Research On Multimodality Medical Image Non-rigid Registration Based On Mutual Information

Medical image non-rigid registration, which plays very important role in medical image information fusion, aided diagnosis, surgical planning, surgical navigation, and so on, is one of the most important technology in medical image processing. With the rapid development of medical imaging technology, many kinds of medical image have been put into clinical application. So it is crucial to investigate multimodality medical image non-rigid registration. To obtain more accuracy results and reduce computational complexity, image similarity measures and regularization technology for non-rigid registration have been intensively studied in this dissertation.Mutual information is the most important similarity measure in multimodality image registration. However, it is time expensive. To overcome this problem, a fast estimation method based on fast Gaussian transform has been proposed and a new truncated error for fast Gaussian transform was proved. The new algorithm has linear time complexity and high accuracy to estimate image mutual information.Most MI-based registration techniques fail to take image spatial information into account and assume that a global statistical relationship exists between a pair of images. In order to resolve these problems, a similarity measure based on local mutual information has been presented. The image space has been divided into many subspaces and the local mutual information has been estimated in each subspaces. The new similarity measure is the weighted sum of all local mutual information. To get rid of low contribution subspace for registration, an adaptive region selection algorithm base on local entropy has also proposed. Experiments show the adaptive local mutual information reduces50percent computational time and has almost no loss of registration accuracy.As non-rigid image registration is an ill-posed problem, the regularization is essential. However, most regularization methods based on physical model are the same in whole image space, which results in an unrealistic deformation and may be over-smoothed in some place. To tackle this limitation, a new adaptive regularization non-rigid image registration algorithm based on nonlinear diffusion has been presented. The local information of image is incorporated into the mathematical framework of non-rigid registration and the diffusion coefficient is chosen to vary adaptively according to image content. So the deformation can be smoothed selectively to preserve the local detail of deformation. In order to reduce time complexity, the additive operator splitting method has been adopted to solve the partial difference equation. Experiments using synthetic images and MRI images show the new method achieves better performance than the non-adaptive regularization methods.Many regularizations methods can cause trouble when large deformations are present. To solve this problem, this dissertation proposed a new method called homeomorphic non-rigid image registration. The necessary condition for homeomorphic map over continuous space and discrete rectangular grid has been analyzed and the image registration has been converted into a constrained optimization problem. The constrained optimization is solved by interior penalty function method. Experiments using artificial and magnetic resonance images show the new method gives better registration results than the no constrains one.