Localized volume matching for the detection of relative displacements in CT or MR images

Our goal was to develop a volume matching (registration) method which is able to manage the failures in the optimization process that resulted from the small size of the target region. The volume matching is applied locally, in a separated region of the scene. It can be used for detecting 3D alignment between a non-movable part and a movable or deformable part of an anatomical object seen on different image scans. The mutual information (MI) registration of different CT or MR scans to a reference scan is based on matching a separated (usually small) target region (subvolume) which represents the non-movable part of the whole volume. The accuracy given outside of the registration volume is critical. Therefore, the accuracy within the subvolume is increased to subpixel range by applying appropriate interpolation methods to estimate voxel intensities. The small size of registration subvolume can destroy the convergence of numerical optimization. We propose a method to minimize this effect by introducing special correction steps in the iteration procedure.