Research On Multi-modality Medical Image Registration Based On Entropy Graph Theory

With the development of biotechnology, image processing technology and computer technology, medical image registration has already became the key technology of modern medical image processing. As the precondition of medical image fusion and other medical image analyses, medical image registration is very important for clinical diagnosis, clinical treatment and treatment effectiveness evaluation. In recent years, it has become one of the most important research topics of medical image processing. In this thesis, on the basis of previous studies, we research and improve the algorithms and technologies of multi-modality medical image automatic registration. The contents are as follows:Firstly, the background, significance and main applications of medical image registration are introduced here. We summarize and analyze the methods of medical image registration which are commonly used, and investigate the basic principles and key technologies of medical image registration in detail. The strengths and the weaknesses of two principal methods based on image pixel and image feature are analyzed. The application and implementation principle of the mutual information theory in medical image registration are deeply researched and their merits and drawbacks as similarity measures are discussed. Then we focused on the Renyi mutual information and its application in the medical image registration. For the computational complexity of mutual information, we study the theory of entropy estimation based on generalized nearest-neighbor graph and use it for multi-modality medical image registration successfully.On the basis of deeply research on the advantages and disadvantages of two principal methods based on image pixels and image features, we combine the strengths of two principal methods and propose an algorithm for the medical image registration with gradient information and generalized nearest-neighbor graph. The algorithm extracts the complementary scale space feature points from images and uses the gradient information around the feature points to describe the feature points like SIFT, then sets the Renyi mutual information as the object function, and employs the generalized nearest-neighbor graph to estimate the Renyi entropy, at last acquires the result through optimization algorithms. The experimental results show that the proposed algorithm can achieve better robustness, higher speed and good accuracy, which is an effective automatic rigid registration method. For the effect of image noise in medical image analysis and on the basis of deep research on the application of image texture in the medical image analysis, we propose a new descriptor which has better robustness to image noise called weighted bi-center-symmetric local trinary pattern. The descriptor combines the advantages of local trinary pattern and center-symmetric local binary pattern. Then we used the new descriptor to describe the feature points in the framework of multi-modality medical image registration based on the generalized nearest-neighbor graph. Finally, the experimental results verify the new descriptor’s robustness to image noise and the robustness and the effectiveness to image noise of medical image registration method using the new descriptor.