Deep Learning for Neuroimaging: Advancing Brain-Based Biomarkers of Autism Spectrum Disorder

Abstract

In the last decade, deep learning (DL) has revolutionised fields like speech and vision through artificial neural networks. Can DL similarly transform biological psychiatry and neuroimaging? This thesis explores that question for Autism Spectrum Disorder (ASD) using MRI data. ASD involves a mosaic of social, communication, cognitive, and sensorimotor differences. Diagnosis relies on behavioural assessments by highly trained clinicians. This process is challenging and resource-intensive, often involving trial-and-error before optimal interventions are identified. MRI offers promise for improving diagnosis and care by revealing ASD's brain bases. But no reproducible biomarkers have emerged, likely reflecting ASD's heterogeneity, small sample sizes, unimodal data, and limitations of standard tools. Recent advances in multivariate predictive modelling could overcome these hurdles. In particular, DL methods from other fields now show potential for neuroimaging applications. Leveraging this opportunity, this thesis had three main aims: 1. Build a DL model for rapid, accurate quality control of structural MRI data, enabling analysis of huge datasets. 2. Predict ASD from raw structural MRI scans without standard template registration, preserving sensitivity to anatomical alterations. 3. Analyse functional MRI data with Transformer models that incorporate spatial and temporal patterns, working toward prediction from 4D data. These projects illustrated successful applications of DL in neuroimaging for ASD, while raising questions around generalisability across confounds like age and gender. Findings emphasise the continued need to refine preprocessing methods for atypical brains and quantify bias from procedural variations. Overall, this thesis advanced reproducible pipelines for potential MRI-based biomarkers of ASD. By openly sharing code and creating a novel tool, it enabled future DL applications in neuroimaging. Follow-up work on multi-modal prediction, optimal sample sizes, expanded categorical labels, and new DL architectures will further realise the promise of neuroimaging to improve psychiatric diagnosis and care.