Deep Vessel Segmentation By Learning Graphical Connectivity.

•A graph neural network (GNN) can learn global vascular structures in medical images.•A CNN only learns local appearances on the regular image grid and thus can be limited.•The vessel graph network (VGN) combines a GNN into a comprehensive CNN architecture.•The VGN jointly learns both local appearance and global structure of vessels.•The VGN structure can be applied to any existing CNN structure to improve accuracy.