Medical image segmentation remains challenging due to inherent characteristics such as low contrast, blurred boundaries, and complex anatomical structures. While state space models, particularly Mamba, have demonstrated remarkable capabilities in capturing long-range dependencies with linear computational complexity, existing Mamba-based approaches for medical segmentation still struggle with precise boundary delineation and effective local-global feature fusion. To address these limitations, we propose BDFNet, a unified architecture that integrates Boundary-Aware Attention (BAA) and Dual-Stream Frequency Attention Module (DFAM) within the Mamba framework. The BAA module explicitly enhances boundary information through multi-branch attention mechanisms and provides auxiliary boundary supervision signals. The DFAM module employs frequency domain analysis to effectively fuse local boundary features with global semantic representations through com plementary attention weighting. Comprehensive experiments on ACDC cardiac segmentation, Synapse multi organ segmentation, and ISIC skin lesion datasets validate the effectiveness of our approach, with notable improvements in both segmentation accuracy and boundary precision. Experimental results show that our method achieves superior performance compared to existing approaches across cardiac, abdominal organ, and skin lesion segmentation benchmarks. Ourcodeispublicly available at https://github.com/apellidole/Mamba-BDF.
Geng et al. (Thu,) studied this question.