Efficient Boundary-Aware Multi-Scale Feature Fusion for Medical Image Segmentation

Accurate medical image segmentation plays a crucial role in clinical diagnosis and treatment planning. Yet, existing U-Net variants often fail to effectively capture both fine anatomical details and global contextual relationships while maintaining efficiency. To overcome these challenges, we propose EB-MSF-UNet, an Efficient Boundary-aware Multi-Scale Feature Fusion U-Net that integrates global reasoning with local detail preservation. The framework employs a dual-pathway encoder combining lightweight convolutional layers for local feature extraction and Mamba-inspired modules for longrange dependency modeling, regulated by an adaptive feature gating mechanism. A cross-attentionbased multi-scale fusion module ensures consistent interaction across feature hierarchies, while a boundary-aware refinement decoder explicitly enhances structural contours through auxiliary boundary supervision. Experiments on multiple medical image benchmarks demonstrate that the proposed method achieves consistently accurate segmentation with improved boundary precision and computational efficiency.