Medical Image Registration Based On Normalized Mutual Information

Medical image registration is an important branch of medical image processing, which can provide more extensive information on medical images for clinical application and research. Registration is mainly to find the best transformation between two images, which makes an image and another image to achieve the same position in space through some transformation. Temporary medical image registration method can generally be divided into two categories:one is feature-based registration, which is simple, but the accuracy depends on the feature extraction step. The other is based on the calculations of pixel gray, which does not need pretreatment and the accuracy of the registration results is relatively high, but the calculation is large. Medical image registration based on mutual information method belongs to the second category, which is important for current medical image registration research. The research of this thesis is based on mutual information registration method. The main work is as following:The first and second chapters introduce the background of the medical image registration and significance of the study, as well as the current research. Medical image registration concepts, principles and specific implementation steps are described, including four important parts of the registration:the geometric transformation, interpolation method, similarity measure and optimization algorithms. The direct impact of these four factors to the results of the registration is investigated and the classification of registration method is mentioned.The third chapter discusses the medical image registration based on mutual information method. First the principle of mutual information and calculation methods are introduced. The principle and processes of the registration based on mutual information are explained and matters needing attention were analyzed. Secondly, to reduce the impact of the image pixel gray background on registration, a calculation method for improving normalized mutual information is proposed. The information of focus region in medical images is emphasized and the weight of background information in the joint histogram is reduced. The redundancies in mutual information formula is eliminated by analyzing calculation formula and the computing is more concise. The fourth chapter discusses the basic principles of optimization algorithms in image registration, classification and scope of the algorithm, focusing on a modified Powell algorithm. First, the chapter introduces local optimization algorithm and global optimization algorithm representative, such as Powell algorithm and particle swarm optimization. Their principles and operating procedures are introduced. Secondly, an improved Powell algorithm is proposed, which can obtain Powell initial registration parameters. This method reduces the numbers of iterations and the probability of the Powell optimization algorithm into local minima, meanwhile the algorithm stability is improved. Experimental results show that the registration combining improved Powell and normalized mutual information can improve the speed and stability.Finally, there is a research about the visualization of medical image application algorithms. In VC++2008platform, using VTK as visualization tools, ITK registration algorithm as the framework and organization package, MFC as the design of software interface, an image visualization platform is implemented, including medical image reading, display and registration.