Default CC radius in antsRegistrationSyN.sh script

[ntustison]

@cookpa and @stnava

I changed the default radius to 2 in ANTsPy in the antsRegistrationSyN equivalent based on the BraTS eval some time ago. However, due to some oversight, I haven't changed it in ANTsR or the original ANTs script. I'd like to keep them coordinated. Any preference? If not I'd like to change the default to 2. Thanks.

[cookpa]

I'm good with 2.

As we are changing defaults, what do you think of boosting the default syn step size from 0.1 to 0.2?

I remember finding a small step size more reliable years ago but I think many of the issues have been solved by better pre processing and affine initialization

[ntustison]

Sure. Sounds good to me. Would you like me to do both at the same time?

[ntustison]

NM, I went ahead and changed them all:

#1829
ANTsX/ANTsR#417
ANTsX/ANTsPy#770

[gdevenyi]

As we are changing defaults, what do you think of boosting the default syn step size from 0.1 to 0.2?

I found this also to be better for human data.

However, does the BRATs challenge include anything non-human? Might this be too aggressive for mice/rats?

[gdevenyi]

P.S. I think semantic versioning means this should be a 3.0 release when it goes out.

[ntustison]

@gdevenyi -- If you're worried about spacing differences, the gradient step takes that into account.

[gdevenyi]

@gdevenyi -- If you're worried about spacing differences, the gradient step takes that into account.

Thanks. I was thinking more about "mice are more similar than human brains and likely need less deformation overall"

As we are changing defaults, what do you think of boosting the default syn step size from 0.1 to 0.2?

How much effort would it be to add this change to the BRATS code @ntustison already tested and see how it goes?

[cookpa]

IDK how forward compatible the evaluation script is. It was based on antspy 0.3.7. Could try updating the base container, but I don't have time to do this

https://github.com/cookpa/ANTsXBratsReg22Singularity

[ntustison]

"mice are more similar than human brains and likely need less deformation overall"

I don't think this is a correct understanding of how the gradient step itself translates into the overall deformation of a registration scenario. It's a bit more sophisticated than that. The step size choice is more about how cautious one should be in updating the total field based on the update field. 0.1 is at the extreme conservative end with some early usage utilizing step sizes of 0.5. So I think @cookpa is correct in wanting to adjust it a bit.

How much effort would it be to add this change to the BRATS code @ntustison already tested and see how it goes?

I don't have time to do this right now but perhaps in a couple weeks.

[gdevenyi]

I don't think this is a correct understanding of how the gradient step itself translates into the overall deformation of a registration scenario. It's a bit more sophisticated than that. The step size choice is more about how cautious one should be in updating the total field based on the update field. 0.1 is at the extreme conservative end with some early usage utilizing step sizes of 0.5. So I think @cookpa is correct in wanting to adjust it a bit.

Thanks, that's indeed how I understand it, I'm just approximating total deformation as (# of iterations X step size) as a max, and 0.1->0.2 is a big change.

We indeed used 0.5 in the ANTS days as well.

[ntustison]

I'm just approximating total deformation as (# of iterations X step size) as a max, and 0.1->0.2 is a big change.

As I alluded to above, I think this is an overly simplistic view which doesn't take into account convergence and the evolution of magnitude of the similarity metric gradient as the deformed objects become more similar. Just consider a scenario in which the fixed and moving image are the same.