Vessel Segmentation via Link Prediction of Graph Neural Networks

The topology of the segmented vessels is essential to evaluate a vessel segmentation approach. However, most popular convolutional neural network (CNN) models, such as U-Net and its variants, pay minimal attention to the topology of vessels. This paper proposes integrating graph neural networks (GNN) and classic CNN to enhance the model performance on the vessel topology. Specifically, we first use a U-Net as our base model. Then, to form the underlying graph in GNN, we sample the corners on the skeleton of the labeled vessels as the graph nodes and use the semantic information from the base U-Net as the node features, which construct the graph edges. Furthermore, we extend the previously reported graphical connectivity constraint module (GCCM) to predict the links between different nodes to maintain the vessel topology. Experiments on DRIVE and 1092 digital subtraction angiography (DSA) images of coronary arteries dataset show that our method has achieved comparable results with the current state-of-the-art methods on classic Dice and centerline-Dice.