I’ve been using fmriprep with ica-aroma to analyze a multiband dataset from a 2 minute long continuous scan, and then applied a separate highpass filter. This procedure yielded an artifact, apparent when performing inter-subject correction of the full timecourses within each voxel. In the image below you can see unusual correlations in voxels that seem to include a lot of CSF, such as in sulci and ventricles (in addition to the expected higher correlations in auditory cortex). This artifact appears whether the filtering is applied using FSL or SPM.
However, when the inter-subject correlation is performed on the original fmriprep ica-aroma output (without the highpass), the artifacts disappear. Also no artifacts manifest when the entire preprocessing pipeline is performed in FSL (using MCFLIRT for motion correction), including the same highpass filter. In addition, no artifacts manifest when performing the same fmriprep+ica-aroma+highpass on a different, non-multiband dataset.
It seems to me like the artifacts are a result of the interaction between ica-aroma, the highpass filter, and multiband data. I read through most of this discussion, but I’m not sure how it will all translate into my specific use case. This is my first time using multiband and ica-aroma, so I might just be missing something obvious. Any leads would be appreciated!
Sorry, just to confirm: You’re not correcting for any of the calculated fMRIPrep confounds, right (e.g., aCompCor, etc) ?
That’s right, no further corrections.
OK, thanks for confirming. Not sure if @jdkent has seen anything like this !
Could you please write the precise outputs of fMRIPrep that you used to generate boxes 1 and 2?
It seems like the high-pass filtering is reintroducing some variance removed by ICA-AROMA. I would cautiously rename this issue to some variation of that statement, because judging by box 2, it seems that fMRIPrep+AROMA’s outputs look just fine.
Thanks for directing me here @emdupre!
I haven’t worked with multiband data yet, and I’m cautious about using ICA-AROMA (which was trained with non-multiband data) on a multiband dataset, but I agree with @oesteban’s interpretation so far and I would also like to see the precise outputs used to generate boxes 1 & 2.
Hi @oesteban and @jdkent!
To create both box 1 and 2 I used fmriprep’s output smoothAROMAnonaggr_bold. The only difference between them is that in box 2 the output also went through a high pass filter (using FSL, 100s cutoff).
I agree that the aroma output looks fine. The artifacts are unique to the combination of aroma with high-pass, but only in multiband data.
Do you happen to know of anyone who used aroma on multiband data successfully? If not, perhaps I should avoid it for now.
Thanks @moregev. At first sight, my impression was that Box 1 does not have any spatial smoothing and Box 2 does have. My hypothesis was that you were using the
smoothAROMAnonaggr_bold outputs for Box 2 and the regular
preproc output for Box 1.
Is there any ways in which the
preproc (unsmoothed outputs) have sneaked into your HP filtering?
@oesteban definitely not, I double checked the filename…