<p>Accurate segmentation of coronary arteries is essential for the diagnosis and treatment of cardiovascular diseases. However, the intricate vascular morphology, small caliber, and low-contrast boundaries of coronary arteries pose persistent challenges for conventional segmentation methods, leading to unstable predictions, particularly along vessel edges. To address these issues, we propose WEM-UNet, a novel multi-view 2.5D segmentation network that integrates wavelet-enhanced feature encoding and multi-scale attention mechanisms. Specifically, the model consists of three core modules: Wavelet Transform Convolution (WTConv), which suppresses high-frequency noise while preserving low-frequency structural details; the Enhanced Adaptive Boundary Awareness (EAGA) module, which refines boundary localization through the synergy of reverse and spatial–channel attention; and the Multi-scale Gated Aggregation (MGA) module, which strengthens contextual representation via dilated convolutions and adaptive gating. A tri-axial fusion strategy further aggregates predictions from orthogonal planes, ensuring spatial continuity and robustness. Extensive experiments on both private and public coronary CTA datasets demonstrate that WEM-UNet achieves superior accuracy and boundary fidelity compared to state-of-the-art methods. These results underscore its potential clinical value in automated coronary artery analysis and preoperative planning.</p>