DAHE-Net: Dual-Attention Mechanism and Hybrid Multi-scale Extraction Network for Enhanced Skin Lesion Segmentation

Accurate segmentation of skin lesions is critical for clinical diagnosis. However, challenges include significant variations in lesion size and morphology, hair occlusion, interference, and low contrast between pathological and normal tissues. To address these, we propose DAHE-Net — a U-Net framework integrating hybrid multi-scale feature extraction and a dual-attention mechanism for feature refinement. The network employs heterogeneous multi-scale convolutional blocks to process input data, capturing local and global features through diverse receptive fields to improve lesion localization. An extended pyramid pooling module (EPPM) is designed to adaptively extract hierarchical semantic information, addressing scale-sensitive issues such as variations in lesion size and morphology. A dual-attention dynamic focus (DADF) mechanism fused with channel-spatial attention suppresses noise interference while focusing on information-rich features, enhancing fine-grained detail capture and occlusion handling capabilities. Extensive evaluations on the ISIC-2018 and ISIC-2017 benchmarks demonstrate DAHE-Net’s superiority. It achieves key metrics of 77.93% IoU (+6.69% over U-Net) and 87.23% Dice (+4.46%) on ISIC-2018, and 67.78% IoU (+3.07%)/80.26% Dice (+2.34%) on ISIC-2017. Cross-domain validation on brain tumor data shows further generalization, with 80.73% IoU (+4.56% against U-Net) and 89.25% Dice (+2.86%). These results substantiate DAHE-Net’s potential for advancing automated skin lesion analysis and computer-aided diagnosis.