WA-NET: enhanced boundary-aware segmentation of skin lesions via frequency-spatial feature fusion and attention-guided edge refinement

Skin cancer remains a major public health concern due to its high morbidity and mortality rates. While automatic segmentation techniques have improved diagnostic accuracy, they continue to face challenges such as artefacts, small lesion detection, and poor contrast between lesions and surrounding tissue. To overcome these limitations, we propose WA-NET, a novel skin lesion segmentation network that integrates a Boundary Refinement module (BRM) and an Enhanced Wavelet Transform (EWT) module. The BRM employs independent edge detection branches to enhance boundary representation, particularly in low-contrast regions. The EWT module adaptively fuses multi-scale, multi-directional sub-band features in the frequency domain to better capture texture and structural details. Furthermore, a composite loss function combining binary cross-entropy, Dice loss, and edge-supervised loss is introduced to improve both global segmentation accuracy and local boundary precision. WA-NET achieves state-of-the-art performance on three benchmark datasets-ISIC2017 (DSC: 0.9395, SE: 0.9357, ACC: 0.9573), ISIC2018 (DSC: 0.9458, SE: 0.9460, ACC: 0.9610), and PH2 (DSC: 0.9517, SE: 0.9593, ACC: 0.9638)-demonstrating strong robustness and superior boundary segmentation under challenging imaging conditions.