I am trying to use some of the confounds generated by fMRIPrep to denoise “task” data within the FSL FEAT GLM, and I am wondering which columns to include as regressors.
I’ve followed along with this previous post so far, but wanted to ask for some further instruction/validation of methods.
I am planning to include the 6 major motion regressors (x,y,z trans & rot) and some combination of a/c/tCompCor regressors to account for physiological noise, but my question lies in which of those to include.
Based on Steven’s comments on the previous post, it is reasonable to include enough components to account for 50% variance. In evaluating the json sidecar file for the fmriprep confound tsv, the components are listed in order of ascending variance explanation. Would it be appropriate to just include the n-th component that explains 50% of the variance by itself for each masked area (combined, wm, csf, edge)? For example below, including just a_comp_cor_45 instead of some combination of a_comp_cor_00 through 44?
Additionally, I am including high-pass filtering in my FEAT setup, which means that I should omit the cosine factors from the input confound file, correct?
The study we are running involves interleaved TMS/fMRI and our “task” is a block design of 12 rounds of single pulses over 6 minutes.
Any suggestions or recommendations are welcome and appreciated.
aCompCor is most common. Given some evidence of resolvable WM bold signal, I think just using CSF mask is okay, but most common practice is CSF + WM.
Keep in mind you also want to account for temporal degrees of freedom. That is, if you do not have a lot of volumes, you don’t want a lot of regressors either. That is why you also sometimes see using top X number (e.g. top 5, top 10) components.
No, the listed variance explained is cumulative.
No, you should not do high-pass filtering in FEAT and instead use cosine regressors.
Hi Steven,
Thanks for the detailed explanation. I followed your recommendations and got much better signal contrast post-GLM than my previous tests that did not include these confounds.
I had an additional question for you about edge components - they are in the same tsv file as the other fmriprep confounds and my PI is interested in using them to reduce edge artifacts.
Would you recommend using a similar approach as above, i.e. only using the first 5-10 (out of 24 listed), or something else? Based on this paper (Patriat, 2017, https://doi.org/10.1016/j.neuroimage.2016.08.051), it seems like using a combination of edge components and the standard motion confounds is a good way to isolate and regress noise caused by motion. Our ROI is in deep subcortical grey matter, so we’re not particularly concerned about signal loss on the outer surface of the brain.
Let me know what you think. Thanks!
Jess
Again, this might depend on your temporal degrees of freedom. I can’t provide an exact recommendation without knowing about your data protocols and analyses goals. And even then there is no one-size-fits-all recommendation. If you have pilot data you can hold out and test various pipelines on, that would be ideal, so you don’t accidentally p-hack.
