Intensity-based image registration is driven by a similarity measure that takes both (or more) input images and computes a numerical value that provides information about the goodness of fit. In rigid registration, only rotations and translations are applied for the spatial transform. It is usually necessary to align images from the same subject without any deformations between the acquisitions or as a pre-alignment for registration methods with higher degrees of freedom. The main contributions of our team are comprised of automatic parameter estimation methods for statistical image similarity measures and performance gains by novel optimization techniques. We also focus on using modern graphics cards for accelerating the computation of the similarity measure.

The following example demonstrates the mismatch due to wrong parameters applied to a rigid registration of a CT-PET image pair using normalized mutual information. For the initial starting position (left) shown as overlay visualization, the result using standard parameters (middle) is compared to the proposed approach with optimal parameter settings (right).