advantages of a multivariate group template for ROI segmentation?

[gjoue]

Hi,

We have T1w-MPRAGE and T2w-TSE images for which we were thinkinging of creating separate group/study templates and registering them to an atlas that has templates in both T1w and T2w. Would it be right to think that a multivariate approach would potentially lead to a more representative group template, especially as the T2w is a higher resolution slab (although I've had problems before with ANTs trying to coregister slabs to whole-brain images)? When would you want to do this/avoid doing it/what are the advantages of creating a single multivariate group template over separate T1w and T2w group templates? And if we did create a multivariate group template, how would you go about also combining the atlas' T1w and T2w templates for automatic segmentation of ROIs from the atlas?

This question branches off of Philip's answer to another question here:
#1878

Thanks in advance,
Gina

[ntustison]

A multivariate template vs. separate templates of different modalities certainly simplifies any normalization and/or analysis strategies. Perhaps we can start off with just one question and continue the conversation from there.

[cookpa]

The template script won't do any coregistration within subject. That part has to be done first, so the T1w and the T2 slab for a subject have to be aligned.

The multivariate template approach will register, simultaneously, subject T1w to template T1w and subject T2w to template T2w. I've not done this with partial coverage, but I think it ought to work. You might consider upweighting the T1w - by default both modalities are weighted equally, but you can change that.

Slab to volume registration has been tricky and does require some careful initialization and more customization of the defaults. We've made some progress recently, for example, making it easier to do "SyN only" registration without any center of mass initialization that was throwing slab registrations off.

[gjoue]

Thank you, Nick and Philip, for your replies! I wanted to do some testing first, but time isn't on my side, so I'm going to take the discussion route, as Nick suggested.

Philip, thanks for the practical details -- these are exactly the functional details I wanted confirmation of. I will try the geometric-center init because I think it must have been the center-of-mass init that made me give up using ANTs to coregister the T2w slab to the whole-brain T1w image. I have several followup practical questions:

I-1). Now, during registration, the image-to-move will be sampled to the reference image, right? If I moved the T2w slab to the whole-brain T1w, that means losing resolution (from 0.4x0.4x2 mm vx to 1-mm isotropic), which I suspect would not be great for some of the small subcortical areas we are interested in. But we can't move T1w to the T2w because we would reduce T1w to a slab, and we are interested in doing VBM analyses and would need some global brain values -- so would the best route be upsampling the T1w's to the T2w resolution (ResampleImage?), then registering the T2w's to the upsampled T1w's, and then creating a multivariate template?

I-2). If the multivariate template approach simultaneously registers subject T1w/T2w to template T1w/T2w, respectively, does that mean that these (study?) templates are the same as what would have been created if you were to create a study template for each modality/image sequence individually, and if so, is there a way of starting the multivariate template creation after the creation of the template for each modality?

And this also leads me back to my initial question about combined multivariate T1+T2 model or separate models for each, but rolled back a bit:

II-1). As we are interested in small subcortical areas that are more defined in the T2w slab, would you upweight the T2w during template transformation? Or -- for certain regions that might be better identified in one of the modalities, would this be a case where having separate templates would be more advantageous?

Thank you thank you!

[cookpa]

Philip, thanks for the practical details -- these are exactly the functional details I wanted confirmation of. I will try the geometric-center init because I think it must have been the center-of-mass init that made me give up using ANTs to coregister the T2w slab to the whole-brain T1w image.

Also time limited but wanted to say about this

If you open a T1w and T2w togther in ITK-SNAP, are they aligned? If so, you can use them for a multivariate template. If not, you need to align them ahead of time.

If you align them ahead of time, geometric centers might not work either. If the slab is close but needs small adjustment, it's best to not use any initialization. Use -r Identity and go from there. You might also need to constrain the registration with a mask.