Abstract

Graph Convolutional Networks (GCNs) are a novel and powerful method for dealing with non-Euclidean data, while Convolutional Neural Networks (CNNs) can learn features from Euclidean data such as images. In this work, we propose a novel method to combine CNNs with GCNs (CNN-GCN), that can consider both Euclidean and non-Euclidean features and can be trained end-to-end. We applied this method to separate the pulmonary vascular trees into arteries and veins (A/V). Chest CT scans were pre-processed by vessel segmentation and skeletonization, from which a graph was constructed: voxels on the skeletons resulting in a vertex set and their connections in an adjacency matrix. 3D patches centered around each vertex were extracted from the CT scans, oriented perpendicularly to the vessel. The proposed CNN-GCN classifier was trained and applied on the constructed vessel graphs, where each node is then labeled as artery or vein. The proposed method was trained and validated on data from one hospital (11 patient, 22 lungs), and tested on independent data from a different hospital (10 patients, 10 lungs). A baseline CNN method and human observer performance were used for comparison. The CNN-GCN method obtained a median accuracy of 0.773 (0.738) in the validation (test) set, compared to a median accuracy of 0.817 by the observers, and 0.727 (0.693) by the CNN. In conclusion, the proposed CNN-GCN method combines local image information with graph connectivity information, improving pulmonary A/V separation over a baseline CNN method, approaching the performance of human observers.