Physiological noise model comparison for resting-state fMRI at 7 T

Functional connectivity measurements underlying the study of resting-state networks using fMRI (rs-fMRI) are compromised by non-neuronal mechanisms producing correlated signal fluctuations across the brain, commonly referred to as physiological noise, and this problem is enhanced at ultra-high-field (7 T). Physiological noise correction can be achieved by estimating the respective sources based on simultaneously acquired cardiac and respiratory signals. Alternatively, image-based methods relying on regressing out the global fMRI signal fluctuations extracted from regions of no-interest, namely cerebral spinal fluid (CSF) and white matter (WM), have also been proposed. However, the extent to which CSF/WM models explain fMRI signal variance related with cardiac/respiratory sources remains unclear. In this paper, we estimated both cardiac/respiratory and CSF/WM models of physiological noise in whole-brain high-resolution rs-fMRI data collected at 7 T, and determined the signal variance explained by each model individually, as well as their shared variance. We found that noise sources estimated from cardiac/respiratory recordings and CSF/WM global signals are, to a great extent, complementary rather than overlapping. Critically, CSF/WM models were found highly sensitive to the mask used to extract the CSF trace, with conservative, ventricle-based masks yielding more accurate correction. In general, model performance varied significantly across brain regions.