Many thanks for this really useful tool, I have already learned which denoising procedures are not effective on my dataset and now I’m trying to settle on the one that will be most appropriate. I wondered if you could answer a few questions about the implementation of these denoising pipelines?
- The confound_regressors file contains 18 HMP rather than 24, are they meant to contain the squared derivatives as well?
- 10 aCompCor components are used, but not the first 10 principal components from fmriprep. I was wondering how these are selected?
- What was the rationale for a DVARS threshold of 3%? I’ve seen this value vary quite a bit in the literature
Thanks a lot!
I’m glad that you like the tool! I’m aware that it’s still not perfect – we plan a large update in January.
It also definitely needs documentation and we are working on this.
You are right, there supposed to be 24 HMP. Now I see that squared derivatives are missing. This is definitely an error and we’ll correct this asap. Thanks for noticing this!
We took the first 5 aCompCor regressors obtained from white matter and CSF. FMRIPrep returns
~desc-confounds_regressors.json file which helps to decode which confound comes from combined WM & CSF, WM, or CSF masks. The first aCompCor regressors from the confounds table were calculated from the combined mask, that’s why we did not use them. For more details check out updated CompCor confounds description.
The number in DVARS threshold corresponds to +/-3 standard deviations. It’s a default in CONN toolbox and I saw this quite often in the literature. Let me know if you found some better recommendations!
Thanks for the quick reply! That makes perfect sense, as you note one of the pipelines is an exact replication of the CONN pipeline. I wasn’t actually aware fmriprep produced principal components separately for WM and CSF signal, so thanks for pointing that out!
Out of interest, do you plan to add any other pipelines in January? I’m particularly interested in the potential of wavelet despiking with ICA denosing and how other censoring and despiking artefact removal strategies compare to spike regression.
Good luck with the next update and Merry Christmas!
Yes, we plan to add more pipelines! You can open an Issue on fMRIDenoise GH and point us to some papers for reference.
We’ll also be happy to have your contribution in implementing this!