Fmriprep: resampling of denoised BOLDs & SUSAN smoothing options

[Elmsch]

Hi everyone,

When running ICA-AROMA within fmriprep 1.2.4, BOLDs are automatically resampled. However, I would like to keep resampling to a minimum (and be able to use the brain masks provided by fmriprep). So I thought of manually smoothing my preprocessed BOLDs (MNI) using SUSAN and then removing fmriprep-derived components using fsl_regfilt.

Can anyone direct me to fmriprep’s implementation of SUSAN? Looking at these defaults (https://rdrr.io/cran/fslr/man/susan.html), I am not sure whether fmriprep used the very same ones (except for sigma, obviously)…

Also: Does anyone know why/at which step BOLDs are being resampled in ICA-AROMA?

Many thanks for your help!

Best,
Elmsch

[effigies]

fMRIPrep uses the nipype wrapper for SUSAN.

The full workflow is here:

poldracklab/fmriprep/blob/624baf34c70d475534642ea774b4e51e7332147d/fmriprep/workflows/bold/confounds.py#L521-L758

BOLD series are resampled to MNI152NLin6Asym before being passed to that workflow. If you’re interested, I can pull out the relevant bits.

[Elmsch]

Many thanks for your help, @effigies!

So fmriprep does not explicitly pass any brightness threshold to the nipype wrapper but must be using some kind of default, right?

Just a quick follow-up question: After resampling of my BOLDs to MNI space, they have a shape of 78 x 93 x 78 x time (MNI152NLin2009cAsym_desc-preproc_bold). When inspecting my BOLDs after running ICA-AROMA, they have a shape of 91 x 109 x 91 x time (MNI152NLin2009cAsym_desc-smoothAROMAnonaggr_bold). So there must happen some additional resampling within the AROMA workflow in fmriprep. Do you know where and why this happens?

Best,
Elmsch

[effigies]

So fmriprep does not explicitly pass any brightness threshold to the nipype wrapper but must be using some kind of default, right

It calculates the median value within a mask and thresholds at 0.75x the median value.

Just a quick follow-up question: After resampling of my BOLDs to MNI space, they have a shape of 78 x 93 x 78 x time ( MNI152NLin2009cAsym_desc-preproc_bold ). When inspecting my BOLDs after running ICA-AROMA, they have a shape of 91 x 109 x 91 x time ( MNI152NLin2009cAsym_desc-smoothAROMAnonaggr_bold ). So there must happen some additional resampling within the AROMA workflow in fmriprep. Do you know where and why this happens?

Unless things have changed I’m not aware of, AROMA outputs are only provided in MNI152NLin6Asym. Each template is likely to have slightly different dimensions.

[ak_poorni]

I would like to clarify, on the inputs to the ICA-AROMA pipeline, as I want to do some additional nonaggr denoising on the preprocessed data, but would like to match the data that are fed into ICA-AROMA, and hence match the temporal courses in the melodic mixing matrix.

Is ICA-AROMA applied to preprocessed data (after smoothing) in native space or MNI space? If I want to do additional denoising (using aroma components), then what are steps that I need to follow?

My take on was

1. Remove non-steady volumes from *space-T1w_desc-preproc_bold
2. Smooth using susan 6mm FWHM
3. denoise using fslregfilt
4. Transform to MNI space
   However, I got confused when I saw in the script here that [ICA-AROMA, @aroma] was performed on the preprocessed BOLD on MNI space (line 629)
   (https://github.com/poldracklab/fmriprep/blob/624baf34c70d475534642ea774b4e51e7332147d/fmriprep/workflows/bold/confounds.py#L521-L758)

Any advice on this would be appreciated.

Thanks
Poornima

[ak_poorni]

@oesteban @effigies I am just wondering if you can confirm the above, specifically if ICA AROMA was applied to BOLD series in native or MNI space?

Thanks very much

[jdkent]

Hi @ak_poorni,

ICA AROMA is applied to the BOLD series in MNI152NLin6Asym (MNI) space.

I do not believe applying the AROMA components from MNI space to “most” other spaces should impact the results significantly (unless the transformation from one space to another is extreme and/or poor).

@dvsmith linked a superthread about different ICA-AROMA conversations, you can see if you find some useful suggestions there too:

Aggressive vs nonaggressive denoising

Thanks, Martin! I think we’re just going to try and make one confound matrix (including the cosineXX regressors for hp filtering) during the first-level analyses in FSL. Based on all the related threads (see links below), it seems like that would be the best way of accounting for shared variance and avoid re-introducing noise. Within that confound matrix, we would include the cosineXX regressors, the non-steady state regressors, and CSF/WM regressors alongside the aromaXX regressors. I’d like t…

Best,
James

[ebelleau]

Hello,

I just wanted to follow up on this question. I just wanted further confirmation, is ICA Aroma performed in Native Space or MNI space? It will be important to know as it will affect our post-processing steps.

I took a look at the fMRI prep code, in one place it says:
#. Return classified_motion_ICs and melodic_mix for user to complete
non-aggressive denoising in T1w space
#. Calculate ICA-AROMA-identified noise components
(columns named AROMAAggrCompXX)

But then in another place following that bit of code it says:
Additionally, non-aggressive denoising is performed on the BOLD series
resampled into MNI space.

Thanks for the clarification!

Emily

[Oliver_Contier]

Just a quick follow up question: What exactly constitutes the mask here, within which the median is calculated? Is it simply the brain mask?

Thanks for your clarifications thus far, it really helped me with similar issues I was having.

[effigies]

@Oliver_Contier

Just a quick follow up question: What exactly constitutes the mask here, within which the median is calculated? Is it simply the brain mask?

It should be brain mask fMRIPrep calculates on the original BOLD series, resampled into the template space, in this case MNI152NLin6Asym.