Physiological Noise Correction in Brainstem Imaging: An Empirical Evaluation of fMRI Data-Driven and Peripheral Physiological Recording-Driven Methods

Abstract

Physiological noise substantially affects fMRI data quality, particularly in areas near fluid-filled cavities and arteries such as the brainstem. Peripheral physiological signals can be recorded alongside fMRI and incorporated into general linear models to remove variance associated with cardiac and respiratory cycles using methods such as Retrospective Image Correction (RETROICOR). In contrast, data-driven noise reduction methods such as anatomical component correction (aCompCor) and independent component analysis based automatic removal of motion artifacts (ICA-AROMA) do not rely on additional peripheral physiological recordings. These methods have shown efficacy in correcting for motion, scanner as well as physiological artifacts. This raises the question of whether logistically demanding peripheral recordings offer added value. We therefore used nuisance regression methods based on peripherally recorded physiology (RETROICOR, heart rate, respiratory volume) as well as fMRI data itself (ICA-AROMA, aCompCor) to account for noise in a resting-state data set. Subsequently, we compared variance explained by the respective methods and improvements in temporal signal-to-noise ratio throughout different regions of interest in the brain. Consistent with prior research, RETROICOR, heart rate and respiratory volume explain significant variance throughout the brain with peaks around areas of strong cardiac pulsations. ICA-AROMA and aCompCor account for a substantial proportion of the variance explained by the methods using peripheral physiology. Nonetheless, the methods based on peripheral physiological recordings retain unique explanatory power. Analysis revealed a pattern of unreliability of ICA-AROMA to consistently identify and remove physiological noise across recordings in each participant, which is compensated by RETROICOR, heart rate and respiratory volume. This unreliability partially results from misclassifications in the noise selection models of ICA-AROMA. Our results suggest that it is advisable to additionally apply cleaning based on peripheral physiological recordings, especially when assessing inter-individual differences and areas with regionally high levels of physiological noise.