ENH dwidenoise2: b=0-based noise level estimation

[Lestropie]

From: https://direct.mit.edu/imag/article/doi/10.1162/imag_a_00049/118324/Efficient-PCA-denoising-of-spatially-correlated

This would involve:

• Prior to denoising, extract b=0 volumes, compute variance, median filter
• Store this in a scratch image as an a priori input noise level estimate
• Perform denoising where the "estimator" just reads from that scratch image

[Lestropie]

On re-reading, what is proposed in that manuscript is different to what was originally listed here. What they actually do is:

• Compute the voxel-wise variance across b=0 volumes
• When denoising, select the median value of this variance from across the voxels within the patch

That would require some different handling. There would still be a scratch image containing the pre-computed b=0 variance data, but without median filtering. This image would need to be provided, in conjunction with the set of voxels involved in the patch, to a new estimator class, which would sample that image across the patch voxels and compute the median of such.

While that could be done technically, this would be somewhat incompatible with the prospect of presence of a variance-stabilising transform.

Given I'm a little sceptical of this technique and its sensitivity to geometric distortions, I'm going to omit this from #3023.