Incorporate VoxelMorph into evaluation of registration methods

[jamesfishbaugh]

The open issue regarding voxelmorph is the output deformation and it's format.

• Voxelmorph documentation refers to it as a deformation field but in my experiments I recall it to be a displacement field. This should be double checked.
• We should double check if the displacement field is in voxel or scanner units and if the evaluation framework is handling the units correctly.

[ebrahimebrahim]

Double checked an output: Indeed it is a displacement field. The voxelmorph documentation probably meant that as "deformation field" is a bit ambiguous.

(Still looking into the voxel vs scanner units question)

[ebrahimebrahim]

In an example, I visually compared the target image, the warped moving image that comes out of voxelmorph, and the warped moving image that I obtained by apply mrtrix's mrtransform to the transform generated by transformation_utils.convert_voxelmorph_transform_to_mrtrix_transform. They all look aligned! Just slightly different interpolation methods between the two warped images. So I would say that confirms that coordinates are being handled correctly (or at least not so incorrectly that it is perceptible).

mrtransform accepts -warp in scanner coordinates, not voxel coordinates. Voxelmorph however seems to be unaware of scanner coordinates... so how is this working out? I think it all works out the same because what we are passing around is the nifti file and we are keeping its header intact which contains the affine. So even though the transform generated by voxelmorph is most likely in voxel coordinates and voxelmorph is unaware of scanner coordinates, mrtrix still knows what to do with the transform nifti file because the header is there.