Individualized Cortical Parcellation Using Masked Semi-supervised Graph Neural Network

Brain parcellation is an important tool for understanding human cognition and function. Group-level parcellation reveals cross-subject stable properties while individual-level parcellation focuses on precise individual differences. Based on research on individual differences, we proposed an iteration scheme to partition stable and unstable regions in cortex, and about 65% of the stable vertices were used to make individualized mask. Our individualized cortical parcellation network is an algorithm to parcellate brain cortex using semi-supervised graph neural network with mask as a personalized criterion for semi-supervision. Resting-state functional magnetic resonance imaging (rs-fMRI) and their connectivity properties provide effective features for brain parcellation. Our parcellation conforms closely to the topological and functional characteristics of brain cortex. First, the individual distribution of brain parcellation converges to the group distribution. Second, within-subject test-retest reliability is significantly higher than that of between-subject. This study provides a new and reliable way for the application of deep learning in individualized brain parcellation.