I know this has been asked many times but looking through all the discussions here and fmriprep’s github issues, I didn’t find a specific answer.
Let me cut through the noise: I want to use the ABIDE2 fMRI for functional connectivitity (voxel-wise brain). The workflow I have seen used previously (see especially Holiga et al):
Smoothing (6mm FWHM)
Bandpass filtering (0.01-0.1Hz)
After that the files are used to calculate different functional connectivity metrics (in my case I am interested in the voxelwise degree centrality). I know pre-processed my data using fmriprep and I used AROMA ICA. Would it make sense to replicate the following as:
as smoothAROMAnonaggr are already denoised and smoothed, can I just bandpass to the respective frequency and use them to calculate my functional connectivity metrics?
I’ll add, that if you end up wanting to use additional confounds, the way to go might be to take your non-AROMA desc-preproc BOLD output, and denoise it with the AROMA confounds and additional ones you want to add (e.g. motion, tissue, global signal, non-steadystate). That way you will not have two steps of denoising, ones that happen before and after smoothing. You can do all your cleaning before smoothing.
Thank you for the insight. I read the (Parkes et al 2018) paper and it seems interesting. Nevertheless, I think the best way for my current analysis would be to use the AROMA-ICA, which as far as I understand it’s pre-smoothed (so it doesn’t need any additional smoothing).
Can’t find any references to relevant papers using fmriprep with this workflow but I expect:
Would be enough to get the files for further connectivity analysis. Is there any obvious problem in that approach (what would reviewer 2 say)?
I suppose in the future I can check your second suggestion (i.e. using the desc-preproc BOLD and denoise using the AROMA cofounds) and compare the two.
Whether or not you smooth is dependent on your connectivity approach. You shouldn’t necessarily smooth for ROI-To-ROI, because the averaging within ROI parcels will serve the same purpose as Gaussian smoothing. Gaussian smoothing in this case would just lead to more signal contamination between nearby areas.
I am calculating whole brain voxel-wise connectivity, not ROI-to-ROI. According to previous work (Holiga et al, 2019) it seems like smoothing is desirable (although again different people have different opinions)