An Improved MeshCNN with Active Learning for Anatomical Labeling of the Circle of Willis

The identification and labeling of the Circle of Willis in cerebral arteries is an important part of cerebral aneurysm treatment, which can assist physicians in making accurate diagnostic analysis. Most studies are based on manual extraction of vessel features, which obtain the low accuracy of labeling vessels with high topological variability in the Circle of Willis. In this paper, we propose a method combining transfer learning and active learning for automatic anatomical labeling of cerebral vessels. Firstly, the neural network model is transferred to automatically extract vessel features. Secondly, we use different sampling strategies of active learning to select some most valuable samples. The information of selected samples is fully utilized to improve the accuracy of machine learning modeling. Finally, in order to achieve automatic labeling of cerebral blood vessels, the downstream task model of transfer learning is integrated into active learning to construct a learning model. The experimental results present that the proposed method improves the labeling accuracy as much as possible while reducing the burden of manual labeling. It can effectively solve the Circle of Willis automatic anatomical labeling problem. In addition, the method can be better applied to vessels with high topological variability.