Benchmarking fMRI Denoising Pipelines

Functional magnetic resonance imaging (fMRI) is a powerful tool for probing neuronal activity in vivo, but fMRI data are inherently noisy. To mitigate this, a wide range of denoising strategies have been developed, including volume censoring, anatomical component-based noise correction (aCompCor), ICA-based methods (e.g., AROMA, FIX), and multi-echo approaches (e.g., ME-ICA, tedana). These techniques are often applied in different combinations and have been predominantly evaluated on single-echo resting-state fMRI data-typically without incorporating more recent methodological advances known to improve modeling, such as order-independent "1-step regression", modeling temporal autocorrelation (pre-whitening), and temporal shifting of physiological nuisance regressors. To fill this gap, we used a framework that incorporates these methods and benchmarked a range of denoising pipelines across task and resting-state, single- and multi-band, and single- and multi-echo fMRI datasets, using different combinations of standard denoising confounds. Pipeline performance was evaluated using temporal signal-to-noise ratio (tSNR) and percentage remaining degrees-of-freedom (DoF), effectiveness of motion correction, and effectiveness of signal preservation. While pipelines only using ICA were insufficient, those that incorporated physiological nuisance regressors performed well. Additional improvements were observed when temporally shifted physiological regressors were accounted for. Based on these results, we provide recommendations for selecting denoising pipelines and emphasize the need for continued benchmarking as new methods are developed or applied in novel contexts.