ANTsPy: Unexpected values with label transformations

Describe the bug I’m trying to align the Allen Institute mouse brain annotation volume to our own volumes using the following sequence: Register the Allen template to our own, transform 1 (Similarity + SyN) Register our own 3D brain acquisition to our own template, transform 2 (Similarity + SyN) After this, I use the transformations to transform the Allen annotation volume to our own acquisition image space. So I apply transform 1 and the inverse transform 2 to the Allen annotation volume to obtain the final aligned annotations. For applying the transforms I use the genericLabel interpolator. This mostly works fine except for a couple of regions. These regions now have values which were not present in the original annotation volume. This is a problem for us with further analysis and image generation as I can’t map the right colour to these values, as these values don’t exist in the Allen Institute structure trees. I suspect that it is summing label values somewhere, but I’m not sure as to the cause of the problem.

I can share the volumes if required.

To reproduce

rigid_transform_template = ants.registration(fixed=lsfm_template, moving=template_allen, moving_mask=None,
                                             type_of_transform='Similarity', write_composite_transform=True)
syn_transform_template = ants.registration(fixed=lsfm_template, moving=template_allen, moving_mask=None,
                                           type_of_transform='SyN',
                                           initial_transform=rigid_transform_template['fwdtransforms'],
                                           write_composite_transform=True)

rigid_transform_image = ants.registration(fixed=lsfm_template, moving=image, moving_mask=temp_mask,
                                          type_of_transform='Similarity',
                                          write_composite_transform=True)
syn_transform_image = ants.registration(fixed=lsfm_template, moving=image, moving_mask=temp_mask,
                                        type_of_transform='SyN',
                                        initial_transform=rigid_transform_image['fwdtransforms'],
                                        write_composite_transform=True)

new_atlas = ants.apply_transforms(fixed=lsfm_template, moving=atlas_allen,
                                  transformlist=syn_transform_template['fwdtransforms'],
                                  interpolator="genericLabel")
new_atlas = ants.apply_transforms(fixed=image, moving=new_atlas,
                                  transformlist=syn_transform_image['invtransforms'],
                                  interpolator="genericLabel")

Expected behavior There should be no integer values in the new image that were not in the original volume. The interpolator should assign only existing label values.

About this issue

Original URL
State: closed
Created 4 months ago
Comments: 40 (11 by maintainers)

Most upvoted comments

Sounds okay to me, too. But I defer to y’all.

ntustison on Mar 19, 2024

Using double should be OK in my opinion.

ncullen93 on Mar 19, 2024

I was confused before, I used numpy’s ‘float’ type which defaults to float64. With 32-bit floats, we see the same behavior

>>> a = np.asarray(614454277, dtype='uint32')
>>> af = a.astype('float32')
>>> af
array(6.144543e+08, dtype=float32)
>>> np.unique(af)
array([6.144543e+08], dtype=float32)
>>> af.astype('uint32')
array(614454272, dtype=uint32)

float32 can’t represent such as large integer.

I think the only ways to work around this are either to remap the labels, or to use antsApplyTransforms on the command line with double precision (don’t add --float).

cookpa on Mar 15, 2024

I think this might be it.

https://github.com/ANTsX/ANTsPy/blob/ff7413a5e7685b94ad7fc9a9b2553656f3bd005e/ants/registration/apply_transforms.py#L111-L114

The moving image is cloned to float, but the output image is a clone of the original without type specified. So antsApplyTransforms is trying to write its float output into a uint32 image array.

cookpa on Mar 14, 2024