Automatic detection of melanoma by deep learning models-based feature extraction and fine-tuning strategy

Abstract

Skin cancer is one of the most common and lethal cancers in the world. Even though melanoma accounts for a small fraction of all skin cancer types, it is currently the leading cause of skin cancer fatalities. In order to accurately identify the disease in the initial stages, dermoscopy was designed to help specialists and increase the detection rate. Deep learning has been found in recent research to be particularly beneficial in the detection of skin cancer and medical diagnostics. Within this context, we propose adopting two deep-learning algorithms to detect benign and melanoma skin conditions. The first option entails extracting features from CNN pre-trained networks and applying machine-learning classification to make predictions, while the second explores using a fine-tuning methodology combined with data augmentation. The goal of this work is to compare the two methodologies in terms of classification performance and time efficiency. A public database including dermoscopic images of melanoma and normal skin lesions is used for the experimental evaluation. The experimental evaluation reveals that pre-trained models offer superior performance in terms of accuracy and processing time.