DWI denoising!




I have tried to denoise my dataset using the default window size. But, on eyeballing the residuals, I found some anatomical information!

A similar issue was discussed previously http://community.mrtrix.org/t/dwidenoise-correct-use/586/2.

I’m looking for some feedback again :slight_smile:

Best regards,


OK, that pattern in the noise map is a matter for concern. It’s hard to explain this away as anything other than a fixed-frequency RF contamination, or similar. I assume your frequency-encode (readout) axis is L-R? If that’s the case, you’ll need to check that the RF shielding is not compromised on your scanner, and whether there is any electronic equipment in the vicinity of your scanner that could emit RF noise at close to the Larmor frequency of your scanner…

The residuals map also look unusual, but this probably corresponds only to the first volume, not the RMS across volumes? So there is a definite pattern for residuals of presumably the b=0 image (?). It’s worth looking through all volumes of your residuals image to see whether this pattern is consistent across volumes, or inspecting the RMS residuals using a command like:

$ mrmath residuals.mif rms -axis 3 - | mrview -


Yes, it seems there is some RF leakage in the images (clearly visible in a couple of slices after scaling). So, I guess denoising should be be best avoided for this data/similar others :confused:

The residuals in the diffusion volumes look ok based on the lack of information.

Looking forward to your comments.



Dear Experts

I have a similar issue to the above. When looking at my residuals I was concerned that I could see a bit of the lateral ventricles and perhaps the outline of the posterior fossa. Can I ask what you think of these snapshots? Do they look as expected or should I be concerned? Thanks as always! Paul


Hello experts,

I have a question about the denoising, is it possible to run the denoising in the one image twice, and the results be slightly different? Thanks in advance.




Hi Manuel,

dwidenoise should be entirely deterministic, so the result should be identical on repeated runs.

Did you find otherwise? Are you sure everything is set up identically?



Hi Daan,

I think I found the answer. When I run it now, I found the same results, but when I compare the image denoised now (version 3.0_RC2-2-gca446aad) with the same image denoised with the version 0.3.15-294-ge8a525c6-dirty, I found slightly differences, is that possible?




Hi Manuel,

Yes, it is possible that you get small changes between different versions. As long as the differences are small you don’t need to worry about it. Let me know if you think they are too large to ignore, then I can investigate.



Hi Daan,

Thanks for your answer! The difference in the subject I was exploring is very small, I don’t think there are reasons to be worried.




Hi All,

I’m adding to this thread, because I too am working with rodent DWI and have some interesting results from the dwidenoise command. This is my first time using dwidenoise, and none of the results I got seem to fulfill the “good residual” requirement of being free of anatomy.

Attached image shows the noise map and rms residual for dwidenoise, using three different settings for extent: 3, 5, and 7. Initially, I ran dwidenoise with default extent, yielding the middle image. It seemed suspicious, since there was definitely a brain outline. After consulting with others, I ran using the other extent setting, resulting in the other two results, which curiously shows even more anatomical structure.

This dataset contains 107 volumes, with 10 b0 images, and is obtained ex-vivo acquired over the course of 12 hours. Is this perhaps why so little noise is estimated with the default setting? This is one of the cleaner data, is this good data actually getting in the way of dwidenoise?