Test-retest reliability of functional networks for evaluation of data-driven parcellation

Abstract

Brain parcellations play a key role in functional connectomics. A set of standard neuro-anatomical brain atlases are in common use in most studies. In addition, data-driven parcellations computed from fMRI data using a variety of clustering algorithms have also been used. Recent studies set out to determine the best parcellation in terms of quality and reliability have remained inconclusive without a clear winner. In this work, we investigated the utility of test-retest reliability of functional connectivity as an evaluation metric for comparing parcellations. Specifically, using data from the human connectome project, we compared a data-driven parcellation and a geometric parcellation using Intraclass Correlation Coefficient (ICC). We also investigated the impact of parcellation granularity on the test-retest reliability. We observed that the ICCs for geometric parcellation are better than those of a data-driven parcellation, suggesting that the FCs computed using regular parcels in the geometric atlases are more reliable than those computed using a data-driven parcellation.