Aiding Alzheimer's Disease Diagnosis Using Graph Convolutional Networks Based on rs-fMRI Data

Abstract

Early detection, diagnosis and treatment of Alzheimer's Disease (AD) is currently the most effective way to delay the progression of the disease. Among various neuroimaging techniques for AD, resting-state magnetic resonance imaging (rs-fMRI) is able to measure neurological activity of brain non-invasively and without radiation, and has become an important tool for studying neurodegenerative diseases. The brain functional network established by rs-fMRI data offers the possibility to quantitatively elucidate brain changes associated with the spectrum of AD pathology. However, the processing method of such graph data is different from that of images, and well-performing convolutional neural networks (CNNs) cannot be directly applied to them, so the graph convolutional networks (GCNs) are required. At present, few studies attempt to extend the classical CNN architecture to the graph data domain to give full play to the feature learning and representation capability of generalized convolutional neural networks. Therefore, this article extends the excellent CNN framework to the analysis of brain networks and proposes a U-shaped hierarchical GCN framework (U-GCN), which includes down-sampling, up-sampling and skip connection operators for graph data, and node additional features based on graph theoretical analysis. The model evaluation is carried out on Alzheimer's Disease Neuroimaging Initiative (ADNI) database, and the proposed method demonstrates superior graph classification performance compared with the methods based on CNN and GCN. Furthermore, an experimental study on the network depth of U-GCN is conducted, and the results can provide a reference for the design of deep GCN model.