Research On Algorithms For Medical Image Rigistration Based On Mutual Information

With the development of medical image engineering and computer technology, medical imaging has become one of the most important parts in modern medical treatment. There are medical images of multimodality in clinical diagnosis that are obtained by different imaging principles, such as Computerized Tomography (CT), Magnetic Resonance Imaging(MRI), Single Photon Emission Computed Tomography (SPECT), etc. These images contain different information about human viscera and pathological tissue. In order to better diagnose, information obtained from different images should be integrated.The problem of image registration must be solved before the quantitative analysis of different images. The task of medical image registration is to find a geometric relation or transformation among the same regions of interest in two or more images. The effect of image registration directly affects the accuracy of diagnosis. So the exact registration is necessary for doctors to obtain the exact diagnosis results.Registration is one of the most important issues in medical image area; it is the combination of information science, computer image and medical areas. The method of registration can be divided into two categories, gray value based method and trait based method. This author focuses on mutual information that belongs to the gray value based category and makes some improvements on the algorithms.Based on theories of wavelet transformation and mutual information, the author combines simplex algorithm and PSO algorithm. Also, some improvements on artificial immune algorithm is made and then used for image registration. Additionally, the author puts all the results into usage for both monomodality registration and multimodality registration. Exactly speaking, followings are jobs done:Firstly, the concept and property of mutual information are introduced. Also, the author explains the theory for its usage in image registration. Concerning its shortcomings of prone to local minimum, normalized mutual information is introduced. As the later one is not so sensitive to the overlap area, it is more robust when used for registration. Test result shows that normalized mutual information reduces local minimum to a large extent. So the author uses it as a metric of similarity in the following chapters.Secondly, wavelet transform is introduced to image registration. Wavelet transform is introduced into registration due to its multi-resolution property. When the low-resolution images are registered, a hybrid algorithm of Particle Swarm Optimization and Simplex is used. And when registration is made on higher resolution images, the author uses only Simplex. PSO excels in global searching while Simplex excels in local searching. When used together, a more accurate result can be got. Meanwhile, a smaller searching space can be reached by using wavelet transform that speeds up registration. A better result is shown in the experiment.Thirdly, an algorithm, which is based on normalized mutual information and improved artificial immune algorithm, is proposed. The algorithm takes advantage of Simplex to produce initial solutions, which provides a faster way to get the final results. What is more, it takes the gravity of images as the immune, and the performance is improved a lot. The results show that the algorithm is fast and accurate.Finally, a conclusion is made and the future research directions in this field are proposed.