Improving Reliability of Subject-­Level Resting State Parcellation with Empirical Bayes

• Activity Number: 185
• Type: Contributed
• Date/Time: Monday, August 4, 2014 : 10:30 AM to 11:15 AM
• Sponsor: Section on Statistics in Imaging
• Author(s): Amanda Mejia*+ and Martin Lindquist and Brian Scott Caffo
• Companies: and Johns Hopkins Bloomberg School of Public Health and Johns Hopkins University
• Keywords: neuroimaging ; resting state functional connectivity ; shrinkage ; empirical Bayes

Abstract

Resting state networks (RSNs) are brain regions that are functionally connected. While originally limited to group-level RSNs, recently the focus has shifted to subject-level parcellation. However, to date the accuracy and reliability of these methods remains largely unreported. We propose a method to improve scan-rescan reliability of subject-level parcellation through empirical Bayes shrinkage towards the group average. Scan-rescan reliability was assessed using the publicly available Kirby21 dataset. We focus on a well-studied region of interest, the pre-central gyrus or M1. We compute the correlation matrix within this region for each subject, estimate the signal and noise variance at each cell, and compute the empirical Bayes shrinkage parameter as the ratio of signal variance to total variance. The resulting empirical Bayes estimates are then computed and employed in clustering for each subject. We find improved scan-rescan reliability of subject-level correlation matrices and parcellations, with mean squared error of the correlation matrices being reduced by 28.2% (95% CI: 18.9%, 37.5%) after shrinkage compared with raw estimates.