<p>The circle of Willis (CoW) configuration varies widely among the population and serves as an important reference for assessing the risk of unruptured aneurysms and ischemic stroke. However, manual classification of CoW configurations demands significant expert knowledge and human resources, and there remains a lack of automated classification methods. To address this, an automated classification method for CoW vessel segment configurations based on vision graph neural network (ViG) is proposed, named LS-ViG3D. Through a light-stem module, the method reduces the number of network patches and adjusts the ViG hierarchical structure while adding multi-task prediction blocks, which enhances information aggregation capability and effectively extracts the topological features of CoW. To reduce training complexity, a prior knowledge-based CoW region of interest extraction algorithm is also proposed, which significantly reduces contrast image redundancy and makes end-to-end automated CoW classification possible. Test results show that this method achieves balanced accuracy rates of 87.63% and 89.91% for the posterior communicating artery configuration 3-class classification task and the posterior cerebral artery P1 segment configuration 2-class classification task, respectively. It also achieves macro-average precision of 89.00% and 90.36%, as well as macro-average F1 scores of 88.07% and 89.88%. The comprehensive performance of this method surpasses that of other common image classification methods and shows promise for supporting clinical applications.</p>