DMF-Net: A Dynamic Fusion Attention Mechanism-Based Model for Coronary Artery Segmentation

Coronary artery segmentation in CTA images remains challenging due to blurred vessel boundaries, unclear structural details, and sparse vascular distributions. To address these limitations, we propose DMF-Net (Dual-path Multi-scale Fusion Network), a novel multi-scale feature fusion architecture based on UNet++. The network incorporates three key innovations: First, a Dynamic Buffer–Bottleneck–Buffer Layer (DBBLayer) in shallow encoding stages enhances the extraction and preservation of fine vascular structures. Second, an Axial Local–global Hybrid Attention Module (ALHA) in deep encoding stages employs a dual-path mechanism to simultaneously capture vessel trajectories and small branches through integrated global and local pathways. Third, a 2.5D slice strategy improves trajectory capture by leveraging contextual information from adjacent slices. Additionally, a composite loss function combining Dice loss and binary cross-entropy jointly optimizes vascular connectivity and boundary precision. Validated on the ImageCAS dataset, DMF-Net achieves superior performance compared to state-of-the-art methods: 89.45% Dice Similarity Coefficient (DSC) (+3.67% vs. UNet++), 3.85 mm Hausdorff Distance (HD, 49.1% reduction), and 0.95 mm Average Surface Distance (ASD, 42.4% improvement). Subgroup analysis reveals particularly strong performance in clinically challenging scenarios. For small vessels (<2 mm diameter), DMF-Net achieves 85.23 ± 1.34% DSC versus 78.67 ± 1.89% for UNet++ (+6.56%, p < 0.001). At complex bifurcations, HD improves from 9.34 ± 2.15 mm to 4.67 ± 1.28 mm (50.0% reduction, p < 0.001). In low-contrast regions (HU difference < 100), boundary precision (ASD) improves from 2.15 ± 0.54 mm to 1.08 ± 0.32 mm (49.8% improvement, p < 0.001). All improvements are statistically significant (p < 0.001).