External Email - Use Caution
Dear FreeSurfer People,
I am hoping to use Synthmorph in a VBM or DBM like longitudinal analysis where I register timepoints to a template. Typically in VBM/DBM type analyses we decompose the affine and nonlinear transforms. The advantages of the joint registration approach may be that its high performance symmetry lessens the need to create an unbiased middle template of the timepoints. However, it produces a single deformation field comprising both affine and nonlinear deformations.
I thought I could initialize using an affine, but it seems doing so still produces a deformation image comprising both affine and nonlinear transforms.
It seems I have two options. 1) Somehow use the affine matrix to subtract (somehow) the affine component from the deformation field. 2) Perform deform only registration on an affine transformed image, in which case creating a template is probably needed.
Juan Iglesias provided code to me once that suggested I compute the Jacobian of the full field, take the logarithm, and subtract the average (which Dr. Iglesias assured me is the contribution of the affine part). However, I am really unsure if that perfectly removes the 12 dof affine components, not that I doubt his expertise, I'd just like a second opinion :)
On a side note, I just bought a gpu for the lab with 16Gb, vram and I'm really regretting not moving that up to one with 24Gb vram.