Less is more: balancing noise reduction and data retention in fMRI with projection scrubbing

Abstract

Functional MRI (fMRI) data are subject to artifacts arising from a myriad of sources, including subject head motion, respiration, heartbeat, scanner drift, and thermal noise. These artifacts cause deviations from common distributional assumptions, introduce spatial and temporal outliers, and reduce the signal-to-noise ratio of the data-all of which can have negative consequences on the accuracy and power of statistical analyses. Scrubbing is a technique for excluding fMRI volumes thought to be contaminated by artifacts. Motion scrubbing based on subject head motion derived measures, while popular, suffers from a number of drawbacks, among them the need to choose a threshold, high rates of scrubbing, consequent exclusion of subjects, and a lack of sensitivity to non-motion related artifacts. Data-driven scrubbing methods, which are instead based on observed noise in the processed fMRI timeseries, may avoid many of these issues and achieve higher sensitivity and specificity to artifacts. Here we present "projection scrubbing", a new data-driven scrubbing method based on a statistical outlier detection framework. Projection scrubbing consists of two main steps: projection of the data onto directions likely to represent artifacts, and quantitative comparison of each volume's association with artifactual directions to identify volumes exhibiting artifacts. Compared with DVARS, which is also data-driven, a primary advantage of projection scrubbing is its use of a common reference to identify abnormal volumes, rather than differences between subsequent volumes. We assess the ability of projection scrubbing to improve the reliability and predictiveness of functional connectivity (FC) compared with motion scrubbing and DVARS.We also compare projection methods, including independent component analysis (ICA), principal component analysis (PCA), and a novel fused PCA method. We perform scrubbing in conjunction with regression-based denoising through CompCor, which we found to outperform alternative methods. Projection scrubbing and DVARS were both substantially more beneficial to FC reliability than motion scrubbing, illustrating the advantage of data-driven measures over head motion-based measures for identifying contaminated volumes. ICA-based projection scrubbing produced the most