MAHSNet: Multi-Scale Adaptive Hybrid Semantic Network for Medical Image Segmentation

Benefiting from self-attention, the Transformer-based methods excel in global context modeling. However, these methods have limitations in capturing local dependencies between pixels, making it difficult to accurately segment small lesions, small tissues, and complex target boundaries. To address this issue, we propose the Multi-Scale Adaptive Hybrid Semantic Network (MAHSNet) for medical image segmentation. The core of MAHSNet is the Efficient Pure Attention Cascaded Decoder (EPACD). Unlike traditional CNN-Transformer decoder architectures, EPACD introduces gate attention, coordinate attention, and boundary attention to accurately segment local details and target boundaries, while significantly reducing the model's parameters and computational complexity. Experimental results on six datasets across two modalities demonstrate that MAHSNet outperforms existing state-of-the-art methods in multiple medical segmentation tasks. For example, the Dice scores on the Synapse datasets reached 83.76%. The parameters and computational complexity were reduced by 73% and 82% compared to TransUnet, achieving dual improvements in segmentation performance and efficiency.