Elastic Medical Image Registration Using Surface Landmarks with Automatic Finding of Correspondences

Abstract

We propose a novel approach to surface-based elastic image registration which at the same time finds the correspondences between source and target landmarks and determines the global transformation from the source image to the target image. In our approach only the coordinates of the sample points on the surface landmarks are required as free parameters to minimize an energy function. Thus, the number of parameters is kept small in comparison to other approaches which sample the whole image data. The energy function consists of the minimal distance between the transformed source and the target landmarks as well as the bending energy of the transformation. During the optimization procedure, more and more local deformations axe allowed. We have investigated the performance of the approach using 2D and 3D synthetic as well as 2D tomographic image data. We found that image registration is achievable even when source and target landmarks differ largely. The registration result can be improved through an increase of the density of sample points. To achieve a better adaption to, e.g., thin dents, we extended our approach for inclusion of predefined landmark correspondences as well as the inverse transformation using sample points on the target landmarks.