Enhanced Skin Lesion Segmentation via Attentive Reverse-Attention U-Net

Accurate identification and segmentation of skin lesions are essential for the early diagnosis of skin cancer. Symmetry is an important diagnostic cue in clinical practice, as malignant lesions often exhibit asymmetric patterns in shape, color, and texture. Therefore, incorporating symmetry-based features into automated analysis can enhance segmentation reliability and improve diagnostic accuracy. However, automated lesion segmentation faces significant challenges, including blurred boundaries, low-contrast lesions, and heterogeneous backgrounds. To address these challenges, we propose a hybrid deep neural network model that enhances the traditional U-Net architecture with an integrated reverse-attention module embedded within its skip connections. This innovation sharpens feature extraction in ambiguous regions, boosting segmentation accuracy, particularly in complex areas. The model employs a multifaceted loss function approach—encompassing binary cross entropy, dice, Tversky, and compound losses—to effectively manage data imbalances while preserving lesion boundary details. Experimental validation on the ISIC2018 and PH2 datasets demonstrates the model’s efficacy, achieving dice similarity coefficients of 88.71% and 93.41% and mean intersection over union values of 87.68% and 90.78%, respectively. These results underscore the potential of our approach for clinical applications.