computer Aided Multi-Parameter Extraction System to Aid Early Detection of SkinCancer Melanoma ",IJCSNS International Journal of Computer Science and NetworkSecurity,Vol. 12 No. 10 pp. 74-86.2012[8] G. Argenziano, H. P. Soyer, S. Chimenti, R. Talamini, R. Corona, F. Sera, M. Binder, L.Cerroni, G. De Rosa, G. Ferrara, R. Hofmann-Wellenhof,M. Landthaler, S.W.Menzies, H.Pehamberger, D. Piccolo, H. S. Rabinovitz, R. Schiffner, S. Staibano, W. Stolz, I. Bartenjev,A. lum, R. Braun, H. Cabo, P. Carli, V. De Giorgi, M. G. Fleming, J. M. Grichnik, C. M.Grin, A. C. Halpern, R. Johr, B. Katz, R. O. Kenet, H. Kittler, J. Kreusch, J. Malvehy, G.Mazzocchetti, M. Oliviero, F. Ozdemir, K. Peris, R. Perotti, A. Perusquia, M. A. Pizzichetta,S. Puig, B. Rao, P. Rubegni, T. Saida, M. Scalvenzi, S. Seidenari, I. Stanganelli, M. Tanaka,K. Westerhoff, I. H. Wolf, O. Braun-Falco, H. Kerl, T. Nishikawa, K. Wolff, and A. W. Kopf,“Dermoscopy of pigmented skin lesions: Results of a consensus meeting via the Internet,” J.Amer. Acad. Dermatol., vol. 48, no. 5, pp. 680–693, 2003.[9] V. T. Y. Ng, B. Y. M. Fung, and T. K. Lee, “Determining the asymmetry of skin lesionwith fuzzy borders,” Comput. Biol. Med., vol. 35, pp. 103–120, 2005[10] M. E. Celebi, Y. A. Aslandogan,W. V. Stoecker, H. Iyatomi, H. Oka, and X. Chen,“Unsupervised border detection in dermoscopy images,” Skin Res. Technol., vol. 13, no. 4,pp. 454–462, 2007[11] S. E. Umbaugh, R. H. Moss, and W. V. Stoecker, “An automatic color segmentationalgorithm with application to identification of skin tumor borders,” Comput. Med. Imag.Graph., vol. 16, pp. 227–235, 1992[12] V. T. Y. Ng, B. Y. M. Fung, and T. K. Lee, “Determining the asymmetry of skin lesionwith fuzzy borders,” Comput. Biol. Med., vol. 35, pp. 103–120, 2005[13] G. Betta, G. Di Leo, G. Fabbrocini, A. Paolillo, and M. Scalvenzi, “Automatedapplication of the “7-point checklist” diagnosis method for skin lesions: Estimation ofchromatic and shape parameters,” in Proc. Instrum. Meas. Technol. Conf. (IMTC 2005), May,vol. 3, pp. 1818– 1822.[14] Tasoulis, S.K.; Doukas, C.N.; Maglogiannis, I.; Plagianakos, V.P.; , "Classification ofdermatological images using advanced clustering techniques," Engineering in Medicine andBiology Society (EMBC), 2010 Annual International Conference of the IEEE , vol., no.,pp.6721-6724, Aug. 31 2010-Sept. 4 2010 doi: 10.1109/IEMBS.2010.5626242 32
  • International Journal of Computer Engineering and Technology (IJCET), ISSN 0976 –6367(Print), ISSN 0976 – 6375(Online) Volume 4, Issue 1, January- February (2013), © IAEME[15] Tim K. Lee*, M. Stella Atkins, Michael A. King, Savio Lau, and David I.McLean´Counting Moles Automatically From Back Images‖ IEEE TRANSACTIONS ONBIOMEDICAL ENGINEERING, VOL. 52, NO. 11, NOVEMBER 2005[16] Faal, M., MiranBaygi, M. H. and Kabir, E. (2013), Improving the diagnostic accuracyof dysplastic and melanoma lesions using the decision template combination method. SkinResearch and Technology, 19: e113–e122. doi: 10.1111/j.1600-0846.2012.00617.x[17] M. E. Celebi, H. A. Kingravi, B. Uddin, H. Iyatomi, Y. A. Aslandogan, W. V.Stoecker, and R. H. Moss, ―A methodological approach to the classification of dermoscopyimages,‖Comput. Med. Imag. Graph., vol. 31, pp. 362–373, Mar. 2007[18] Xiaojing Yuan; Zhengyu Yang; Zouridakis, G.; Mullani, N.; , "SVM-based TextureClassification and Application to Early Melanoma Detection," Engineering in Medicine andBiology Society, 2006. EMBS 06. 28th Annual International Conference of the IEEE , vol.,no., pp.4775-4778, Aug. 30 2006-Sept. 3 2006 .doi: 10.1109/IEMBS.2006.260056[19] Steven McDonagh, Robert B. Fisher and Jonathan Rees,"Using 3D information forclassification of non-melanoma skin lesions",Available[20] M. EmreCelebi, Hassan A. Kingravia, Y. Alp Aslandogan, William V.Stoecker,"Detection of Blue-White Veil Areas in Dermoscopy Images Using MachineLearning Techniques",Proc. of SPIE Vol. 6144.doi: 10.1117/12.655779[21] M. F. Healsmith, J. F. Bourke, J. E. Osborne, and R. A. C. Graham- Brown, “Anevaluation of the revised seven-point checklist for the early diagnosis of cutaneous malignantmelanoma,” Br. J. Dermatol., vol. 130, pp. 48–50, 1994[22] R. B. Mallet, M. E. Fallowfield, M. G. Cook, W. N. Landells, C. A. Holden, and R. A.Marsden, “Are pigmented lesion clinics worthwhile?” Br. J. Dermatol., vol. 129, pp. 689–693, 1993[23] Dangare, Chaitrali S. and Apte, Sulabha S., A Data Mining Approach for Prediction ofHeart Disease Using Neural Networks (November 14, 2012). International Journal ofComputer Engineering and Technology (IJCET), Volume 3, Issue 3, October-December 2012.Available at SSRN:[24] S. ShenbagaEzhil and Dr. C. Vijayalakshmi. "PREDICTION OF COLON-RECTUMCANCER SURVIVABILITY USING ARTIFICIAL NEURAL NETWORK",InternationalJournal of Computer Engineering and Technology (IJCET),Volume 3, Issue 1, January- June(2012), pp. 163-168[25] Marghoob AA, Braun RP, Kopf AW. Atlas of dermoscopy. London: Taylor & Francis;2005[26] Ru-Song Meng, Xiao Meng, Feng-Ying Xie and Zhi-GuoJiang,"Early diagnosis forcutaneous malignant melanoma based on the intellectualized classification and recognitionfor images of melanocytic tumour by dermoscopy",Journal of Biomedical Graphics andComputing, December 2012, Vol. 2, No. 2. pp 37-47[27] S. Aruna, L.V. Nandakishore and Dr S.P. Rajagopalan, “A Novel Lns Semi SupervisedLearning Algorithm For Detecting Breast Cancer” International journal of ComputerEngineering & Technology (IJCET), Volume3, Issue1, 2012, pp. 44 - 53, Published byIAEME[28] S. Shenbaga Ezhil and Dr. C. Vijayalakshmi, “Prediction of Colon-Rectum CancerSurvivability Using Artificial Neural Network” International journal of ComputerEngineering & Technology (IJCET), Volume3, Issue1, 2012, pp. 163 - 168, Published byIAEME 33