learn2reg_nlst_paired_lung_ct.ipynb keypoints warping

[coolteemf]

Hi, I'm working on a new registration method, and decided to use learn2reg_nlst_paired_lung_ct.ipynb as a base. However, the conversion from keypoints from voxel coordinates to warping coordinates in the range [-1, 1] seems incorrect to me. It is done like this (in the forward function):

offset = torch.as_tensor(fixed_image.shape[-3:]).to(fixed_keypoints.device) / 2
offset = offset[None][None]
ddf_keypoints = torch.flip((fixed_keypoints - offset) / offset, (-1,))

But in the Warp class of Monai, the conversion is done like this:

for i, dim in enumerate(grid.shape[1:-1]):
    grid[..., i] = grid[..., i] * 2 / (dim - 1) - 1
index_ordering: list[int] = list(range(spatial_dims - 1, -1, -1))
grid = grid[..., index_ordering]  # z, y, x -> x, y, z

Which seems correct to me, since voxel coordinates should max out at dim - 1 for indexing.

The two are almost equivalent, except that in the notebook, image.shape is used and not image.shape - 1, so I think the notebook is wrong. Feel free to correct me if I'm wrong, because I'm lacking info on what exactly are those keypoints and how they were obtained.

[ericspod]

Hi @coolteemf it's possible the notebook is incorrect, could you provide a correcting code snippet? @nvahmadi could you have a look at this issue? Also the images the notebook used in the description are no longer available, would have these elsewhere? Thanks!

[flecomet]

Hi @ericspod correcting code snippet:

offset = (torch.as_tensor(fixed_image.shape[-3:]).to(fixed_keypoints.device) - 1) / 2
offset = offset[None][None]
ddf_keypoints = torch.flip((fixed_keypoints - offset) / offset, (-1,))

I added the -1 to the offset computation, which is now equivalent to the dim - 1 in the Warp function