Research On Multimodality Medical Image Registration Algorithms

Medical image registration is one of the most important research topics of medical image processing. Multiple medical images are registered and fused by image processing techniques and the result image provides more information for medical analysis and treatment. Although registration is highly appreciated and widely applied in medicine, existed registration algotithms have many disadvantages, the accuracy and speed need to be improved. Based on basic registration theories, this paper discusses recent registration techniques, studies and improves multimodality medical image registration algorithms.Multual information is a popular criterion to register rigid objects. A global optimization method based on normalized mutual information is proposed for multimodality medical image registration. First, external surfaces are extracted from various image modalities and the Levenberg-Marquardt ICP algorithm is adopted to initially align unregistered images. Then the registration is performed by maximization of normalized mutual information using a deterministic global optimization algorithm named Dividing Rectangles. The surface based matching is used to provide a good start point for Dividing Rectangles in order to fully utilize its high efficiency in small search space. The results of experiment on Vanderbilt university three dimensional multimodality human brain data show that the algorithm achieves subpixel accuracy and avoids local minimums efficiently. It is faster and more accurate compared with a local optimization and a global stochastic optimization registration.A new elastic registration model based on spring mass system is proposed concerning local deformation. Source image is Delaunay triangulated, and then modeled as a spring mass system. The system deforms under the influence of forces derived from gradient of the mutual information registration criterion, while specific geometrical structures are restricted to predefined shapes. When the mesh vertices reach equilibrium, each triangle undertakes a local affine transformation. Experimental results show that the algorithm controls the deformation rapidly and precisely, efficiently registers both intro-modality and inter-modality images.