I am just updating our diffusion processing pipeline to include the dwidenoise and mrdegibbs steps. Our raw data is organised in two DICOM directories, one containing the main diffusion-weighted data, and one containing reversed phase-encode data. I am wondering whether I am ok to perform the dwidenoise and mrdegibbs steps only on the main diffusion volume only? Or should I also run the steps on the reversed phase-encode series?
My data is in a similar format as yours, and the first thing I always do is to create the file with 2 B0s, one in each phase encoding direction, to use it in dwipreproc (topup), then I apply dwidenoise and mrdegibbs to the actual data. But you could also concatenate everything at the beginning and run dwidenoise and mrdegibbs all together, and then split it again. Both approaches seem reasonable to me. I hope this helps.
Thanks for the swift reply @mblesac. That’s really helpful. I will try the latter approach.
Just to follow up on this, I’ve run the preprocessing steps (dwidenoise & mrdegibbs) on;
- the concatenated data, including b0s in both phase-encoded directions
- the main data, including b0s in the phase-encoded direction only
The residuals look better when the steps are run on the main data only. I’ve seen other threads where concatenating is advised e.g. here
I’m struggling to see why though. Would the distortions not be in the opposite directions and could this not result in anatomical details being mapped differently?
My personal intuition with regards to use of concatenated data with different phase encoding directions for
dwidenoise is that as long as the absolute noise level does not differ “substantially” between the anatomical locations from which the image intensities in a particular voxel originated, then the mechanism should work fine.
However I’m reconsidering the point now having been going through the process of determining new DWI protocols: with the gravitation toward high MB factors and disabling of in-plane acceleration, the resulting more severe distortions may bring this into question.
Certainly if you have empirical observations from your own data that the concatenation approach may be somehow problematic, I’d be curious to see it.