Based On Statistical Models Of Multi-modality Medical Image Registration

The need for new approaches to image analysis has become increasingly important and pressing as advances in imaging technology enable more complex objects and processes to be imaged and simulated. The medical image registration technique is an important discipline of medical image processing which has been developing rapidly since 1990's. It is widely used in modern clinical application and medical research.Firstly, this thesis introduces the present situation of the research and methods of medical image registration. The major classes of algorithms and relative techniques are reviewed and several specific examples of each class of algorithm are described. This thesis concludes the questions about the technique of image registration and analysis in clinical using. These questions also tell us the direction of our technological development.Secondly, the thesis analyzes several statistical models of image based on statistical knowledge in detail. Then, we propose our way of voxel-based multi-modal image registration on the strong theory base. We compare the traditional maximum likelihood metric and the new mutual information metric and find the new metric is the farther development of the traditional metric. To inspect this way of registration, we realize our method using the non-parameter model, Parzen Window. In order to avoid the disadvantage of large computation, the multi-resolution optimization strategy is used for this registration algorithm. This strategy accelerates the speed of the computing clearly. Experiments are presented that demonstrate the approach registering multi-modal biomedical image using the data base of Vanderbilt University, USA. The results tell us this method is very suitable for multi-modal images and has high precision and robust. We also design a multi-modal image fusion prototype. Using this, we give excellent example to describe the function of image registration in clinical environment.Finally, we present some suggestions of future work and point the statistical learning theory will be a very promising direction in this field.