Identification Of Skin Lesions Via Pattern Recognition Techniques

This thesis proposes a generic method of description of skin lesion types and recognition of Melanocytic lesions in Dermoscopic images.These images are acquired by dermoscopic instruments publically available for experimental analysis of skin cancer identification.Scientific judgments show that cancerous lesions demonstrate various characteristics such as streaks,texture or pigmented networks as compared to normal or benign lesions.Our aim is to confirm those judgments by carrying out proper segmentation,features extraction and classification.Segmentation phase of PH2 Public dataset was done by means of active contours.Features extraction stage included implementation of Local Binary Patterns and its extension to study textural features more deeply.Finally,K nearest neighbor and Support vector machine were used for classification.This choice of methods is motivated by scientific importance of differential configurationally structures in the melanomas.The experiment of the offered algorithm is done on both sets of images,manually segmented(segmentation done by physicians followed by Features extraction and classification done by us)and automatic segmented images(all the steps performed by us)using performance metrics of sensitivity,specificity and Accuracy.The highest result of the system was attained by manually segmented images,i-e SE=89.2,SP=84,Accuracy=89.2.Whereas,images with automated segmentation showed SE=89.3,SP=75.4,Accuracy=88.2.