Hello MRtrix community,
I am using dwidenoise on diffusion data acquired on ex vivo mouse brains placed in a syringe filled with fluorinated oil (Galden). I have multiple b-values, directions and diffusion time values, amounting to a total of 620 volumes in the 4th dimension which I use jointly to do the noise estimation. When using the default kernel size (9x9x9), I observed on the noise map a sort of a halo all around the brain with very high sigma values. This halo decreases in size as I go to lower kernel size (eg., 7x7x7 and 5x5x5). Since this effect is really just at the border of the brain and is dependent on the kernel size, is it correct to assume that the issue comes from the very large difference in signal and noise between the voxels inside and the outside the brain that are in the same patch when at the border of the brain?
Top row, L to R: sigma map for kernel 5x5x5 and sigma map for kernel 7x7x7; Bottom row: sigma map for kernel 9x9x9; The contrast of all maps was adjusted between 0 and 400 for comparison
If I look at the noise map, for the 5x5x5 kernel size the maps seem to be a bit more homogeneous, the halo effect is less pronounced, however in the most inner parts of the brain the estimated noise is two times lower than the one estimated with kernel size of 7x7x7 and 9x9x9 (these two are closer one to another). Another peculiar observation is the zero values on the sigma map resulted from denoising with 9x9x9 kernel size (all the dark voxels that you see inside the brain). I mention that I am using the -mask argument. Here is an example of the mask I am using. I tried more restrictive and less restrictive masks, and the result was still the same.
b0 image with overlayed mask
I am really not sure what denoising strategy would be best in this case. Did anyone run into a similar issue? Any suggestion would be very helpful!
Thank you,
Best regards,
Andreea Hertanu