Vessel-MAYON: A 3D direction-aware deep network for intracranial vessel segmentation from MRA Images

In magnetic resonance angiography, intracranial vessels exhibit complex tree-like structures with strong anisotropy. Accurate segmentation remains challenging because of discontinuous signals in small vessels and weak contrast in low-signal regions. To overcome these limitations, we propose the 3D Vessel Multi-Directional Attentional Yoked Optimization Network (Vessel-MAYON). The method incorporates two key components. First, the 3D Multi-Directional Core unit captures anisotropic vascular patterns by combining directional convolutions from coronal, sagittal, and axial planes with isotropic cubic convolutions. Second, the Vessel Yoked Attentional Fusion module enhances feature integration through learnable channel weighting, spatial attention, and adaptive residual fusion. Evaluations on the MIDAS dataset and the Cerebral Artery Segmentation Challenge 2023 show Dice Similarity Coefficients of 0.6981 and 0.8476, respectively, outperforming leading approaches in both accuracy and continuity of vascular reconstruction. These results demonstrate that Vessel-MAYON provides reliable, high-precision vessel segmentation and supports clinical application on GPU-equipped workstations. The implementation is publicly available at https://github.com/Brainsmatics/Vessel-MAYON.