The C-PAC documentation says -
“There are three main approaches to motion correction: volume realignment, using a general linear model to regress out motion-related artifacts (i.e. regression of motion parameters), and censoring of motion confounded time points (i.e. “scrubbing”)”
My doubt is - Among the 3 approaches mentioned, which one is the best? Why?
Do people use all three together? Why?
And most importantly, after I have aligned the brain volumes using, let’s say, fsl’s MCFLIRT, why do I need to regress out the motion parameters again?
you can check out this answer to a similar question:
A tool like MCFLIRT focuses on aligning each of the images so that each voxel’s index refers to the same part of the brain. But, the article linked in that reply (Friston, et al. 1996) details that another part of the issue with motion is a voxel’s history of excitement by the RF pulses. Without motion, the nuclei excited within a given voxel will have a chance to relax back to some consistent value. With motion, a voxel might become re-excited before completely relaxing, so the signal it gives will be dependent in some unwanted way on its history (i.e., in a way that doesn’t really depend on cognition/neurophysiology). That extra dependence can be mitigated by including the estimated movement parameters in the GLM as regressors.
Everything @psadil said (and that I said in the linked post). One thing we did not address is motion censoring.
I’m not sure if this is the way everybody thinks about this, but sufficiently large motion produces artifacts that are induced by the resampling implicit in motion correction, and will not be fully eliminated by regressing out the motion parameters. In this case, you avoid these time points “infecting” your signal estimates by (in some way) removing them from your GLM.
In modeling hemodynamic responses to discrete events (as opposed to, e.g. rsfMRI), the strategy I’ve used is to zero-out the design matrix for the time points in question, and add a nuisance regressor (a column that is all zeros, except for a one at the time point in question) for each “censored” time point. In this case, I used ArtifactDetect to find time points with >(0.5 voxel-width) frame-wise motion or >3 sd frame-wise intensity jumps.
So to answer the question: Yes, I would use all three. Motion-correction is a baseline. For motion regression, note that the Friston 1996 paper recommends a 24-parameter autoregressive model (the DPARSF write-up contains a succinct description). And finally, you’ll need to choose a censoring strategy appropriate for your analysis, if there is one. The one I described above is not necessarily relevant to all analyses.
OK.
So, just to make sure - I can create a short method that will read FD and DVARS from the tsv file (created bu fmriPrep) and remove regress out every scan that has high levels of each of them (one scan back, if I’m correct). Is that a reasonable practice?
I’m not sure I understand what process you’re describing. For fMRIPrep confounds, FD[i] is the displacement between volume i and volume i-1, and similarly with DVARS. I would censor specifically the volumes that show high motion relative to the predecessor, so simply thresholding FD or DVARS will get you the correct volumes to censor without shifting.
The Power et al. (2012) paper suggests removing volumes before and after a motion peak, and the evaluations of censoring have been done with that kind of procedure. You may want to adapt both the threshold on FD as well as the number of volumes being removed on the TR of the sequence. In particular the values suggested by the original Power paper will not scale to ultra-short TR (<1s) acquired using simultaneous multi-slice sequences.
As a quick note for task fMRI, the recommendation of Siegel, et al. (2014) (PMID:23861343), which explicitly compared the effects of different censoring strategies, was:
[…] we do not recommend removing volumes proceeding or following high-motion volumes, and we do not do so in the other data sets used in this article.
Hi, effigies, thank you for your wonderful explanation, I’m kind of new to fMRI, I also used your way of zero-out the design matrix for the timepoints with high head motion (FD>0.5), so if in this way, should i included FD in design matrix or not? I don’t think so, since nuisance regressors generated, they have some colinearlity with FD? greatly appreciated if you could answer this question!
Collinearity among nuisance regressors isn’t a problem, unless for some reason you want to attempt to interpret the betas. It’s collinearity between a variable of interest and any other variable that causes problems.
I would personally probably include the original motion regressors, not FD, which is just a weighted sum of the rotations and translations.