Structural changes in autism reflect atypical brain network organization and phenotypical heterogeneity: a hybrid deep network approach.

Autism Spectrum Disorder (ASD) is a developmental disorder characterized by heterogeneity in social and emotional responses, language, and behavior. Assessments such as the Social Responsiveness Scale (SRS) can quantify this variability but understanding underlying mechanisms and identifying distinct and shared atypical organization and function of brain networks remains a challenge. Convolutional Neural Networks (CNNs) have been used to analyze imaging data. However, the relationship between structural brain changes observed in structural MRI (sMRI), the affected brain functional networks inferred from these structural changes, and their connection to ASD phenotypes and scores still requires systematic investigation. In this study, we ensembled 3D CNNs with other artificial intelligence (AI) methods to conduct a comprehensive analysis of macrostructural changes in ASD. We found consistently dominant involvement of (a) the left hemisphere, (b) the frontal and temporal lobe, and (c) the default mode, salience, and language networks in ASD. Our findings highlighted brain network similarities and differences between high and low severity ASD and showed that typically developed brains fall at the low-severity end of the high-to-low severity spectrum of ASD. Our systematic AI approach utilized the phenotypic heterogeneity and spectral nature of ASD to uncover significant structural changes across brain regions and functional networks, correlating the structural, functional, and phenotypical heterogeneity of individuals with ASD. This enabled us to identify known and novel global and local brain region and network changes in ASD in relation to phenotypes and clinical scores that can guide diagnostic subtyping.