Performance Evaluation of U-Net Based Methods for Lesion Segmentation from Dermoscopy Images

Abstract

Skin cancer accounts for approximately half of all cancer cases worldwide, making it one of the most prevalent types of cancer. Melanoma, which develops from melanocytes that give skin its color, is the most lethal among skin cancers. Early diagnosis of melanoma, particularly through dermoscopy images, is of vital importance. To this end, automated diagnostic systems significantly aid dermatologists in their decision-making processes. In recent years, advancements in deep learning and machine learning have improved diagnostic accuracy. Specifically, CNN-based deep learning algorithms are utilized for medical image analysis and skin lesion segmentation. While traditional methods struggle to capture fine details and broader context, the U-Net architecture overcomes these challenges, providing more accurate segmentation. This study evaluates U-Net, Residual U-Net, and Attention U-Net models for skin lesion segmentation. The performance of the models is measured using Dice Score, Jaccard Index, and train loss metrics. The results reveal that Attention U-Net demonstrates the highest performance, with a Dice Score of 0.8063 and a Jaccard Index of 0.7203.