Hi,
I ran dwidenoise on a 4D diffusion weighted image volume (100 x 100 x 20 x 200). However, in each volume, I see that the first and last slices are not denoised. I tried with different extents but to no avail. Is this expected?
Hi Manu,
dwidenoise uses a sliding-window approach to estimate the noise distribution locally, based on its neighbouring voxels. At the image boundaries, less neighbourhood information is available and the noise distribution is difficult to estimate. In this case, Veraart et al. take a conservative approach in which the data is not corrected (not denoised).
Therefore, you should see a similar effect at all bounding slices and it is perfectly normal. If you only see it in slices across a single axis, let me know so that I can investigate.
Cheers,
Daan
@dchristiaens: Is there anything theoretical precluding the use of a wrap-around window? Would have thought that’d be appropriate for readout & phase axes if it’s the first processing step applied subsequent to the FFT; would hope it’d be better than nothing in the slice direction…
I reckon the noise level would be expected to differ significantly between different sides of the FoV, which would be a serious enough problem that you probably wouldn’t want to do this…
Maybe for the phase encode (in that the FFT genuinely would wrap around), but not for the readout axis: most modern systems oversample along that axis by default, giving twice the actual FoV, then crop out the outer edges (means you can have homogeneous signal across the FoV without a really sharp bandpass filter).