Regional supervised learning of inhibitory control strength from cortical sulci

Joël Chavas, Aymeric Gaudin, Denis Rivière, Jean-Francois Mangin

Keywords: deep learning, cortical sulci, folding patterns, inhibitory control

Abstract:

The human cortical brain is folded and is highly variable among individuals.
We ultimately want to quantify how cortical folding relates to clinically relevant parameters. Here, using supervised convolutional networks on the human connectome project (HCP) dataset, we learn to predict inhibitory control strength, using all the information from the cortical folds. As we want to focus on the shape of the folds (which are supposed to remain identical throughout adult life), we don't use the full MRI images but the cortical skeletons, which are negative casts of the brain. As we expect putative folding patterns to be local, we scatter the supervised learning over 24 sulcal bilateral regions on the two hemispheres and apply an ensemble method to each region. We found the strongest significant correlations between inhibitory control and cortical sulci in the frontal marginal, the central sulcal, and the cingulate regions.

Submission Number: 127