Is there a danger associated with using fMRIPrep GSR/WM/CSF regressors with fMRIPost-AROMA motion regressors?

, , ,
1

Background

XCP-D has a wonderful feature where you can use confounds from other derivatives like fMRIPost-AROMA as regressors.

However, this statement from Wang et al. 2024 is concerning:

An early version of this work did include an ICA-AROMA+GSR strategy, which performed very poorly … We strongly recommend fMRIPrep users to avoid fMRIPrep-generated GSR when using the ICA-AROMA strategy. It is also worth noting thatfMRIPrep will drop the support for ICA-AROMA from version 23.1.0

With the appendix further elaborating:

In fMRIPrep, the whole brain global signal regressor and the estimated head-motion parameters were calculated on the output from their regular pipeline (i.e., before denoising), which is inconsistent with the original proposal (Pruim, Mennes, van Rooij, et al., 2015) and not suitable for combining with ICA-AROMA. Based on the original implementation, the GSR should be calculated from the time series after removing the variance of ICA-AROMA regressors (Pruim, Mennes, van Rooij, et al., 2015). Otherwise, there is a possibility that the global signal regressor reintroduced motion to the data (Lindquist et al., 2019).

Question

So, I’m a little confused w.r.t two perspectives and would like your thoughts on which, if any, are valid for using relatively recent versions of these tools (fMRIPrep 24.1.1, fMRIPost-AROMA v0.1.dev1+g999b3f5, XCP-D: v0.10.1)

  1. This danger still holds since GSR, WM & CSF regressors are calculated during fMRIPrep, prior to any ICA-AROMA. XCP-D should not have listed aroma_gsr nor aroma as strategies since they use the confounds from fMRIPrep. Users should re-calculate GSR/WM/CSF from denoised outputs of fMRIPost-AROMA (or feature request?) and use them in XCP-D over those listed in fMRIPrep confounds.

  2. This danger is no longer a concern in current versions since it was contextualized during a time when ICA-AROMA was a part of fMRIPrep and the worry was running GSR/WM/CSF regressors on those partially-denoised outputs. Using the aroma_gsr or aroma strategy listed in XCP-D is fine.

2

I’d say the second one is accurate.

XCP-D will denoise the minimally preprocessed data in a single step, using the GSR, WM, and CSF regressors from fMRIPrep and the noise components from fMRIPost-AROMA. If XCP-D denoised the partially-denoised fMRIPost-AROMA outputs, then the problem you’re referring to would stand, but I don’t think it’s a concern with the way XCP-D works currently.