A comprehensive Framework for the detection of melanoma From dermoscopy images based on modified Total dermoscopy score

Abstract

Melanoma is a malignant skin lesion. Melanoma is newlinedistinguished from nevus using Total Dermoscopy Score (TDS) that newlinequantitatively reflects the geometric, texture and colour features of the newlinelesions, computed from dermoscopy images. Subjective computation of newlinethe TDS is time-consuming and prone to inter/intra operator variability. newlineHence, an exhaustive image-processing framework for the automated newlinediagnosis of melanoma is proposed in this thesis. newlineThe framework is composed of modules for pre-processing, newlinesegmentation, feature extraction and classification. The pre-processing newlinemodule is composed of algorithms for denoising, illumination newlinecorrection, contrast enhancement, sharpening, reflection removal and newlinevirtual shaving, cascaded as a pipeline. newlineThe Expectation Maximization (EM) algorithm is used for newlinelocalizing the lesions. To distinguish melanoma from nevus, a refined newlineform of the traditional TDS known as Modified TDS (MTDS) is given newlineas the input to the Support Vector Machine (SVM) classifier. MTDS newlinescore is computed as the weighted sum of distinct features that reflect newlineasymmetry of lesions, strength and continuity of their borders, newlinepigmentation, presence of dermoscopic structures like globules, streaks, newlinenon-textured regions etc. and geometrical features like area, solidity, newlinecircularity etc. The weights corresponding to each feature is computed newlinevia logistic regression. newlineImages from ISIC 2016 and PH2 datasets are used for newlinevalidation. The EM segmentation exhibited the highest Dice Similarity newlinevi newlineIndex (DSI) than Fuzzy C-Means, Isodata thresholding, k-means, Otsu s newlinethresholding and region growing. The SVM with MTDS as input newlineexhibited the highest values of accuracy, sensitivity and specificity, newlinecompared to Gaussian Naive Bayes, Logistic Regression, Linear newlineDiscriminant Analysis, k-Nearest Neighbor and Decision Tree newlineclassifiers. newlineThe high values of DSI exhibited by EM ensure agreement of newlinelesion areas segmented by EM with manual ground-truths. The proposed newlineframework eliminates the subjectivity of melanoma diagnosis.