Abstract
Accurate skin lesion segmentation is a critical foundation for computer-aided dermoscopic diagnosis. Existing CNN-Transformer hybrid methods still fall short in achieving deep fusion of global semantics and local details, as well as in task-specific boundary guidance mechanisms, especially given the core challenges in dermoscopy images: highly blurred lesion boundaries, complex and variable internal textures, and ubiquitous interference from hair occlusion and illumination artifacts. To address these issues, this paper proposes a novel segmentation network named DA-CTransNet, which deeply integrates three core modules: domain-knowledge-aware adaptive attention, multi-scale collaborative affine calibration, and cascaded boundary refinement. Specifically: First, we construct a domain-knowledge adaptive module that encodes clinically validated key visual diagnostic patterns (e.g., textures and vascular structures of specific lesions) into learnable feature prompt vectors via a lightweight sub-network, and adaptively injects them into the self-attention computation of the Transformer to achieve targeted enhancement of semantic features. Second, we propose a multi-scale collaborative affine calibration mechanism that, based on cross-spatial and cross-channel attention interaction, explicitly estimates and compensates for statistical distribution and spatial geometric discrepancies among features from different levels, constructing an intrinsically consistent and aligned feature pyramid. Finally, we design a dual-pathway cascaded boundary refinement decoder: the global pathway performs coarse segmentation, while the boundary pathway introduces a spectral-aware uncertainty mapping module to locate blurred boundary regions, combined with a high-order edge-preserving loss for end-to-end learning to achieve sub-pixel boundary reconstruction. Extensive experiments on the ISIC 2017 and ISIC 2018 datasets demonstrate that, compared to the baseline model, DA-CTransNet improves the Dice coefficient by approximately 2.3%, with significant improvements in Boundary F1 score and HD95 metrics; compared to existing mainstream methods, the Hausdorff distance is reduced by approximately 15%, and the model exhibits stronger robustness on images containing artifacts. This study provides a new technical pathway for building high-performance clinical intelligent auxiliary systems.