registration advice

[joue.ni...@gmail.com]

Hi,

I have high-res T2w structural slabs and T1w whole brain structural images to run ASHS with the ASHS-PMC atlas and would like to get advice from the ASHS community on what registration tools people have successfully used to coregister these images before I go trying a whole bunch of things that the community knows doesn't work anyway. The ASHS Google Sites doc says that the T1w and T2w should be "roughly aligned in space" -- not sure how "rough" is still okay, but I assume if I'm not even in the MTL in one of the images when I click on somewhere in the MTL in the other corresponding image, then it's not "roughly aligned" enough?

I have tried the following so far in my current dataset:

* SPM 12 -- straightforward to be able to realign without resampling/reslicing, but for my current data set, the resulting registration is pretty bad for everyone (though this had resulted in decent registrations for another dataset I had in the past with T2w-TSE and T1w-MPRAGE on which I subsequently ran ASHS, which produced pretty good-looking Hc segmentations).

* ANTS -- rigid/linear registration with antsRegistration, moving slabs to T1w. Registration looks great, but ANTS resamples the registered images to the resolution of the target image, and there does not seem to be an option to modify image headers only (Nick Tustison has confirmed this). This resampling causes ASHS to create pretty bad segmentations and quite a few aborted attempts. Advice on how others have dealt with this would be much appreciated. Is the key to move the whole-brain to the high-res slabs? I would have thought that the clipped FOV/reducing whole-brain T1w to slabs would affect ASHS?

I haven't tried FSL's flirt yet. The flirt user guide notes the difficult in coregistering slabs with whole brains and recommends moving the slab to whole brain. My question would be whether it is possible to do registration with no resampling in one flirt command, or do you have to do it in several steps? If it has to be in a few steps, would the following be correct? (1). flirt slab -> whole-brain, which resamples; (2). create some dummy reference image with the high-res slab resolution and dimensions; (3). flirt whole-brain -> dummy reference; (4). convert_xfm -concat the transformations in (1)+(3); (5). flirt to apply combined transformation in (4) to slab.

Any advice and in particular specifics (like command options) very much appreciated!

Gina