Talk title: Individual variation in neuroanatomy within autism
Abstract: Autism spectrum disorder is very heterogeneous. Diagnosed individuals vary in terms of genetics, neural measures, and behavior. Heterogeneity has two important implications for autism research. First, it makes it more difficult to identify findings that replicate, hindering the progress of research. Second, different individuals are likely to need different, personalized interventions. Therefore, understanding individual variation within autism is a key research objective. A major obstacle to this objective is the fact that individual variation related to autism is mixed with unrelated variation, which also occurs in the general population. We find that a recently developed deep learning method -- Contrastive Variational Autoencoders -- makes it possible to disentangle autism-related individual variation in brain anatomy from variation that is shared with the general population. In turn, this process reveals previously hidden relationships between individual differences in brain anatomy and individual differences in behavioral symptoms. Finally, we use counterfactual AI to identify interpretable anatomical sites of individual variation within autism, associating differences in behavioral symptoms with changes at specific anatomical loci. This approach can also be applied to other data modalities and other disorders, offering a technique to study individual variation specific to a particular population of participants.