Split rigid and overall deformation

[antonfrancois]

Dear authors,

I am trying your model and find it very interesting. I successfully registered images from the command line and retrieved the .hdf5 file with the transformation information. Upon inspection, I noticed two rigid and one non-linear deformation within the file.
It is not clear to me (and I didn't find a documentation) which rigid deformation does what.
I would like to apply the rigid deformation only to my images to test and evaluate the contribution of the rigid vs the non-linear deformation.

I made a itk code so far from #21 but I am not very familiar with this library. Can you help ?

Anton

[HastingsGreer]

Hi @antonfrancois ! This has been a longstanding goal of our codebase, and is currently available in beta. It requires a different architecture than uniGradICON, which we have just finished training. To get it, run

pip install git+https://github.com/uncbiag/uniGradICON.git@unicarl

unicarl-register --fixed your_fixed.nrrd --moving your_moving.nrrd --transform_out trans.hdf5 --warped_moving_out output.hdf5

Note that with the unicarl model, it is not necessary to specify mri vs ct, and the same model is used for multimodal and monomodal registration.

The resulting transform file will have four components: the components will be (physical coordinates to network coordinates, affine transform between images, deformable residual, network coordinates to physical coordinates)

To get just the affine transform, simply remove the deformable residual. Sample code to do this is

affine_Transform = itk.CompositeTransform[itk.D, 3].New()

affine_Transform.PrependTransform(phi_AB.GetNthTransform(3))
affine_Transform.PrependTransform(phi_AB.GetNthTransform(1))
affine_Transform.PrependTransform(phi_AB.GetNthTransform(0))

The architecture allowing this is documented in

https://openaccess.thecvf.com/content/CVPR2025/html/Greer_CARL_A_Framework_for_Equivariant_Image_Registration_CVPR_2025_paper.html

The training run generating the "unicarl" weights by training carl on a composite dataset has not yet been accepted for publication, but it should be out soon.