Question regarding midpoint spine registration

[jqmcginnis]

Hey ANTS community,

I am currently working on longitudinal lesion quantification in MS for spinal cord images, and I have been wondering if you see any of the ANTs tools as suitable for this setting to perform this analysis in mid-space.

Right now, I am using the Spinal Cord Toolbox, which wraps ANTs to do the registration (first rigid alignment based on disc labels (not images), second centroid-based alignment using the spinal cord masks (again not using the images)), and then naively halving it, which is, I guess, not technically correct, but gives nice (emperical) results.

I just discovered the SyN-based / template-based example scripts, but I am unsure how to combine this with a segmentation label-based approach. Do you have any advice? This would be greatly appreciated!

Thank you!

[ntustison]

So you can go the template route using the template construction scripts. I have used it for constructing templates using label images but If you're simply looking for the midpsace, that might be a bit overkill. I'm pretty sure there's a midspace generation script in the scripts directory which wraps antsRegistration. It's going to be intensity-based which you can modify for labels but I don't know how easy that would be.

More recently, though, there's a couple developments in ANTsPy/ANTsR which might be relevant for your work. They all stem from using anatomical labels for driving registration motivated by the fact that segmentation is so quick and easy now. The links below are from a set of code snippets which demonstrate usage located here.

• label image registration: pairwise registration using paired label images (or multiple sets of paired label images) and, optionally, intensity images. It handles linear (using label centroids) and SyN-based transforms.

Although the SyN algorithm does have an underlying midspace, with transforms to and from both the fixed and moving images, making that available to the user would've been more difficult than we were willing during the ITK development years ago. However, I did develop some pointset-based transforms which do expose that (demonstrated here under the "More complicated example").

• pairwise pointset registration
• longitudinal pointset registration

Something very much related was a recent publication for a longitudinal registration scenario (github repo here). A collaborator had developed a set of templates using ANTs tools showing the longitudinal development of the mouse brain and wanted to be able to link them together through spatial transforms and was hoping for something more sophisticated than pairwise SyN transforms between timepoints. So I used pairwise SyN transforms between time points to construct a time-varying velocity field (also using related ANTsPy/ANTsR tools which can then be used to construct the transforms between any two time points within the bounds of the temporal domain).

Although requiring a two-step process, these approaches do provide a way of exposing the transforms you might want that aren't available to the user directly from ANTs tools.

[jqmcginnis]

@ntustison Thank you so much for the great feedback. I have been running some of the scripts you mentioned and will report back to let you know which one we ultimately opted for 👍

[ntustison]

Great. Let me know if you have any questions or come across any of the examples which don't work as expected.