Abstract
Resting-state functional magnetic resonance imaging (rsfMRI) is widely used to study brain-wide functional connectivity (FC). However, the resulting signals are highly noise sensitive, and the best strategies for mitigating this noise remains unclear. In 358 healthy individuals, we compared 60 multi-echo (ME) and 30 single-echo (SE) rsfMRI preprocessing pipelines across six measures of data quality and associated effect sizes in FC-based prediction models of personality and cognition (cross-validated kernel ridge regression). ME pipelines generally outperformed SE pipelines, but no single pipeline excelled at both denoising and behavioral prediction. Using a heuristic scheme to rank pipelines across benchmarks, ME optimum combination (OC) with ME independent component analysis (ICA), ICA-FMRIB's ICA-based Xnoiseifier (FIX), and with head motion, cerebrospinal fluid, and white matter and gray matter signal regression, performed best when only considering denoising efficacy metrics. ME OC with ICA-FIX and head motion parameter regression performed best when only considering behavioral prediction results. ME OC with Automatic Removal of Motion Artifacts (AROMA) ICA, head motion parameter regression and Regressor Interpolation at Progressive Time Delays (RIPTiDe) performed best when aggregating across all evaluation metrics. These results favor ME acquisitions but show that no single denoising pipeline should be considered optimal for all purposes.