The tools/GridTransformer.h methods for resampling grids from one transform to another uses a per-thread copy of the target grid and a scatter algorithm that is well-suited for scalar and vector grids, but performs poorly with mask and bool grids that are required to use a nearest-neighbor interpolation. GridResampler::transformBBox() in particular is a bottleneck.

It is worth experimenting with different configurations, a low-res grid being resampled into a high-res grid and vice-versa for example. However, I expect an improved implementation could be a gather algorithm for this particular use case.