Use of Graph convolutional neural networks in fMRI data for characterization of psychiatric disorders.

Abstract

The period of childhood and adolescence is essential for comprehending brain function mechanisms, especially within multiple neuropsychiatric conditions. Once most psychiatric disorders begin at this phase, where the brain passes through intense changes that result in the consolidation of brain connectivity networks. In particular, the study of neurophysiology during the resting state (i.e., the state in which the brain does not consciously receive any internal or external stimulation) at the macroscopic level is of great interest, by its complex patterns that microscopic components cannot explain. Different techniques can be employed to measure brain function at the macroscopic scale, such as functional magnetic resonance imaging (fMRI). With fMRI, it is possible to study functional brain organization, which enables the search for predictive biomarkers for neurodevelopmental and neuropsychiatric disorders to elucidate their underlying mechanisms. Recently, neuroimaging studies have employed machine-learning techniques, that enable statistical inferences of neurophysiological characteristics of multiple disorders. In that context, this project aims to use a state-of-art machine-learning model called graph-convolutional neural networks (GCN), which operates on a non-euclidian domain to capture patterns in graphs that more classical models cannot. Providing a network-level analysis that captures topological information within brain networks from multiple neuropsychiatric disorders.