A Multi-View Clustering-Based Method for Individual And Group Cortical Parcellations with Resting-State FMRI

Cortical parcellation provides an important tool for revealing the organization of cerebral cortex. Despite the increasing number of attempts to developing parcellation algorithms using resting-state fMRI, generating reliable, functionally coherent brain parcels at both subject-level and group-level remains challenging due to the difficulty in balancing individual variability and group consistency without prior information. To overcome this challenge, we proposed to treat each subject as a view of population data and use multi-view clustering approach to learn individual and group parcellations. Specifically, it integrates spectral embedding and tensor learning into a unified optimization framework to optimize individual embedding matrices (for individual parcellation) and group consensus matrix (for group parcellation) jointly. In this process, an optimal balance between subject-specific and group-consensus parcellation can be achieved in an adaptive manner. Experiments on a test-retest dataset from Human Connectome Project showed that our method outperformed the existing state-of-the-art algorithms.