Robust Multimodal Medical Image Elastic Registration Using RPM and Mean Shift

Abstract

Landmark matching is widely applied in multimodal medical image registration. In this paper, a novel landmarks corresponding estimation combining robust point matching (RPM) and mean shift algorithm is proposed. At first, corners of images are detected to be the landmarks sets. Robust point matching algorithm is used to estimate the coarse transformation parameters between two landmark sets. Next, mean shift iterations are adopted to search the exact corresponding point positions in the two images based on the edge structure feature. Moreover, mutual information between two local regions and the topology relations between landmarks are analyzed to eliminate mis-matching landmarks. Finally, the deformed image is transformed by compact support thin-plate spline interpolation. Experiments show that the precision in location of corresponding landmarks is satisfied. The proposed technique is feasible and robust shown in the experiments of various multi-modal medical images registration.