Identifying Malignant Skin Diseases Through Deep Learning
Abstract
Early and accurate identification of malignant skin diseases is critical to improving treatment outcomes and patient prognosis. Current diagnostic methods heavily rely on visual assessments by medical professionals, followed by biopsies for confirmation. These approaches can be subjective and require significant expertise, while access to specialized care remains limited, particularly in developing regions. Recent advancements in artificial intelligence, specifically deep learning, present an opportunity to improve dermatological diagnostics. This paper explores the application of two deep learning models, ResNet and Vision Transformer (ViT), for the automated detection of malignant skin conditions. Both models were trained on the PAD-UFES-20 dataset, which comprises clinical skin lesion images captured via smartphones. Our findings suggest that these models can offer real-time diagnostic support at reduced costs, with accuracy levels comparable to or exceeding human performance, with the best experiment achieving the F1 score of 79.89% and an accuracy of 80.63%.
