COMPUTER-ASSISTED SKIN LESION DIAGNOSIS AND MELANOMA CANCER DETECTION

Abstract

Skin cancer is among the most common cancer types that affect people across the world. According to the American Cancer Society statistical data, an approximate increase of new melanoma cases will be 5.8%, and the death rate will increase by 4.8%. Medical specialists believe that early diagnosis is the key to effective treatment. The preliminary visual examination of skin lesions by dermatologists had a low accuracy rate (60-70%). However, melanoma-type cancer is hard to distinguish from nevus. Machine learning and CNN applications on dermatoscopic images for the early detection of melanoma cancer produced tremendous outcomes in the past few decades. The primary objective of this thesis is to develop a robust CAD system for the analysis of skin lesion images. At the same time, the secondary objective is to study the methodologies adopted in many research articles on skin cancer identification and classification. In this thesis, we proposed a modified capsule UNET architecture for segmentation tasks and FixEfficientNet as a classification model and performed analysis on ISIC 2017 and PH2 datasets. Moreover, we present a brief discussion of recent approaches evolved to detect and classify skin cancer and perform a comparative analysis of performances achieved by different systems developed for skin cancer detection and classification. The proposed work achieved an 88.3% dice score, 93.4% accuracy, 94.7 % sensitivity, and 96.3% of specificity. Even though these approaches achieved a significant result, there are many challenges for successful diagnostics. We also discussed these challenges and proposed potential advancement as future work to overcome these challenges. The comprehensive analysis of different methods suggests that advanced deep learning algorithms provide a robust technique for the multi-class classification of pigmented skin lesion images.