Our group at UBC is getting acquainted with fMRIPrep. One thing we have noticed is that the number of components identified in ICA AROMA when it is implemented in fMRI prep is much smaller than when we run ICA AROMA via FSL. To unpack that, in fMRIPrep, we see about 25-30 components identified but in FSL it is more like 150-200. Any reason why they differ? The preprocessing is otherwise the same (e.g., same smoothing). Is it possible that fMRIPrep implements a thresholding so that less components are outputted?
Just to be clear, are these differing component numbers coming from the same data? The number of components returned will largely vary based on TR and scan length. fMRIPrep does limit the number of components it can return (which can be overridden), but this default is 200.
I am one of the members of the UBC group she mentioned.
*** CORRECTION TO PREVIOUS POST ***
I was mistaken, it was not the same fmri run/scan. In fact the one with a lower number of components and none lower than ~1.4% variance was actually a multi-echo sequence and the one with ~200 components was single echo. After reviewing the ICA output in the fmriprep HTML report from single-echo data form a different study I now see over 70 components and lower % variance components as well.
This begs a followup question though, does it make sense that multi-echo data would have lower dimensionality?
*** Original and incorrect post below ***
To clarify, yes the exact same fmri run/scan is producing a large difference in components via melodic in FSL natively, and that built into fmriprep. When dealing only in FSL, we follow the same preprocessing steps (motion correction, 6mm smoothing, no high pass filtering until after AROMA) but this difference persists.
We did notice that in FSL we get components with very small % variance, but in fmriprep we noticed no component has less than ~1.4% variance. We arent sure why, but out assumption is that the fmriprep implementation is somehow doing a component variance constraint or using a different threshold? We arent sure here since we are also not adjusting the 200 component constraint built into default fmriprep.
Thanks for your reply Steven. This is already helpful information and points us away from thresholding.
Per my colleague below, indeed we do not yet have the very critical apples-to-apples comparison of the exact same dataset run through ICA-Aroma via fMRIPrep versus FSL. Our observations were more qualitative, based on various datasets from our scanner, but as my colleague nicely points out, some of the fMRIPrep datasets we looked at were ME.
We will perform that apples-to-apples test and report back! We will try to do it for a multi-echo and single-echo dataset. I would not expect them to be completely identical across software programs but hopefully they are reasonably close. Time will tell.