Testing different noise models output from fmriprep - unclear results with ICA-AROMA

ica-aroma
fmriprep
afni
3dtproject

#21

Thank you again @fMRIstats for sharing your insightful paper. I have re-run a couple versions of my pipeline to test the effects of sequential (regression of 6 motion parameters, followed by high-pass filter) versus simultaneous (regression of 6 motion parameters and stop-band regressors) de-noising procedures.

Sequential motion regression–> temporal high-pass .009 Hz:

Simultaneous motion+stopband (.009 Hz) regression:

As evidenced by the plots, regardless of the sequential vs. simultaneous method taken, the association (-log[p]) between Functional Connectivity strength and Frame-wise Displacement is greater when the 6 motion regressors are included (labeled “06P”) versus when they are not (“0P”). I do not point this out to disagree with the arguments made in your paper, but rather I don’t think this to be the bases for the anomalous findings in my data.

I wonder if spatial smoothing (and/or ROI time-series extraction, another form of spatial smoothing) following motion/stop-band regression might introduce noise? E.g., although the time-series of one voxel might be orthogonal to motion parameters, might averaging it with its neighbor reintroduce some association with motion? Is the “best bet” to smooth before motion regression? Or rather to conduct the motion regression on the extracted ROI time-series? I am working on a way to test my speculations, but if anyone has any insight into this question, it would be greatly appreciated!


#22

Just a side note: you can limit the number of ICA components in fMRIPrep using --aroma-melodic-dimensionality 150


#23

Interesting! Thanks for checking. I’ll have to think more about this…


#24

Thank you @Michael_Hallquist for the detailed explanation on the non-aggressive (vs. aggressive) regression filtering processes conducted in AROMA. True, the 3dTproject program does not take into account the “signal” components in the regression model and only regresses out “noise” components.

Nonetheless, @jforawe’s above point (posted on 2018-08-30) about functional connectivity from the “06P” model (6 motion parameters as “noise”) having greater association with mean Framewise Displacement than the “base” model (i.e., first 5 volumes removed; 0th, 1st, and 2nd order detrend, filter) is the most puzzling. So, part of my confusion is due to issues independent of AROMA. IF the voxel-wise BOLD time-series are indeed made orthogonal to the 6 motion time-series, wouldn’t one expect functional connectivity (inter-voxel correlation) to have a reduced association with Frame-wise displacement (a linear combination of the 6 motion parameters)? (Note: I’m not directing this question at you specifically, just reiterating the confusing aspect :slight_smile: ). I found a related 3dTproject discussion on the AFNI discussion boards from fall 2016, but I’m not sure it’s the same issue I’m seeing. I also re-ran a highly similar preprocessing pipeline, where detrending was done (AFNI’s 3ddetrend), followed by regression of 6 motion parameters using FSL’s fsl_regfilt, followed by smoothing (AFNI’s Merge), and get basically the same result.

  1. 0P, no regression conducted, “base” model
  2. 06P, regression in AFNI
  3. 06pFSL, regression in FSL

As I posted earlier today, I wonder if part of the issue stems from the fact that, at least with the “6P” data, spatial smoothing (and ROI time-series extraction) occurs after the noise-regression (I’m not sure when its done in the AROMA sequence), and I am curious if spatial smoothing among voxels might reintroduce some of the motion associations? I have no evidence of this now, but hope to in time.