David Moodie Undergraduate Dissertation 2017/18

Diagnosing Autism using Deep Learning and Brain FMRI

Supervised by H.Lu

Abstract

The primary goal of this project was to use Deep Learning techniques to create a classifier which can diagnose autism spectrum disorder (ASD), based solely on a subject's brain activations. ASD is a neurodevelopmental disorder characterised by impaired social skills and restrictive, repetitive patterns of behaviour. ASD has a prevalence of approximately 1 in 68 children in the United States, according to data from the Centre for Disease Control. The current methods of diagnosing ASD are behaviour-based. These methods lack objectivity and can lead to inaccurate diagnoses due to ASD's heterogeneous nature. Previous work has been done using fMRI data for ASD diagnosis, however, most of these studies have focussed on a carefully selected subset of single-site data which results in classifiers that have poor generalisability and minimal clinical use. In this dissertation, a large multi-site brain-imaging dataset known as ABIDE, was used to create a classifier for ASD diagnosis using deep neural networks. The tuned SVM classifier, which was intended to be a baseline, performed better than the neural network, achieving an accuracy of 70% which is on par with the best known classifier produced by Heinsfeld et al. (2018). The results show how optimal hyperparameter tuning and use of a recently published functional connectivity measure can improve fMRI classification. The neural network classifiers underperformed, likely due to time and resource constraints.