Structure-preserving lightweight skin lesion segmentation with dual-resolution attention and boundary-aware fusion
Xu Yang, Duwei Dai, Caixia Dong, Liang Wang, Pengyu RenObjective
Dermoscopic skin lesion segmentation suffers from low contrast, artifacts, and blurred boundaries in lightweight models due to lossy downsampling and rigid tokenization. We propose Structure-Preserving Representation Learning (SP-RL), a lightweight framework designed to preserve geometric topology and continuity for stronger representation learning.
Methods
SP-RL integrates four components: 1) Structure-Preserving Sampling Strategy (SPSS) using a bijective pixel-folding mechanism to redistribute spatial details into channel dimensions; 2) Lite Multi-Scale Aggregation Block (LMA-Block) to expand receptive fields at high resolutions; 3) Dual-Resolution Attention (DR-Atn) maintaining high-resolution queries while compressing keys/values via hierarchical pooling; and 4) Dual-Gated Multi-Scale Fusion (DG-MSF) to suppress low-level artifacts and inject noise-resilient Sobel gradient priors into progressive deconvolution paths.
Results
Evaluations across 23 models show SP-RL achieves a 0.8976 Dice score and the lowest HD95 (15.1829) on ISIC2018 among lightweight alternatives. It exhibits superior robustness across four public datasets (ISIC 2016/2017/2018,
Conclusion
SP-RL offers an optimal trade-off between boundary accuracy, reliability, and efficiency. By retaining high-frequency energy and domain-invariant morphology, it provides a dependable solution for resource-constrained diagnosis.