Multi-modal Medical Image Registration Of CT-MRI Based On Mutual Information

Medical image registration is a hot research point in the field of image processing in recent years, and the multi-modal medical image registration is a difficult. Because of the fixed image and moving image are from different imaging devices, image acquisition time, pixel size. image resolution and the reference coordinates are all different, and the patient's physiological status at different times and shooting errors caused by body position, that many successful similarity algorithms in single-modal medical image registration can't applied in multi-modal medical image registration.This paper from the interdisciplinary cooperative research project "Knowledge-based Model of Medical Image Registration", collaboration by Institute of Computer Science and Technology of Jilin University and Second Hospital of JiLin University, focus on the multimodality image registration based maximization mutual information, the main data are CT and MRI brain images. Our standard experimental data are from the open source ITK(The insight toolkit for medical image registration and segmentation), and Second Hospital of JiLin University provided some diseased CT. MRI images of real patient, and other images from the Vanderbilt University retrospective image registration evaluation project RIRE.First, we introduce the research background and significance, concepts and the current research situation. Then gives a common framework by the basic process of medical image registration, which is divided into four basic modules by the main issues involved, including geometric transformation, image interpolation, similarity measure and optimization of function, and gives a detailed description of the registration process by the framework, and then discuss the features of the framework, the classification of medical image registration methods, and algorithms of each module respectively. There are two confusing issues in the framework, in contrast with the intuition, the fixed image is mapped to the corresponding coordinates of moving image space after the geometric transformation in the registration framework, and another point is the whole registration process is performed in the physical coordinate system rather than the logic coordinate system of image. The paper then focuses on the theory of maximization of mutual information, mutual information calculation formula, the Parzen window, histogram estimation and the image pixels sampling methods used to estimate the value of mutual information, and the sampling methods mainlv include Full. Grid. Random. Random-Coordinate four methods. A two-dimensional multi-modal medical image registration and fusion experimental system is designed and implemented, the system consists of three parts, including user interface subsystem, registration subsystem and fusion subsystem, and the fusion subsystem is implemented by other group members using the wavelet transform theory. The registration subsystem is implemented according to the above common registration framework, based on the open source library ITK. main achievements are implementation of similarity measure and optimization function two modules, combined with the ITK's Command/Observer model, used to observe every step's statuses of the registration process. User interface using the open source library VTK for data visualization, the interface part is designed and implemented by MFC. and the three subsystems are independent of each other. ITK is an open source basic library dedicated to the registration and segmentation of medical image, which provides a number of basic classes for medical image registration, we use one of the geometric transformation and image interpolation method. VTK is an open source library for data visualization. people usually design and implement medical image analysis software combined with the ITK and VTK. and there are many well-known medical image analysis software is implemented based on ITK and VTK.An important purpose of this paper is to find the parameters' best value of the similarity measure and optimization function module, which impact the registration results and computation time of the mutual information, to fit the most situations. Because most of the time of the registration process is the calculation of mutual information, in order to more clearly understand the operation involved in the computation of mutual information, the calculation process icon and pseudo-code of one mutual information calculation process is presented in the begin of fourth chapter, and a process of mutual information calculation is demonstrated using one most simple example. We first use two simple standard images with size of 221*257 pixels, the moving image is obtained by translating the fixed image (13,17)mm in x-axis and y-axis. According the previous process of mutual information calculation, our experiments focused on the samples of pixels and the number of histogram bins two parameters, we first select as samples by 50% of the fixed image total pixels, and change the number of histogram bins, the experimental results show that the bins take about 32 or 50 make better performance. Then fixed bins to 32 and 50. select samples from 5% to 100% of the fixed image total pixels, the results show that when the sample percent is more than 20% of total pixels, experimental results are all very well. that is. no need to use too many samples, too many samples just increase the calculation time. Based on the observation of experimental process and the analysis of optimization function's nature, we get several critical parameters' value of optimization function, including the max step MaxStepLength. min step MinStepLength and relaxation factor RelaxationFactor. Then use the real patient images, the parameters values are the result of the previous standard experiments, the geometric transformation model is the affine transformation, interpolation strategy is 3 orders B-spline interpolation. And gives the graphs of mutual information and X-axis translation values with the number of iterations changes, experimental results also verify the parameters obtained from the previous standard image experiments.In order to improve the accuracy, finally combine the framework with the multi-resolution strategy, first introduced the basic idea of multi-resolution strategy and the image pyramid. and gives the detailed registration processes using multi-resolution strategy, and the Gaussian image pyramid is used in our system. And then use the same real patients data with the previous experimental, affine transformation,3 order B-spline interpolation strategy, and the multi-resolution layers is setted to 3, experimental results show that for complex images, multi-resolution strategy plays a significant role in improving the registration accuracy, while for the simple brain images may be not obvious, but will increase the computation time.In summary. this paper study the multi-modal brain image registration of CT-MRI, design and implement a medical image registration and fusion system, which can be used for two-dimensional medical experiments, obtained some critical parameters" value of similarity measure and optimization function by experiments, and did experiments using real patients' data.