Accurate Vascular Segmentation via Wavelet-Guided Edge Attention and Multi-Level Feature Fusion

Abstract

Accurate vascular segmentation plays a critical role in medical image analysis, supporting disease diagnosis, surgical navigation, and treatment planning. However, existing methods often struggle to preserve fine vascular structures and maintain topological continuity due to low contrast, complex backgrounds, and scale variations. To address these challenges, HWEA-UNet, a novel hybrid network for robust multi-scale vascular segmentation, is proposed. The framework integrates wavelet-guided feature decomposition with edge-aware attention to enhance structural integrity and boundary precision. Specifically, a hybrid encoder combines convolutional operations and MLP-based modules to capture both local details and global dependencies. A wavelet pooling strategy is introduced to preserve high-frequency information during downsampling, while an edge-focused encoding block and multi-level feature refinement modules further improve boundary representation and suppress background noise. Extensive experiments on coronary angiography, cerebral angiography, and retinal datasets demonstrate the effectiveness of the proposed method. On the coronary DSA dataset, HWEA-UNet achieves a Dice score of 0.8913 and an IoU of 0.8041 while requiring only 2.21M parameters. On the brain DSCA dataset, it achieves a Dice score of 0.8789 and reduces HD95 to 1.0000, outperforming existing methods in both segmentation accuracy and structural consistency. These results demonstrate that HWEA-UNet effectively preserves fine vessel structures and maintains vascular continuity across different imaging modalities.