A self-supervised learning framework for discovering cortical folding patterns under genetic influence: Application to the Anterior Cingulate Cortex.

Cortical folding patterns can serve as a macroscopic probe for hidden events that occur during the development of the human brain. However, the large inter-individual variability of these patterns makes them very difficult to associate with developmental pathologies. We propose a framework for discovering candidate patterns under genetic influence and illustrate its use in the anterior cingulate cortex region, where the paracingulate sulcus shapes have aroused interest in psychiatry. This framework is based on a comprehensive regional representation of fold variability inferred from a self-supervised deep learning algorithm applied to 36,000 white British ancestry subjects from the UK Biobank database. This representation can be used to train linear classification models, in order to learn to discern folding patterns from labeled databases and generalize to larger databases. In our case, the generalization of paracingulate sulcus labeling on UK BioBank subjects did not lead to clear genetic associations. Secondly, we show that new loci associated with cortical folding patterns can be discovered directly from the representation. We find in the discovery cohort 4 loci for the right hemisphere and 10 loci for the left hemisphere related to specific folding patterns of the anterior cingulate cortex (