Abstract
Skin lesion segmentation is a crucial component of dermoscopic computer-aided diagnosis, yet challenges such as boundary ambiguity, morphological diversity, and noise interference under complex imaging conditions still limit the accuracy and robustness of existing methods. To address these issues, we propose a dual-prior hybrid segmentation network that integrates both boundary priors and shape priors. In the encoder, a gradient-driven Boundary-augmented Hybrid Attention module is constructed to jointly capture long-range contextual information through explicit boundary enhancement, self-attention, and state space-inspired modeling. In the decoder, a Multi-scale Lesion Shape Prior module is designed to impose global structural constraints on the segmentation mask via multi-scale shape priors and a unified loss formulation, thereby balancing fine-grained contour precision with overall morphological consistency. Evaluated on three public datasets-ISIC2018, HAM10000, and PH2-the proposed method achieves IoU/DSC scores of 92.4%/96.0%, 87.3%/93.2%, and 95.2%/97.5%, respectively, outperforming the strongest baseline by an average margin of 1.4 percentage points in IoU while reducing HD95 and ASD by approximately 0.8 and 0.06 on average. Moreover, with only 3.79G FLOPs, Our method surpasses a range of state-of-the-art Transformer and CNN-Transformer hybrid architectures, demonstrating its comprehensive advantages in accuracy, boundary quality, and computational efficiency.