Recently, someone asked (via e-mail) about the registration optimization method used in AFNI. The method is outlined in the paper

RW Cox and A Jesmanowicz.
Real-time 3D image registration for functional MRI.
Magnetic Resonance in Medicine, 42: 1014-1018, 1999.

You can find a PDF copy of this at [afni.nimh.nih.gov] (along with some other papers). I believe the method actually goes back to Ref 10. It is very similar in concept to the method used in SPM (Ref 4).

The details are important, of course. I cannot speak for SPM or other programs. AFNI uses a weighted least square criterion to measure how well the rotated image matches the base image. The weights can be input by the user, but by default they are a smoothed version of the base image. The purpose of this is to give little weight to the non-brain parts of the image. Also, AFNI allows 2 phases to the optimization; first, with the base and target images smoothed, and second with unsmoothed images. The first phase allows for a more rapid convergence to the region of the correct optimum; the second phase (slower) gives a better result. And of course AFNI lets you set the convergence criterion. In the past, people have commented that AFNI's registration has an accuracy "floor" -- this is because they did not use the program correctly and set the convergence criterion to what was appropriate for their tests. Alas, most people don't understand such issues.

bob cox