Improved functional cortical parcellation using a neighborhood-information-embedded affinity matrix

Cortical parcellation of the human brain typically serves as a basis for higher-level analyses such as connectivity analysis and investigation of brain network properties. Inferences drawn from such analyses can be significantly confounded if the brain parcels are inaccurate. In this paper, we propose a novel affinity matrix structure based on multiple kernel density estimation for cortical parcellation. Neighborhood functional connectivity is embedded into the affinity matrix, which serves the dual purpose of allowing self-adaptive adjustment of voxel affinity values and providing robustness against noise. The proposed affinity matrix can be used with any parcellation method that takes an affinity matrix as its input. In our validation tests, we apply normalized cuts on our proposed affinity matrix to evaluate performance. On synthetic and real data, we demonstrate that the use of our proposed affinity matrix in lieu of the classical definition better delineates spatially contiguous parcels with higher test-retest reliability and improved inter-subject consistency.