Structure-aware neural modeling of Circle of Willis vessels from noncontrast MRA data

Cerebral artery diseases have always been an important topic in medical research. In recent years, many studies have focused on vessel segmentation and completion, especially for the arteries around the circle of Willis. However, most existing methods are based on CTA images, which require contrast agents that may be harmful to the human body. In addition, these methods often result in incomplete topology or incorrect connections after vessel completion. In this paper, we use MRA data, which is safer and non-invasive, to complete missing vessel structures. We collected data from 125 patients, along with manually labeled vessel segments, focusing on 13 segments arteries in the circle of Willis. A neural network combining PointNet++ and Graph Convolutional Networks (GCN) is used to learn the structure and relationships of the vessels. Our experimental results show that the proposed method can effectively complete the vessel structures and reduce topological errors. This suggests that our approach is a promising alternative to CTA-based methods for vessel analysis.