Research On Mutual Information Based Rigid Registration Of Medical Images

Image registration is widely used in many image analysis and processing fields, and is a classic research topic with scientific significance in image processing. Medical image registration is an important technique in the field of medical image processing and analysis, and plays a key role in medical diagnosis, clinic therapy, etc. Multimodal medical images acquired based on different imaging principles, provide internal organ information from different aspects. Modern medicine usually requires integrating the complementary information contained in multimodal images in order to improve the diagnosis and therapy level. However, before implementing the integrated analysis, the processed images should have been aligned correctly. The internal organs of patients show different in different modals, in addition, positioning, image resolution and contrast are often different, which make it time-consumed and inaccurate for doctors to align related images only basing on observation, and make image registration research necessary. Medical image registration is the process of finding the geometric transformation between corresponding points of the same anatomy location in two images.The author takes on the development of the image and graphics processing system, in which the multimodal image registration module is a practical application of medical image registration. This system is a sub-system of Multi-Leaf Collimator Dimension Conform and Intensity Modulated Radiation Therapy System Project, which is supported by a grant from the National High Technology Research and Development Program of China (863 Program) (No. 863-306-ZD 13-03-6).Robustness, precision and speed are essential for image registration; furthermore, human intervention should be reduced or avoided. Due to the property of needing no hypothesis on image gray, mutual information (MI) is especially suitable for multimodal image registration, and has better robustness and flexibility than other conventional registration methods, which make it one of the research focuses. This dissertation analyses and concludes the fundamental theory, method and work flow of MI based registration method of medical images. Aiming to solve these problems inhered in MI based registration, such as time-consumed, neglect of spatial information, and multi-artifacts, this dissertation is devoted to the study and improvement of MI based rigid medical image registration. The main works are summarized as follows: 1. Aiming to solve the time-consumed problem of conventional MI based registration method, this dissertation proposes a coarse-to-refined hierarchical registration method by introducing macroscopical spatial information. Image gradient describes the gray change among pixels and reflects the shape features of the image in some extent. Shape information such as centroid and principal axis, extracted from the gradients of images to be aligned, is used to calculate the coarse registration results. Then MI based registration is utilized to refinedly align the coarse aligned images. The coarse step can be realized without any optimization searching, at the same time, it can help the refined step avoid suffering from the local maxima of MI because of inappropriate initial transform parameters. Experiment results show that this hierarchical method can reduce time consumed by 97% while achieving close results with conventional MI method.2. To overcome the limitation that MI based registration doesn't contain gray spatial information, this dissertation proposes the concept of gradient similarity, and combines it with MI to generate a new similarity metric (GSMI). Gradient similarities in magnitude and orientation between each corresponding pixel pairs are extracted first, on which the total gradient term is calculated. The product of the gradient term and mutual information is used as the new similarity measure in medical image registration. Experimental results show that the GSMI method is more robust than typical MI because of the combination of gradient information.3. The method proposed above simply multiplies MI by a gradient term, and doesn't expand MI per se. Therefore, a new 2-order MI metric based on gray and co-occurrence gradient is addressed. In this metric, a pixel gray and its gradient information are incorporated to estimate the joint histogram, thus makes it possible for this 2-order MI to reflect image gray and spatial information simultaneously, and expands the connotation of MI in image registration. Statistic results from amount of experiments show that, compared with typical MI and gray level co-occurrence 2-order MI, image registration based on the new metric can get higher correctness rate, and higher precision in those experiments succeeded in all three metrics.4. The reasons why PV (partial volume) interpolation induces local maxima in mutual information function are analyzed, and aiming to this problem, the dissertation proposes an improved PV interpolation method, namely BHPV (Blackman-Harris PV) interpolation. The differences between conventional PV interpolation and BHPV interpolation include two aspects. One is that PV interpolation utilizes pulse function as its interpolation kernel function, while BHPV uses an approximative form of Blackman-Harris windowed Sine function instead, and the other is that the most count of neighborhood pixels involved is different,4 for PV and 16 for BHPV. The differences make BHPV interpolation avoiding the generation of local maxima in mutual information function. Experiments show that BHPV based MI is more robust than PV based MI.These methods above improve typical mutual information from different aspects, and the latter 3 methods can be used in the second phase of the first method, to achieve rapid, robust and correct results. Besides, these methods can collate each other.