Using resting state functional MRI to build a personalized autism diagnosis system

Autism spectrum disorder (ASD) is a neuro-developmental disorder that affects social functionality and communication abilities of affected subjects. Studying brain functional connectivity is believed to be one of the trending techniques used in diagnosing and understanding ASD. In this work, resting state functional MRI data for 202 (78 autistic and 1245 typically developed) subjects are used to build a novel automated deep learning based autism diagnosis system. It uses power spectral density of time courses corresponding to the spatial activation areas as raw features to be fed to a stacked autoencoder followed by probabilistic support vector machine for diagnosis. The system achieved very high accuracy, sensitivity, and specificity of 88.5%, 85.1% and 90.4% respectively. The system proved to have very good generalization ability, especially among high-risk population. In addition to its ability to differentiate between autistic and healthy controls subjects, and the global differences found between the two groups, the proposed system generates a full personalized report for each individual subject that shows the impacted areas and to what extend they are impacted. From the clinical point of view, this report could be of great value, as it helps to predict and understand autistic subjects' behavior, in addition, it helps in developing a personalized treatment plan for each subject individually, which is an important step towards personalized medicine in autism, the ultimate goal of our group.