I am aware of using the --use-aroma flag, to run ICA AROMA. However, I’m curious whether there is an option to choose to run agg vs non-aggressive ICA aroma. If not, is there a way I could use the outputs from the --use-aroma flag to perform aggressive ICA aroma?
Yes that is possible. To perform aggressive denoising after running FMRIPREP you need to regress out the aroma_motion_* regressors found in _desc-confounds_regressors.tsv file from the _desc-preproc_bold.nii.gz file.
I just want to clarify something. If we’re using the aggressive denoising strategy on the _desc-preproc_bold.nii.gz data, do we also need to include the motion parameters (trans_x trans_y trans_z rot_x rot_y rot_z) in our nuisance regression, or will they already be regressed out by the ICA-AROMA confounds?
I realize that we would want to include these motion parameters in a nuisance regression if we were just using the _desc-preproc_bold.nii.gz data without ICA-AROMA denoising; however, does it make sense to include them with _desc-preproc_bold.nii.gz + aggressive ICA-AROMA confounds, (or even for analyzing the non-aggressive smoothAROMAnonaggr_bold.nii.gz data?).
As described in the original paper and fMRIPrep’s documentation, the “aggressive” mode of ICA-AROMA results in a series of confound regressors.
In principle, there is nothing theoretically incorrect in using the aggressive ICA-AROMA regressors with motion parameters. I would recommend you to check on the correlation matrix to see whether the motion parameters are adding any new information or they are basically captured by some other confounds. That means that the GLM will probably “favor” one over the set of highly correlated confounds to explain noise, and that does not have any further implication (unless you are interested in that fitting particularly).
You are right here. Those time-series have gone through denoising, so a second step adding motion regressors may actually re-induce motion artifacts in your signal.
Correlation with the 6 realignment parameters (trans_x trans_y trans_z rot_x rot_y rot_z) is an explicit criterion in the AROMA classifier; components that are more correlated with motion are more likely to be classified as nuisance by ICA-AROMA.* Furthermore, in the aggressive model, the nuisance components aren’t competing with signal-of-interest components to explain variance in the dataset; thus, all of the variance that can be explained by nuisance components should be removed in the regression step.
For the above reasons, there is likely to be substantial collinearity between nuisance components and realignment parameters; fitting realignment parameters in a separate model could be redundant or could remove additional temporal degrees of freedom from the data for little benefit. The original implementation of ICA-AROMA recommended supplementing the ICA-based denoising procedure with a confound model that incorporated mean signal from white matter and cerebrospinal fluid compartments (white_matter and csf in fmriprep), but didn’t include the 6 motion parameters. We’ve found, along with others, that adding the mean global signal into the model can also substantially improve denoising performance, but some reviewers might still find it controversial.
So, in short, in addition to Oscar’s warning about potential reintroduction of artefactual variance, there might be other reasons not to include the realignment parameters in the nuisance regression.
*Technically, the criterion uses a linear discriminant analysis on a 2D feature space including (i) the maximum temporal correlation with any of the motion parameters and (ii) spatial overlap with a mask indicating voxels at the edge of the brain. The details are in Fig 2A in the original Pruim article.
The original implementation of ICA-AROMA recommended supplementing the ICA-based denoising procedure with a confound model that incorporated mean signal from white matter and cerebrospinal fluid compartments ( white_matter and csf in fmriprep ), but didn’t include the 6 motion parameters. We’ve found, along with others, that adding the mean global signal into the model can also substantially improve denoising performance, but some reviewers might still find it controversial.
I’m assuming though that including the confounds white_matter ,csf, and global_signal requires recalculating them into the AROMA-processed MNI space?
Actually, after reading this post (#817), the impression I’m getting is that csf,white_matter,global_signal, (and perhaps the consine##) can be included in nuisance regression, even though they’re derived from the pre-AROMA denoising data
If you’re using aggressive denoising, it’s certainly valid to include the mean WM, CSF, and global time series (as well as discrete cosine frequencies) in the nuisance model alongside the components identified as noise by ICA-AROMA and perform the regression in a single step. In this case, the output time series is orthogonalised with respect to all regressors in the model.
The situation becomes more complicated for the case of nonaggressive denoising. In particular (as discussed in the issue that you linked):
any further regression steps will reduce the overall variance in the data, as measured by the residual sum of squares (less relevant here), and
Intuitively, we might expect that – in order to prevent reintroduction of such structured noise – we would want to re-extract the mean CSF, WM, and global time series after performing the nonaggressive denoising, since they will already be orthogonal to any sources of variance removed during the nonaggressive denoising step.
Interestingly, however, previous simulations indicate that using the signal extracted prior to ICA-AROMA denoising confers a slight benefit in recovering a known ground-truth signal. I can’t justify this result theoretically, and it defies intuition to some extent. I’m sorry that I don’t have a more conclusive answer for you, but this remains an active area of research at this time.
This thread and many others like it are super helpful! With respect to denoising, ICA-AROMA and FMRIPREP, we are now trying to shift to doing everything as part of a single regression in fsl_regfilt. But I’m still trying to wrap my head around a couple of issues related to aggressive vs. non-aggressive denoising.
With the non-aggressive denoising, we are basically using a matrix that contains both signal and noise, and we are saying that any shared variance between signal and noise will not be removed. But, if there is shared variance between columns that are denoted as “noise” (e.g., perhaps a CSF regressor is highly collinear with a white regressor and the motion parameters), then the shared variance here would be removed (because it is just noise anyway)? Is that correct?
Assuming I am on the right track with how I’m thinking about the non-aggressive denoising, I am curious how this would play out in the analyses. For example, if we have stimulus-correlated head motion, then it would probably not be removed in the case of non-aggressive denoising? Would this then bias our stats/inferences? It seems like the motion parameters would need to be included in the analysis to prevent this from happening, but would their inclusion lead to the reintroduction of noise that was removed earlier?