What you are describing sounds to me like an extension of "voxel based morphometry", the subject of John Ashburner's PhD dissertation. There are some tools in SPM for this, but as far as I know, no one has tried to do this in AFNI by clever combinations of the basic tools such as 3dRegAna. You could possibly use a random-effects ANOVA as well. You'd have to convert each raw MR image type to something more structurally relevant, such as GM/WM density, diffusion tensor fractional anisotropy, etc. Probably then blur these measurements (as is done in VBM) to allow for inter-subject variability.

Speaking of diffusion tensors, we now have a program 3dDWItoDT for computing the diffusion tensor from diffusion weighted images. At present, this program does the linear solution for the DT (i.e., log(I(q)/I(0))=-bD), but we'll soon be upgrading that to solve the nonlinear least squares problem directly. The nonlinear solution works better in areas with highly anisotropic D, which of course are the most interesting areas. In particular, the method used to solve the nonlinear problem always gives a positive-definite D, which the linear method does not.

I'm working on getting a summer intern or two in here, and the GM/WM segmentation issue is one that I'm thinking of assigning them.