Graph Convolutional Neural Networks for Brain Connectivity Analysis

Abstract

We explore the potential and limitation of Graph Convolutional Neural Networks (GCNs) for classification of brain graphs derived from MRI measurements of subjects with Alzheimer’s disease (AD). GCNs differ from regular Artificial Neural Networks (ANNs) in that they operate directly on graph structures by defining convolutional operators in a non-Euclidean space. We show that GCNs perform well on graphstructured data, where regular ANNs typically fail due to the arbitrary ordering of nodes. Different GCN architectures are examined and compared to a Fully Connected Feedforward Neural Network. Tests are initially performed on simulated graphs mimicking the human brain. The simulated brain graphs were generated by following an algorithmic approach parameterized by graph measures known to be important for characterizing brain graphs. We demonstrate that GCNs are vital in accurately classifying the simulated brain graphs. The GCNs’ performance is evaluated on structured MRI-data, displaying cortical thicknesses for 68 regions in the brain of patients with Alzheimer’s disease and a healthy control group. On the structured brain data, both GCNs and regular ANNs are shown to be able classifiers. Finally, we show that the performance of regular ANNs is completely dependent on a fixed ordering imposed by the brain graph derivation from the MRI-data. In contrast, we show that GCNs perform well independent of node order.