Documentation on using empirical noise scans with NORDIC

[tsalo]

The actual math goes way over my head, but it doesn't seem like patch-denoising includes a way to use noise scans with the NORDIC denoiser. The documentation mentions "pre processing the image and dividing it by an externally available g-map". Does that mean that calculating the g-map is outside the scope of this package?

[paquiteau]

Hi again,

If you look closely at the NORDIC MATLAB code, they have routines to estimate the noise level per patch in the absence of g-map (I tried to reproduce this in Nordic_Raw in patch-denoise).

G-map estimation goes beyond the scope of this package indeed, but our team is tackling this challenge in another repository with https://github.com/mbertrait/ggrappa (cc @mbertrait)

[paquiteau]

Actually, the maths are not that hard. I will soon write my thesis, and patch-denoise will be a chapter of it, I will take this opportunity to detail the maths more (and add it here in the documentation)

[tsalo]

That would be amazing. For context, I would really like to get NORDIC added to fMRIPrep, but using the MATLAB version would dramatically increase the size of the container per @bpinsard, and there are also license and code quality issues. Being able to use the noise scans within patch-denoise or having solid documentation so we could write an interface to do that on fMRIPrep's side would be awesome.

[paquiteau]

Turns out, I have some code in place to estimate the noise statistics from noisy scans 🥳 !

https://github.com/paquiteau/patch-denoising/blob/master/src/patch_denoise/space_time/utils.py#L206

And even better, I have nipype bindings (which fMRIprep uses IIRC) for it

https://github.com/paquiteau/patch-denoising/blob/master/src/patch_denoise/bindings/nipype.py

For the documentation, I have no ETA yet, but contributions are welcome to improve it ;)

https://paquiteau.github.io/patch-denoising/denoisers.html

[bpinsard]

Amazing, thanks for the nice package! I am glad that this interesting MRI filtering technique won't die as an old, barely-readable MATLAB code with restrictive license.

[tsalo]

I've been reviewing the MATLAB code in detail and it turns out that the measured noise is a single scalar, not something calculated on a patch-wise basis. Also, in the MATLAB code they only use the first noise scan. It's clearly a bug, but the author doesn't think it merits fixing.

[paquiteau]

Hi @tsalo, that's a good catch, this might explain the discrepency between the result of the python vs MATLAB implementation (note that in the benchmark paper, I used the MATLAB NORDIC for fair comparison ¹). The NORDIC authors are still using their implementation as is (and apparently with good results ²), so maybe the Python version should also use this "bug".

Footnotes

1. It was a PITA to make MATLAB, Nipype and NORDIC to work together smoothly: https://github.com/paquiteau/retino-pypeline/blob/master/src/retino_pypeline/interfaces/nordic.py ↩

2. https://mitp.silverchair-cdn.com/mitp/content_public/journal/imag/jam/10.1162_imag_a_00426/2/imag_a_00426.pdf?Expires=1738316200&Signature=QTQ5ZQtxBjopglXJYziw1tBv2EkodPPlRvAhXUcpp~0ArnI-zVOGPo7CdWik4kwYRwpS6Hasdkk0IU-lwgUStqRFG0DhuazrrIdbS0YtjCn-sQ747B64pDcni2WxO~Y-wwsnpl7hNckBVVRJQGtd1KkRvuUjhcCFM36jMczibQMv4-r6GBGGwcktgbqugE47Wb1QE7xuqq9gWBNUIyo5EYo8ADFsoFzQfgu94J-CGdoEjOuQZp0EWEncaqb0MfmfRji8jDfNnWqbiPu3goqJmvHBc45-Egq3PEu9D0UDgXZUZhVYmbafSVunq7iEvyVaHqhSKcmzEBHv9lpiZVWRWQ__&Key-Pair-Id=APKAIE5G5CRDK6RD3PGA ↩

[tsalo]

I don't think the bug (using just the first noise volume) will have much of an effect, but it is definitely a bug that should be fixed. Even if only so that people aren't acquiring multiple noise volumes for no reason.

I have a line-by-line Python translation that seems to work pretty well. I still need to write unit tests for it, but I'm optimistic. My ultimate goal is to incorporate it into dwidenoise, but I'd love to do what I can to bring the patch-denoising implementation in line with the MATLAB implementation.