Hi, I find myself in a situation where my collaboration provides only preprocessed eddy-corrected dwi data (no denoising/deringing).
I’m aware that based on this post andthis post the recommended preprocessing order seems to be
- dwidenoise
- (mrdegibbs)
- dwipreproc
- dwibiascorrect
Nevertheless, I wonder if there are any definite arguments against running
- eddy
- dwidenoise
- mrdegibbs
- dwibiascorrect
in a situation where only eddy-corrected data is accessible. I understand that dwidenoise operates based on the uncorrelated noise assumption in p and q space, so it will perform worse due to some possible interpolation from the previous command. Visual inspection shows an improvement after the denoising, so I think it should still be better than not running it at all. Would be very interested in experiences or opinions on the topic.
There was a recent thread on the mrtrix forums that said that we should actually be running dweidenoise after eddy correction, and while searching for that post I came across this one and thought I’d link to the answer for what it’s worth (even though it’s a year later): Eddy/topup on denoised data
From that thread, someone posted “part of a discussion with Jesper Anderson (he is one of the authors of eddy)”:
Yes, I don’t recommend using denoising prior to eddy. I have seen very poor results on denoised data. It tends to mean that one ends up with unrealistic estimates of hyperparameters for the Gaussian Process, making “everything” look like valid signal variance and nothing look like noise. It would also interfere with eddy_qc and lead to inflated estimates of CNR.
I would also have thought that any denoising would work much better after eddy, where for example outliers have been detected and corrected.