Multi-relation graph convolutional network for Alzheimer’s disease diagnosis using structural MRI

Abstract

Structural magnetic resonance imaging (sMRI) is widely applied in Alzheimer’s disease (AD) diagnosis tasks by reflecting structural anomalies of the brain. Currently, most existing methods solely focus on pathological changes in disease-affected brain regions and ignore their potential associations and interactions, which provide valuable information for brain investigation. Meanwhile, how to construct effective structural brain graphs composed of nodes and edges remains appealing. To tackle these issues, in this paper, we propose a novel multi-relation reasoning network (MRN) to learn multi-relation-aware representations of brain regions in sMRI data for AD diagnosis, including spatial correlations and topological information. We frame distinguishing different disease statuses as the graph classification problem. Each scan is regarded as a graph, where nodes represent brain regions with abnormal changes selected by group-wise comparison, and edges denote semantic or spatial relations between them. Specifically, the dilated convolution module learns informative features to provide discriminative node representations for constructing brain graphs. Multi-type inter-region relations are then captured by the local reasoning module based on the graph convolutional network to provide a reliable basis for AD diagnosis, including geometric correlations and semantic interactions. Moreover, global reasoning is employed on the learned graph structure to achieve information aggregation and gradually generate the subject-level representation for AD diagnosis. We evaluate the effectiveness of our proposed method on the ADNI dataset, and extensive experiments demonstrate that our MRN achieves competitive performance for multiple AD-related classification tasks, compared to several state-of-the-art methods.