Learning Subject-Specific Functional Parcellations from Cortical Surface Measures

Cortical parcellations that are tailored to individual subjects have been shown to improve functional connectivity prediction of behavior and provide useful information about brain function and dysfunction. A hierarchical Bayesian (HB) model derived from resting-state fMRI (rsfMRI) is a state-of-the art tool for delineating individualized, spatially localized functional parcels. However, rs-fMRI acquisition is not routine in clinical practice and may not always be available. To overcome this issue, we hypothesize that functional parcellation may be inferred from more commonly acquired T1- and T2-weighted structural MRI scans, through cortical labeling with deep learning. Here, we investigate this hypothesis by employing spherical convolutional neural networks to infer individualized functional parcellation from structural MRI. We show that the proposed model can achieve comparable parcellation accuracy against rs-fMRI derived ground truth labels, with a mean Dice score of 0.74. We also showed that our individual-level parcellations improve areal functional homogeneity over widely used group parcellations. We envision the use of this framework for predicting the expected spatially contiguous areal labels when rs-fMRI is not available.