Correspondence of closest gradient voxels--a robust registration algorithm.

Abstract

A robust, automatic volume registration algorithm based on intensity gradients is presented. This algorithm can successfully perform registrations under conditions of unrelated intervolume voxel intensities, significant object displacements, and/or significant amounts of missing data. It also allows the user to visualize the registration convergence, clearly illustrating any source of registration errors. This algorithm consists of a matching algorithm based on iteratively finding the correspondence of the closest voxels containing a high three-dimensional intensity gradient magnitude. This algorithm was tested by registering T2-weighted MR volumes that had undergone varying displacement transformations to simultaneously acquired proton-density volumes. These transformations involved rotations of up to 25 degrees followed by translations of up to 25 mm along the axis of rotation. For all registrations, the mean registration error was less than one-fifth of a voxel and the mean registration time was less than 30 minutes. In conclusion, this algorithm is shown to be a powerful method of sequence-independent MR volume registration that is simple to both use and understand.