Which stages is proper to resize dwi image?

preprocessing

#1

In the reconstruction of the dwi data, our scanner pad some zeros to the K matrix, then we get a 256x256 image, the voxel size = 0.8x0.8. If I want to back to my original voxel size(1.5x1.5), which stages is proper to do this ?
0, befeore any preprocssing
1, after dwidenoise
2, after dwipreproc
3, after dwibiascorrect
4, after dwi2fod

Thanks ,
Zonglei


#2

Ideally, you’d disable zero-filling on the scanner - it’s detrimental to image quality and introduces problems. This is primarily because the magnitude transform ‘rectifies’ negative signals and introduces odd patterns in noisy regions – of which there are plenty in high b-value dMRI… Unfortunately, some scanners makes it harder than others. It’s trivial to disable on Siemens (in fact, it’s rarely enabled by default), a bit more obscure on Philips, and Guru-level on GE systems…

Unfortunately, once the magnitude transform has been done, there’s not much you can do to reverse the operation – at least not perfectly. If I was to do this, I’d probably opt for doing it right at the start. The main thing to check is whether there is any difference in performance in dwidenoise if you denoise first and resample after, or vice-versa: this command is likely to be the most sensitive to changes in the noise properties that the resampling might introduce. But this is just my gut feeling, maybe others have more direct experience with this…?


#3

Thanks for your clear explanation, Donald.
Unfortunately, our scanner is GE system. Although the issue has been fixed now, a lot of data have been scanned. I am analyzing some of these old data. tough life!
I’ll follow your suggestion to resample the data and see what will happen.
Zonglei


#4

Hi all,
See bellow for the resampling effects in dwidenosie. The apparent voxel size of my data is 0.87x0.87x2 and the actual voxel size is 1.5x1.5x2. I dowmsampled the raw data to its actual resolution using mrresize, and denoised it wiht dwidenosie.
It seems that the resampling is OK. and the SNR is improved to some extent because of the large voxel size.
Besides, after downsamping, it save a lot time to run dwidenosie.

@jdtournier what do you think?

Zonglei