We first consider the problem of registering PD images to transformed
T2 as a means of testing the registration algorithm. Each T2-weighted
scan was rotated by
and translated 5 cm and then
registered with the corresponding PD scan. Since these T2 and PD
images were acquired at the same time and are originally in the same
coordinate system, by perturbing one scan by a known amount, we have
accurate ground-truth upon which to validate the registration process.
All 36 test cases registered with sub-voxel, and in most cases with
sub-millimeter accuracy. Note that the method has a large region of
convergence and thus does not require the starting pose to be very
near the correct solution.