It's currently possible to run dwidenoise on DWI data containing a tiny number of directions. I expect that determination of the MP threshold will be highly discretised in this case, which may make the influence of denoising quite inconsistent between adjacent voxels. Perhaps if the number of volumes is smaller than what would result in use of a 5x5x5 kernel, i.e. fewer than 27, such that a 3x3x3 kernel is utilized, this should be stated at a warning level, along with the potential consequences of such. This would also catch people trying to denoise individual volumes, which probably happens.