Gauging interest on spatial transforms

[oesteban]

I'm about to start working on python (mostly numpy) based geometric transformation of 3D images. I feel nibabel could be the best host to a new module nibabel.transforms.

I would start by implementing affine transforms (which is almost there already with resample_from_to) and free-deformation fields (e.g. generated by ANTs registration).

Next steps would include taking in collapsed transforms (hdf5), and also parametric transforms (e.g. BSpline).

WDYT?

[oesteban]

Here there is an example: master...oesteban:enh/transforms. This code makes it easy to apply an affine transform on images:

import nibabel as nb
ref = nb.load('image.nii.gz')
xfm = nb.transform.Affine([[1, 0, 0, 4], [0, 1, 0, 0], [0, 0, 1, 0], [0, 0, 0, 1]])
xfm.reference = ref  # Set a reference image
moved = xfm.resample(ref)  # Resample ref into reference space through xfm
moved.to_filename('moved_4mm_image.nii.gz')

The purpose of the module would also include reading linear and nonlinear transforms from different formats (mainly FSL, ITK/ANTs, SPM, FreeSurfer, etc.).

cc @effigies

[effigies]

This looks useful. I've done some work already with FreeSurfer's LTA format in #565.

So a Transform is defined to be RAS-RAS?

[oesteban]

#565 is definitely related to this.

In principle, Transforms are Voxel-to-voxel. However, they must have an internal parameterization of the transform. That internal model should be defined to be RAS-RAS and accessible through the map_point method.

[oesteban]

Sorry for the second post. For RAS-RAS transforms with the xfm object of my example would be:

>>> xfm.map_point([10.0, -10.0, 20.0])
array([14.0, -10.0, 20.0])

However, the xfm.resample method needs to work voxel-to-voxel.