Just to fill in some details -

Greg is correct, there is no GPU accelerated registration in CMTK, only MR bias field correction and binary level set segmentation.

For these two, GPU code actually resulted in quite significant speedups.

The reason there's no registration is that we found at the time (~10 years ago) we'd have to sacrifice significant features, such as configurable interpolation kernels or some similarity metrics.

With current GPU APIs, it might now be feasible to revisit the feature/performance trade-off, but I have neither time but plans to do that.

Best,
Torsten