Segmentation-aware template building

[tmspvn]

Dear expert,

I have a question regarding template building. I was considering to build a study-wise template to run a whole-brain voxel base morphometry analysis. However, using an image registration algorithm for the gray matter is non-trivial if one aims for accuracy.

I was wondering then if including a segmentation (e.g. just the GM ribbon or aseg from freesurfer) could improve the template building process. The plan was to include T1w, FA and segmentation with equal weights and re-apply the warps and recompute the segmentation template once the call finishes so to have the warps applied with nearest-neighbor instead of linear (or just skip the refinement).

Maybe masking only the GM during the registration could be another option and then refine it without masking in a second call?

How do you feel about it? Do you have any suggestions? What could be a better way to improve the GM registration without using surfaces?

Thanks a lot.

[gdevenyi]

The template builder can support multimodal inputs so in theory you can just add a label set as part of the inputs.

What's maybe missing is a blurring/similarity process that is "label aware" as label sets don't act quite the same as other images. Best thing to do is just try it out and compare.

[tmspvn]

Hi gdevenyi, thanks for answering.

Do you have any suggestion on a quantitative metric to evaluate the quality of the templates with and without the segmentation?

[ntustison]

Interestingly, and coincidentally, we're working on this exact question in the context of some relatively recent ANTsX publications such as this one. We hope to have put something together within a couple weeks or so.