Hi Torsten,

I think you are providing good points here. Those thresholding or masking options are definitely important for registrations.

About my issue on irtk, I think I find the real issue. It is not about isotropic resampling, but the starting point, i.e., the position of the moving image to start with. The subject I tested on before turns out a special case.

What I did this time was to apply an identity transformation matrix first using FLIRT, save the temporary image, and then use irtk to do the registration. And the registrations look pretty good to me. I guess that irtk uses some header information to initialize. Maybe they just align the origins of two images, which may always work for brain images, but not really the case for abdominal scans. In some (actually quite a lot) cases, the moving image starts outside the FOV of the target image, and that's why I saw a lot of empty registered images, where neither the multi-level registration nor the default step size of irtk can get the moving image back.

Interestingly, if I convert an identity matrix into their transformation format "*.dof", and use that "*.dof" as the intial transformation of registration, the registration still fails. I think this proves my inference about the origin-based initialization. So my solution to this problem s to enforce the origin of the nifti images to be aligned before I do anything with irtk, which works well.

Thank you very much for advice anyway! It's a really nice discussion here.

Zhoubing