I’ve just looked at the code for this, and yes, that would be the reason. As it says in the command synopsis, it looks for b=0 volumes, then computes the mean b=0 intensity for each voxel in the mask, which will give Not-a-Number due to the divide by zero.
More to the point though: these days we would advocate the use of mtnormalise to deal with both intensity normalisation and bias field correction – even for single-shell data where we’d advocate the use of multi-tissue CSD with WM & CSF responses only (see e.g. this post, amongst others). I think we’ll eventually end up removing dwinormalise altogether…