This paper presents MDSVM-UNet, a novel two-stage framework for accurate coronary artery segmentation from CTA images. The proposed method integrates multi-view deformable convolution with Visual Mamba state space models to effectively capture long-range dependencies in vascular structures while maintaining linear computational complexity. This approach addresses challenges in segmenting complex multi-branching coronary arteries with severe class imbalance, enabling improved diagnosis and treatment planning for cardiovascular diseases. The synergy between deformable convolutions and Mamba architecture provides an efficient solution suitable for clinical deployment in resource-constrained environments.