Add optional NORDIC denoising step

[tsalo]

What would you like to see added in fMRIPrep?

If possible, it would be great to support NORDIC denoising.

NORDIC is currently written in MATLAB, which might be a problem.

Do you have any interest in helping implement the feature?

Yes

Additional information / screenshots

BIDS now supports separate noRF scans for the no-excitation pulse volumes people often acquire at the ends of their runs for NORDIC.

[effigies]

Would be good to know how NORDIC fits into the fit and transform workflows. Does it need to be done before/after HMC/STC? Is it deterministic/nondeterministic? What's the runtime? Is there a compact set of parameters that can be saved, or is it basically just the noRF files are the parameters?

[tsalo]

Pinging @madisoth for his thoughts, but my understanding is that it needs to be done before HMC, STC, or basically anything else. I don't think it's deterministic, unfortunately.

Most parameters are determined from the data (e.g., how many noise volumes, whether there's phase data or not). The only one I can think of that users might want to vary is phase_filter_width, but there's a strong recommendation to use a value of 10 for fMRI data, so even that could be hardcoded.

Not sure about the runtime since I've only ever run it on large numbers of runs in jobs.

[mvdoc]

Just popping in to say that re the MATLAB implementation, I recently found this package with a python implementation of NORDIC: https://paquiteau.github.io/patch-denoising/ (and associated paper: https://hal.science/hal-03895194/file/isbi2023_denoise.pdf)

I did not have a chance to try it myself (or NORDIC itself), so buyer beware.

[tsalo]

Thanks @mvdoc! The repo looks really promising, and even if it's not completely up-to-date with SteenMoeller/NORDIC_Raw (which it very much could be), it would still be a great jumping off point.

I also wanted to note that, for multi-echo data, NORDIC could be run on each echo independently (which I believe is how most folks do it), or we could try the HYDRA-OC approach from this OHBM abstract by Marco Flores-Coronado and Cesar Caballero-Gaudes. In that approach, it looks like you concatenate the multi-echo in a spatial direction (Z in their case), run NORDIC, and then split up the denoised data.

[tsalo]

I looked at the patch-denoising code and it doesn't seem like it uses noRF scans for its NORDIC implementation, so I'm not sure how close it is to the "official" method.

[tsalo]

I think we might need to use the MATLAB version.

@Remi-Gau do you know any tricks for using MATLAB in containers/BIDS apps?

[Remi-Gau]

Yeah. Make sure the code is Octave compatible and use Octave.
Otherwise you are down the road of compiling Matlab code...
Where is the Nordic code base?
If it's the one I am thinking of I may have "opinions" about it...

[tsalo]

This one is the one I was planning to use, since it's the one others seem to be using. I don't know if it's Octave-compatible though. It does rely on the nifti/20130306 libary.