If there are a sufficient number of volumes per DWI input series, then it may be preferable to denoise each individually, since that provides a greater guarantee of constant noise level across all volumes within a patch. This is the current behaviour.
If however the input images are aggressively split, e.g. few total volumes across many phase encoding directions, then the number of volumes per series may be relatively low, to the point where the rank of the data matrix becomes quite small and therefore denoising efficacy is reduced.
So it may be the case that in such a scenario, it is in fact better to concatenate the DWI series prior to denoising. This does introduce the risk of having a non-constant noise level across all volumes, but would increase the rank of the data matrix such that the MP-PCA thresholding process is not as detrimentally affected by quantisation. To me a logical threshold would be if the number of volumes per series is equal to or lower than the number of voxels in a 3x3x3 neighbourhood (27), ensuring that at least a 5x5x5 neighbourhood is used.
If there are a sufficient number of volumes per DWI input series, then it may be preferable to denoise each individually, since that provides a greater guarantee of constant noise level across all volumes within a patch. This is the current behaviour.
If however the input images are aggressively split, e.g. few total volumes across many phase encoding directions, then the number of volumes per series may be relatively low, to the point where the rank of the data matrix becomes quite small and therefore denoising efficacy is reduced.
So it may be the case that in such a scenario, it is in fact better to concatenate the DWI series prior to denoising. This does introduce the risk of having a non-constant noise level across all volumes, but would increase the rank of the data matrix such that the MP-PCA thresholding process is not as detrimentally affected by quantisation. To me a logical threshold would be if the number of volumes per series is equal to or lower than the number of voxels in a 3x3x3 neighbourhood (27), ensuring that at least a 5x5x5 neighbourhood is used.
Curious if you have any thoughts @dchristiaens.