A Dual-Path Fusion Network with Edge Feature Enhancement for Medical Image Segmentation

This paper proposes a Dual-path Feature-enhanced Fusion Network (DPF-Net) for medical image segmentation to address limitations in existing methods, including insufficient edge feature extraction, semantic gaps among multi-scale encoder features, and significant semantic disparities between the encoder and decoder in U-Net architectures. To this end, we design a symmetric encoder–decoder structure based on U-Net and introduce three core modules: the Edge Feature Gating (EFG) module, which extracts and integrates edge features from shallow encoder layers to enhance edge integrity; the Cross-channel Fusion Transformer (CCFT) module, embedded in skip connections, to achieve comprehensive and semantically balanced multi-scale cross-fused features; and the Dual-path Feature Fusion (DPFM) module, which combines channel and spatial attention mechanisms to effectively bridge the semantic gap between encoder and decoder features while improving spatial resolution recovery accuracy. Experimental results demonstrate that DPF-Net achieves superior performance on six public datasets covering gland segmentation, colon polyp segmentation, and skin lesion segmentation tasks, significantly outperforming existing methods in terms of both mDice and IoU metrics. The conclusions confirm that the proposed method not only comprehensively improves the overall accuracy and edge segmentation quality of medical image segmentation but also enhances the model’s generalization capability across different tasks.