Abstract
Proper and accurate classification of skin lesions is essential for melanoma diagnosis. Distinguishing lesion regions from surrounding healthy skin is difficult due to blurred tissue patterns, irregular shapes, and faint boundaries. Conventional segmentation techniques often lose high-frequency structural details during downsampling, which compromises boundary precision. This article proposes WIDENet, a lightweight segmentation framework that preserves fine lesion boundaries while maintaining computational efficiency. WIDENet combines a multi-resolution wavelet, a transformer-based contextual encoder, and an attentive feature refinement mechanism. In contrast to classical encoder-decoder systems, spatial downsampling is performed in the frequency domain via wavelet decomposition, thereby reducing feature dimensionality while preserving multiscale structural information. The encoder based on transformers enables long-range semantic perception. In contrast, the attention-learned refinement framework focuses on information related to boundaries and suppresses background noise. We tested WIDENet using three benchmark dermoscopic datasets, including$\mathrm {PH}^{2}$, ISIC 2017, and ISIC 2018. Comparison with state-of-the-art models using U-Net, U-Net++, DeepLabV3(+), PAN, PSPNet, and LinkNet demonstrates that WIDENet consistently outperforms them. In particular, it achieves Dice scores of 0.9384 on$\mathrm {PH}^{2}$and 0.8752 on ISIC 2017, and high precision, recall, and IoU across all datasets. These results demonstrate that WIDENet is not only correct but also highly computationally efficient, making it a promising tool for real-world melanoma screening and clinical decision support systems.