BMANet: Boundary-aware Multi-Scale Attention Network for medical image segmentation

Medical image segmentation remains challenging across modalities because weak object boundaries, scale-varying anatomical structures, and heterogeneous appearances often degrade feature representation and prediction accuracy. To address these challenges, we propose BMANet, a Boundary-aware Multi-scale Attention Network for accurate and efficient medical image segmentation. BMANet is built upon a lightweight encoder–decoder architecture and consists of four complementary modules. First, the Multi-scale Contextual Aggregation Module (MCAM) combines pyramid dilated convolutions with a large-kernel branch to capture hierarchical contextual information while reducing spatial information loss caused by repeated down-sampling. Second, the Edge Contour Enhanced Attention Module (ECEA) integrates low-level edge cues with high-level semantic representations to enhance edge-sensitive feature learning. Third, the Adaptive Channel-aware Spatial Attention Module (ACSA) dynamically calibrates channel and spatial responses, improving feature discrimination under varying contrast and texture conditions. Finally, the Prior Multi-scale Feature Refinement Decoder (PMFR) employs reverse attention and progressive feature refinement to emphasize hard-to-segment regions and recover fine spatial details. Extensive experiments on seven publicly available datasets covering polyp, nuclei, skin lesion, and cell segmentation demonstrate that BMANet achieves competitive or superior performance compared with state-of-the-art methods while maintaining favorable computational efficiency. These results indicate that BMANet provides an effective and practical framework for robust medical image segmentation across diverse imaging scenarios.