The Research On Mutual Information Based Multimodality Medical Image Registration Methods

The registration methods of multimodality medical images provide doctors the possibility to utilize the information of different modal medical images from the same patient, so the accuracy of medical diagnosis was improved greatly. The mutual information (MI) based registration method is of high accuracy and robustness without the need for preprocessing of images, Hence it represents the trend of registration. This thesis mainly deals with MI based multimodal medical image registration methods.This thesis firstly introduces the conception and process of registration. Then the major registration algorithms and categories are described. The knowledge about MI and the affection of interpolation, outlier strategy and grey levels was discussed in detail. Unfortunately, the MI based registration has lots of disadvantages: MI is computing expansive, the registration process is slow, and the mutual information function is generally not a smooth function but one containing many local maxima, which has a large influence on optimization.Secondly, the intelligent optimization algorithms are used as search strategy. To overcome the genetic algorithm (GA)'s premature convergence, the Metropolis scheme was introduced into GA. The accuracy and robustness of experiment using GA, ant colony algorithm (ACA) and genetic-simulated annealing hybrid optimization algorithm was improved preferably. The experiment results and analysis of data show the methods proposed are satisfactory.Finally, to overcome the shortcoming of computing expansive, we applied wavelet-based multi-resolution approach in registration. Two novel registration methods were proposed: a method using hybrid GA and Powell's method and a method combined ACA with Powell's method. We register the sub-images decomposed by wavelet transformation using GA or ACA, then the results were applied as the start point for the registration of fine-image using Powell's method. These registration methods could efficiently restrain local maxima of MI, speed up the registration process and improve optimization ability in local search. Experiment results show subvoxel accuracy can be achieved and the feasibility and efficiency of algorithms are verified.