The current lack of trustworthy biomarkers for autism spectrum disorder (ASD), constrains the range of viable diagnostic strategies to the observation of a set of behavioural characteristics defined in the Diagnostic and Statistical Manual of Mental Disorders (DSM) V of the American Psychiatric Association: persistent deficits in social interaction and communication, repetitive or restrictive patterns of behaviour, among others. These characteristics are usually revealed by the Autism Diagnostic Interview-Revised (ADI-R), and the Autism Diagnostic Observation Schedule (ADOS).
Convolutional neural networks make it possible to find patterns in high dimensional data with a complex structure. In this work, we apply these capabilities to resting state functional magnetic resonance (R-fMRI) obtained from an international multisite database (ABIDE I), containing ASD individuals as well as controls. By training a particular type of networks termed graph convolutional neural networks (hence adapted to the domain at hand), we reach the state of the art in terms of performance for this dataset. An additional virtue of these models is that, once trained, one can use them to investigate which aspects of the image determine the model’s predictions. This method hence provides a new route to find ASD markers based on neuroimages to, not only assist in the diagnosis of this disorder, but also to better understand it.