Exceptionally fast non-linear 3D image registration using GPUs

We have developed code to test the potential of GPUs for speeding up typical 3D Medical Image registration tasks. We have used NVIDIA's CUDA programming language because this is very close to C and hence such code is likely to have lasting value. We find registration using both affine and B-Spline based non-linear transformations can be greatly speeded up. A typical very high resolution registration using up to one B-Spline control point every two voxels can be completed in about 5 minutes using a Normalized Correlation cost function and in about eight minutes using a Normalized Mutual Information cost function.