I was wondering if anyone had any suggestions about best practices for using fmriprep + ICA-AROMA processed data for a PPI analysis?
I have already seen a few threads about using data and noise regressors generated by the fmriprep + ICA-AROMA pipeline but I haven’t found a definitive answer to my question. I think the main issue here relates to the CSF and WM regressors, which are particularly important for PPI. I understand that in fmriprep these are calculated pre-AROMA, hence using them on AROMA-processed data might re-introduce noise. On some of the forum discussions I’ve seen people suggesting re-extracting CSF and WM from the *space-MNI152NLin2009cAsym_desc-smoothAROMAnonaggr_bold.nii.gz file, however, since this image is already smoothed I believe extracting CSF and WM signals from it is not ideal.
I was considering following the suggestion presented here but I am not sure if that was suitable for a PPI analysis.
Actually, ICA-AROMA will not give you those regressors. Running fMRIPrep with default options will run a/t CompCor (anatomical and temporal CompCor). You’ll be interested in the aCompCor outputs in particular.
That may be a solution for people interested in having both (CompCor regressors) and ICA-AROMA denoising. And still, if you use ICA-AROMA I would discourage using the CompCor regressors as they can effectively reintroduce some of the noise variance.
I might be wrong (@jdkent and @rastko are invited to police my response), but I think your best option here is reusing the aCompCor regressors.
If you erode the WM/CSF masks more than the FWHM of the smoothing kernel (6mm), this should be a fairly safe operation in the sense of avoiding capturing gray matter responses. Unfortunately, there’s no good way to avoid smoothing before running ICA-AROMA, since the classifiers are trained on smoothed data.
If I use ICA-AROMA (*space-MNI152NLin2009cAsym_desc-smoothAROMAnonaggr_bold.nii.gz file), should I use the WM and CSF regressors instead of aCompCor components in the *confounds_regressors.tsv? But there are also suggestions that ICA-AROMA algorithm is trained to pick up motion artifacts specifically, while CompCor is trained to pick up physio-related noises (e.g., cardiac and respiratory related-noise). Therefore, if I use ICA-AROMA, I want to discourage using motion-related variables, and use the first 5 aComcor confounds in the post first- and second-level analysis.
If you use ICA-AROMA nonaggressively denoised images, I believe you should not use any other regressors or you’ll be risking re-introducing spurious signals. Is that correct, @rastko?
As Oscar said, when performing sequential regression steps there is at each step always the risk of reintroducing variance that was removed in previous steps. Notably, the CompCor decomposition is performed on the BOLD data before AROMA-based denoising and the resulting component time series could potentially include variance that has already been removed by ICA-AROMA. As shown by jdkent’s previous simulations using the mean WM and CSF time series, it’s possible that this might not pose a major problem, though.
While I accordingly can’t make a blanket recommendation against regressing CompCor on the nonaggressively denoised ICA-AROMA data, to my knowledge that isn’t (as of June 2019) an approach that is broadly used or whose efficacy has been validated. If you’re planning to use a denoising strategy that hasn’t been previously evaluated, I would strongly recommend reporting a measure of residual noise, like dataset QC-FC correlations.
I would also like to regress out WM and CSF from the nonaggressive denoised data. @oesteban, you wrote
And still, if you use ICA-AROMA I would discourage using the CompCor regressors as they can effectively reintroduce some of the noise variance.
Do you also discourage regressing out mean time series of WM and CSF from AROMA’s nonaggressive denoised data? Because I thought that was the original recommendation in the Pruim et al paper, wasn’t it? And also the first scenario suggested here.
means you recalculate the WM/CSF signal from the post-denoised data (fmriprep has the WM/CSF signal from the pre-denoised data). That is also the reasoning for @oesteban’s comment.
Does that make sense?
Sorry if that was not very clear before.
Thanks @jdkent, now I understand. But then there is no way to avoid ýour con, ie, recalculate WM/CSF post AROMA from a smoothed image, because AROMA works on smoothed images as @effigies pointed out in his comment above.
@mri you are correct that to identify the noise components (i.e., do the classification) using ICA-AROMA the input image must be smoothed. However, I believe you can apply the non-aggressive denoising procedure to the unsmoothed bold image, and then calculate the WM/CSF signal from the unsmoothed denoised bold image.
I believe @mmennes also says it is okay to denoise using unsmoothed data, but I’m pinging him here to make sure I’m not putting words in his mouth.
Thanks @jdkent, that was really very helpful!
I’m using fmriprep with the -use-aroma flag. So if I understood correctly, I can take the *space-T1w_desc-preproc_bold.nii and perform non-aggressive denoising in T1w space with fsl_regfilt as discussed here. Then I take the resulting *_space-T1w_desc-AROMAnonaggr_bold.nii.gz and calculate WM/CSV signal from it, is that it? @mmennes could you please confirm that it is indeed ok to denoise the unsmoothed data? I’ve seen many posts that had this same issue so I believe it would benefit others as well to know this. Thanks!
yes, you can ignore the auto-generated AROMA denoised file and use the component selection AROMA provides to denoise another dataset - in your case the unsmooth data. Note, that technically you would have to run the first part of dual regression using the smooth component maps against the unsmooth data to get corresponding timeseries from the unsmooth data. However, I think the impact of this will be minimal, unless you smoothed with an excessively large kernel (we advise 1.5x the voxel size).
There’s still something unclear to me. In the original paper, was the WM/CSF mean signal extracted before or after ICA-AROMA? If I’m reading the paper right, I think it was after, but in that case the WM/CSF mean signal was computed from smoothed functional data. Was that the case @mmennes?