I am using AFNI’s 3dTproject in nipype to bandpass rsfMRI data between 0.01 and 0.1 Hz and simultaneously perform confound regression. However, it calculates 171 stopband regressors, leaving few DoF after regression of 24 head motion parameters, 10 aCompCor components, and spikes from 266 time-points (as in the standard Conn pipeline). This means I have a maximum of 58 DoFs, a number of cases that receive warnings for too few DoFs remaining (<30 and <20), and some cases result in errors where the total number of regressors exceed 266.
I was wondering if this is specific to temporal filtering in AFNI and whether there are alternative filtering methods to mitigate this problem? If this is an inherent problem with filtering are there any recommended guidelines for minimum DoF?
This is not specific to AFNI. Band passing uses up many DoF, though in some cases people do it separately (somewhat incorrectly) and do not realize the cost (in terms of DoF).
If TR=2s, then the highest contained frequency (Nyquist) is 0.25 Hz. When you apply an upper limit of 0.1 Hz and throw out all of 0.1 through 0.25, that is throwing out 60% of the DoF right there (e.g. ~160 out of 266).
As TRs get shorter, the cost (fraction of total DoF lost) will go up.
Thanks for the explanation, I think you may be right that many may bandpass without considering the effect on DoF. It is rarely mentioned in papers and I’m not sure if other softwares will provide these helpful warning and error messages, especially when running sequentially as you say (e.g. nilearn clean_img).
I wonder if ICA-AROMA is a suitable replacement for low-pass filtering as it automatically identifies components with high frequency content as noise? They don’t use a low pass filter for this reason (Pruim et al., 2015), and the number of regressors is far far fewer than those required for the low-pass filter (e.g. Ciric et al., 2017).
I found that ICA-AROMA didn’t effectively mitigate motion artefacts in my dataset, but perhaps in combination with censoring this could be more effective.
As the TRs get shorter, it seems likely that component regression will become almost necessary. The cost of bandpassing may already have it at the limit of usability (if TR <= 1.0 s, 80% of DoF are gone off the top, which is probably too much to be useful anymore).
Unfortunately, component regression is dangerous, too. The more individual components are polluted with BOLD signals (which most are, to some degree), the more distorted the results can be (by moving “noise” and “signal” to different parts of the brain, before regressions or correlations are computed). But the advantage in removing some troublesome noise components with not-so-many regressors is pretty big.
I am not qualified to say much about ICA-AROMA in particular, but motion is the bane of FMRI. We do our best to mitigate the problem, but do not expect any method to do a great job with it. The effects are too varied and hard to predict.
Yes, I have planned to apply ICA to censored data, but have not gotten to it. It does not seem effective to have the ICA method deal with motion spikes in time points that we already “know” are bad. It might be preferable to perform the ICA without those time points included, and then to properly account for time continuity outside of that operation. This is on “the list”…
Thanks for the interesting discussion here!
Just to add to your ideas about censoring and ICA denoising: some people have been using wavelet despiking (BrainWavelet or SpeedyPP) in combination with ICA, which sounds potentially very useful. I’ve yet to get around to it, but I hope to try it in the near future.
That’s interesting, @JoffJones, thanks!