Melanoma Detection in Skin Cancer Images using Deep Learning Architectures

Abstract

Melanoma is widely recognized as the most lethal variant among various types of skin newlinecancer, primarily affecting individuals with pigmented skin. An early diagnosis of newlinemelanoma can significantly increase the likelihood of a successful course of therapy. newlineHence, an accurate early detection technique is critically needed. The current auto- newlinemated algorithms for melanoma detection primarily rely on color images. Contempo- newlinerary developments have evolved to perform early detection of melanoma automatically. newlineDue to the initiation of dermoscopy, clinical diagnostic abilities have been augmented newlineso that detection of melanoma can be done at the earliest stages. Emerging techniques newlinein Image Processing(IP) and Machine Learning(ML) have demonstrated the potential newlineto differentiate between Malignant Melanoma(MM) and Benign(B) cases. However, newlineseveral newly developed systems have shown a wide range of implementations, indi- newlinecating the need for further technical advancements to improve accuracy and reliability. newlineAs there are many variations in diagnosis, it is significant that the diagnosis must be newlinedone at the proper stage and time. Metastasis is a critical stage in cancer progression newlineand can significantly affect treatment options and patient prognosis. This thesis aims to newlineexplore and enhance the techniques used for the automated analysis of skin lesions thus newlineexpanding the capabilities of existing techniques employed with different phases of the system for automatic diagnosis. newlineThe objective of the research is to conduct a comprehensive investigation into auto- newlinemated skin cancer diagnostic systems. This includes enhancing and advancing segmen- newlinetation, feature selection, and classification methods to effectively handle the intricate newlinestructures found in dermoscopic/digital images. newlineDuring this thesis, several algorithms were examined. It presents a highly effective newlineevolutionary algorithm (EA) that enables the simultaneous optimization of deep neu- newlineral network (DNN) architecture and hyperparameters. There are several enhancements newlinemade