I am working on an analysis pipeline for a new study, and am intrigued by the edge CompCor regressors.
In my previous studies using fMRIPrep, I included CSF, WM, and 5 aCompCor regressors, as well as motion outlier volumes and motion regressors.
It’s implied by the abstract describing the new crown detection methods (https://osf.io/hz52v/) that it could make sense to use the edge_comp regressors instead of aCompCor regressors (and maybe also instead of WM/CSF?). But it doesn’t seem like any validation is available yet. Does anyone have advice on whether replacing aCompCor regressors with these new regressors actually improves analysis quality?
Edit to add: As I’ve thought about this issue more, I’m realizing that it may not make sense to use the edge components instead of the aCompCor regressors. But it may still make sense to add them as additional confound regressors. Has this been tested yet?
I would agree with this. My understanding is that the edge components are particularly useful for looking at noise from motion, as there should only be signal in the crown region if a head moves enough to put the brain there. On the other hand, aCompCor is primarily for physiological noise (and thus may be a better substitute for WM/CSF). Although, the preprint you link does talk about advantages of using edge components over WM due to partial volume effects.
I am also interested in how to best design first-level models with these (and hope that bumping this will get some more attention). In my data, I see between 20-25 edge components per run. So, in terms of degrees of freedom, it is about as costly as a 24 parameter head motion component expansion (i.e. 6 HMP + squared terms + derivatives).
I do not believe these regressors have been benchmarked yet. Definitely interested in seeing results of such benchmarking.