Comparative Analysis of UNet Plus Plus and Attention UNet for Skin Lesion Segmentation

Abstract

Accurate skin lesion segmentation is a fundamental step in computer-aided melanoma diagnosis because it enables precise localization of lesion boundaries before further clinical analysis and classification. Deep learning-based semantic segmentation models have demonstrated significant potential in automating this task, with U-Net-based architectures being among the most widely adopted approaches. This study presents a comparative analysis of two advanced architectures, U-Net++ and Attention U-Net, using dermoscopic images from the International Skin Imaging Collaboration (ISIC) 2017 skin lesion dataset. To ensure a fair and objective comparison, both models were trained and evaluated under identical preprocessing procedures, training configurations, hyperparameters, and evaluation settings. Model performance was assessed using three segmentation metrics: Accuracy, Dice Similarity Coefficient (Dice), and Intersection over Union (IoU). Experimental results show that U-Net++ achieved superior overall segmentation performance, obtaining a Dice score of 0.7872, an IoU of 0.6955, and an accuracy of 0.9068. In comparison, Attention U-Net achieved an accuracy of 0.9039, an IoU of 0.6842, and a Dice score of 0.7757. Although both models produced competitive results, U-Net++ demonstrated better region overlap and segmentation accuracy, whereas Attention U-Net generated smoother lesion boundaries in some cases. These findings indicate that U-Net++ is more effective for skin lesion segmentation on the ISIC 2017 dataset and emphasize the importance of maintaining consistent experimental settings to ensure reliable and meaningful model comparisons.