Using points and surfaces to improve voxel-based non-rigid registration T. Hartkens, D.L.G. Hill, A. D. Castellano-Smith, D.J. Hawkes, C.R. Maurer, Jr. , A. J. Martin, W.A. Hall, H. Liu, C.L. Truwit Computational Imaging Sciences Group, Guy´s Hospital, King´s College London,UK Department of Neurosurgery, Stanford University, Stanford, CA Department of Radiology, University of California San Francisco, CA Departments of Radiology and Surgery, University of Minnesota, Minneappolis, MN Abstract Voxel-based non-rigid registration algorithms have been successfully applied to a wide range of image types. However, in some cases the registration of quite different images, e.g. pre- and post-resection images, can fail because of a lack of voxel intensity correspondences. One solution is to introduce feature information into the voxel-based registration algorithms in order to incorporate higher level information about the expected deformation. We illustrate using one voxel-based registration algorithm that the incorporation of features yields con- siderable improvement of the registration results in such cases. 1 Introduction Increasing number of studies focus on detecting temporal anatomical changes in the brain by non-rigidly registering tomographic images, e.g. [2][15][14]. The resulting de- formation field is used to quantify the volume change or the displacement of the tissue. The applied registration algorithms can be divided into feature-based approaches which use point, curves, and surface information to drive the registration, or voxel-based ap- proaches which operate directly on the image intensities and define voxel similarity measure to compare the images. While feature-based registration algorithms can re- liably align anatomical boundaries and therefore can quantify the change of certain anatomical structure with high precision, voxel-based registration algorithms use the intensities throughout the whole images and therefore yield deformation values based on the image content also in regions where it is difficult to detect distinct features. We demonstrate in this paper that incorporating feature information in a voxel-based regis- tration algorithm combines the advantages of both approaches. The voxel-based non-rigid registration algorithm with which we have the greatest experience works well on a wide variety of data including pre-and post contrast MR mamograms [11], serial MR images of the brain [7], and pre-and post resection brain images [6]. However the algorithm sometimes fails, especially when there are large changes between the images. In particular, in a series of 24 pre- and post- resection MR images of the brain, we found that 3 clearly failed on visual inspection to align corresponding features.