Coronary Artery Disease (CAD) is one of the leading causes of death worldwide, and its accurate diagnosis relies on precise segmentation of Coronary Arteries (CA). Although Computed Tomography Coronary Angiography (CTCA) is the clinical gold standard, manual segmentation suffers from inefficiency and subjectivity. To address the challenges of low contrast and complex vascular structures in CA segmentation, we propose DPNet, a dual-perception fusion network with parallel branches: (1) The Edge Perception Branch (EPB) extracts fine-grained vessel details through 3D convolutions, effectively handling low-contrast boundaries, and (2) The Structural Perception Branch (SPB) captures long-range dependencies via self-attention mechanisms, enhancing segmentation robustness for intricate vascular topologies. We propose two novel components: (i) The Bilateral Feature Communication Module (BFCM) that enhances cross-scale feature interaction via channel attention, and (ii) The Structural Reconciliation Module (SRM) that bridges semantic gaps between EPB and SPB to correct the shape of CA. Extensive experiments on the ImageCAS dataset demonstrate that DPNet achieves a Dice Similarity Coefficient (DSC) of 0.831, outperforming the second-best method (TransUNet) by 2%, with a 41% reduction in Hausdorff Distance (HD). Validation on ASOCA further confirms the model’s generalizability. Ablation studies verify the significance of BFCM and SRM (p<0.05).