Beware detrending: Optimal preprocessing pipeline for low-frequency fluctuation analysis

Resting-state functional magnetic resonance imaging (rs-fMRI) offers the possibility to assess brain function independent of explicit tasks and individual performance. This absence of explicit stimuli in rs-fMRI makes analyses more susceptible to nonneural signal fluctuations than task-based fMRI. Data preprocessing is a critical procedure to minimise contamination by artefacts related to motion and physiology. We herein investigate the effects of different preprocessing strategies on the amplitude of low-frequency fluctuations (ALFFs) and its fractional counterpart, fractional ALFF (fALFF). Sixteen artefact reduction schemes based on nuisance regression are applied to data from 82 subjects acquired at 1.5 T, 30 subjects at 3 T, and 23 subjects at 7 T, respectively. In addition, we examine test–retest variance and effects of bias correction. In total, 569 data sets are included in this study. Our results show that full artefact reduction reduced test–retest variance by up to 50%. Polynomial detrending of rs-fMRI data has a positive effect on group-level t-values for ALFF but, importantly, a negative effect for fALFF. We show that the normalisation process intrinsic to fALFF calculation causes the observed reduction and introduce a novel measure for low-frequency fluctuations denoted as high-frequency ALFF (hfALFF). We demonstrate that hfALFF values are not affected by the negative detrending effects seen in fALFF data. Still, highest grey matter (GM) group-level t-values were obtained for fALFF data without detrending, even when compared to an exploratory detrending approach based on autocorrelation measures. From our results, we recommend the use of full nuisance regression including polynomial detrending in ALFF data, but to refrain from using polynomial detrending in fALFF data. Such optimised preprocessing increases GM group-level t-values by up to 60%.