A paper was recently published in Human Brain Mapping that compared task-based denoising pipelines. They suggest efficacy of an “optimized” aCompCor method. Here is what the authors write
Compared to previous evaluation studies, we optimized aCompCor by extracting PCs from tissue signals orthogonalized with respect to the confounding variables that composed the model, yielding a set of PCs with a greater noise prediction power compared to the standard variant. The benefits of this optimization are illustrated in Figure S11 for the aCompCor50% variant of the CF dataset. Figure S11a–c shows QC-FC plots for the standard aCompCor50% approach and for two different optimizations obtained by preorthogonalizing WM and CSF signals with respect to either the sine/cosine basis functions (i.e., filtering the signals before computing PCs) or to the sine/cosine basis functions plus realignment-derived variables (the actual model used for pipeline comparisons). Both preorthogonalization schemes produced benefits compared to the standard approach, with the complete orthogonalization yielding the best results, particularly in shifting the QC-FC correlations toward zero. A second and critical benefit of such an optimization is specific to the aCompCor50% variant. When the optimization was used, we saw a marked reduction of the number of extracted PCs, that is, a reduction of the components required to fulfill the 50% variance criteria, as illustrated in Figure S11d. This result shows that the majority of the components extracted without the optimization explained variance that was already accounted for by the other regressors within the model, particularly by the sine/cosine basis functions.
Their MATLAB implementation is linked in this GitHub repo.
They did not use fMRIPrep in this paper. I was wondering if this is something that fMRIPrep does already. It looks like a promising way to preserve DOF.
Thank you,
Steven