Improving reliability of subject-level resting-state fMRI parcellation with shrinkage estimators.

摘要: Abstract A recent interest in resting state functional magnetic resonance imaging (rsfMRI) lies subdividing the human brain into anatomically and functionally distinct regions of interest. For example, parcellation is often a necessary step for defining network nodes used connectivity studies. While inference has traditionally been performed on group-level data, there growing parcellating single subject data. However, this difficult due to inherent low signal-to-noise ratio rsfMRI combined with typically short scan lengths. large number approaches employ clustering, which begins measure similarity or distance between voxels. The goal work improve reproducibility single-subject using shrinkage-based estimators such measures, allowing noisy subject-specific estimator “borrow strength” principled manner from larger population subjects. We present several empirical Bayes shrinkage outline methods when multiple scans are not available each subject. perform raw inter-voxel correlation estimates use both produce parcellations by performing clustering we standard spectral approach, our proposed method agnostic choice can be as pre-processing any algorithm. Using two datasets — simulated dataset where true known test–retest consisting 7-minute resting-state fMRI 20 subjects show that produced have higher reliability validity than those estimates. Application data shows increases motor cortex up 30%.