I have been trying to formalise the next steps of my analysis after fMRIprep but I am rather new to fMRI analysis so I would really appreciate any help.
My main questions are related to high-pass filtering, ICA-AROMA and using regressors to denoise data.
I know that ICA_AROMA denoises the data without applying a high pass filter to the data. If I were to apply a highpass filter to the fMRIprep output without the ICA_AROMA and then run ICA_AROMA on the filtered data, would the results be worse or better compared to the ICA_AROMA cleaned data produced automatically by fMRIprep? Would that influence the data in my case, where I have multiband data with 2000 slices (and thus a lot of components in ICA)?
I do not seem to understand exactly what output to use in a subsequent fMRI analysis. Would you use the ICA_AROMA denoised data and the motion regressors already calculated in a GLM? Would you altenatively need to do a high-pass filter of the ICA cleaned data, and would that introduce any noise in the data? Would using the motion regressors in combination with ICA introduce new noise in the system?
I believe you would want to perform the opposite (i.e. apply highpass filter to the ICA-AROMA denoised fMRIPrep data). See this post for more information.
I assume by “slices” you’re referring to volumes. With such large data, fMRIPrep provides an --aroma-melodic-dimensionality option where you can specify the number of components generated (e.g. --aroma-melodic-dimensionality 100. See this post for guidance.
If you’re using the ICA-AROMA denoised fMRIPrep output data then you don’t need to include the motion parameters; doing so may reintroduce noise.
I should also note that ICA-AROMA was not designed on multi-band data, so be aware of that. However, from what I’ve been on Neurostars and with my own data, ICA-AROMA works reasonably well on multi-band data.
Thanks a lot for this reply! Really usefull points I just had some clarifications/additional questions:
Essentially in this question I am trying to investigate how good is ICA by using the automatic clasiffier compared to me manualy selecting components (given that I have a multiband sequence). Would I need to high pass filter them before running the ICA in order to compare the data with the ones from fMRIprep (ICA_AROMA cleaned)?
Should the CSF, WM and Global Signal timeseries regressors be added to the ICA-aroma output? Would simply using the cleaned ICA-Aroma output, be enough for having good quality data in principle?
The ICA-AROMA denoised fMRIPrep data isn’t highpass filtered, so if you’re comparing against the fMRIprep output then you wouldn’t need to include highpass filtering on the other ICA data. Worth noting is that fMRIprep provides them in the ~_desc-confounds_timeseries.tsv file(s), under the Cosine_XX columns.
I personally include all three for additional nuisance regression; however, there’s some debate in the field regarding global signal regression, so it’s a bit dependent on what you’re examining. Regardless, it’s best to perform your regression simultaneously with whichever regressors you choose.
Thank you very much. I did indeed look through all the documetation I could find and there seems to be a lot of discussion on the subject.
In reality I am trying to formulate an analysis pipeline but I was wonrdering if there is a suggested one by fMRIprep (using ICA_AROMA on multiband data). My dataset is from an experiment that containts electrick shocks being delivered at a few times (which I do not need to include in the final analysis but might contain a bit of motion).
May I also ask some clarification on the latter point around using everything as regressors (WM,CSF,GS + cleaned ICA_data).
Would that constitute as an aggressive ICA_AROMA or not?
I am under the impression that aggressive ICA would contain the regressors of the ICA instead of the cleaned data but I am not very sure.
Honestly I really appreciate all your help and time.
fMRIprep is agnostic with regards to analysis steps following pre-processing, though the documentation will sometimes provide warnings regarding which confounds should or shouldn’t be incorporated in certain circumstances.There are some papers that evaluate analysis pipelines, though typically with regard to resting-state data. For example, see Parkes et al, 2018.
Not that I’m aware, ICA-AROMA only pertains to removal of noise from motion. The original paper that details ICA-AROM (Pruim et al., 2015) incorporates white matter, csf, and a linear trend in addition to their ICA-AROMA denoised data.