fMRI-based data-driven brain parcellation using independent component analysis

Background

Studies using functional magnetic resonance imaging (fMRI) broadly require a method of parcellating the brain into regions of interest (ROIs). Parcellations can be based on standardized brain anatomy, such as the Montreal Neurological Institute’s (MNI) 152 atlas, or an individual’s functional activity patterns, such as the Personode software.

New method

This work outlines and tests the independent component analysis (ICA)-based parcellation algorithm (IPA) when applied to a hypertension study (n = 48) that uses the independent components (ICs) output from group ICA (gICA) to build ROIs which are ideally spatially consistent and functionally homogeneous. After regression of ICs to all subjects, the IPA builds individualized parcellations while simultaneously obtaining a gICA-derived parcellation.

Results

ROI spatial consistency quantified by dice similarity coefficients (DSCs) show individualized parcellations exhibit mean DSCs of 0.69 ± 0.14. Functional homogeneity, calculated as mean Pearson correlation value of all voxels comprising a ROI, shows individualized parcellations with a mean of 0.30 ± 0.14 and gICA-derived parcellations’ mean of 0.38 ± 0.15.

Comparison with existing method(s)

Individualized Personode parcellations show decreased mean DSCs (0.43 ± 0.11) with the individualized parcellations, gICA-derived parcellations, and the MNI atlas having decreased homogeneity values of 0.28 ± 0.14, 0.31 ± 0.15, and 0.20 ± 0.11 respectively.

Conclusions

Results show that the IPA can more reliably define a ROI and does so with higher functional homogeneity. Given these findings, the IPA shows promise as a novel parcellation technique that could aid the analysis of fMRI data.