Research On Medical Image Registration

With the development of computer technology and biomedical engineering technology, imageregistration has been widely used in the medical field, and the requirements of registration accuracyand speed also will be gradually increased in the disease diagnosis and clinical treatment. Accordingto the characteristics of internal features, the registration method can be divided into two categories:the feature-based medical image registration and the gray-based medical image registration. Thefeature-based medical image registration requires pretreatment and can not achieve automaticregistration. Because of these limitations, the gray-based medical image registration gains widelyattention recently. The registration based on Mutual Information (MI), which is one of thegray-based image registrations, has been widely recognized as the most effective and most widelyused method. This method does not require priori knowledge of the images to be registered, withouthuman intervention. Meanwhile, this method has strong robust, high registration accuracyand it issuitable for multi-modality medical image registration. Therefore it has been widely applied.Firstly, this paper gives an introduction of the basic conception and classification of medicalimage registration and it gives a detailed analysis and instruction of the current research focus, keytechnologies and key issues of medical image registration.Then, we lead to the research emphasis ofthe paper----research on the similarity measure and search strategies of medical image registration.Meanwhile, it also gives a brief introduction of the key technique and idea which may be used inthe paper.Secondly, because of the low accuracy of the similarity measure for the medical imageregistration based on mutual information, this paper constructed the weighted factors for mutualinformation and lead into gradient information. With all the above, we can accomplish a new thesimilarity measure named GWNMI. The similarity measure uses a normalized structure, andimproves the gradient information correspondly. The difference-value between gradient moduluswhich can reflect intensity of gray change makes the similarity measure more effective, so itimproves registration accuracy. From the experimental simulation result, we can see whencomparing to using the mutual information directly or combining mutual information and gradientinformation, this new method has an improved attribute in aspects of the more obvious climaxes ofparameters’curves, the finding of the extreme point, higher accuracy and stronger robustness.Finally, because of Powell algorithm easy to be entrapped in the local extremum, this paper proposes a modified Powell algorithm that makes a judgment when the direction set is updated incase of the degeneration of the search direction. Meanwhile, this paper introduces a multi-resolutionstrategy that uses the result of the low-resolution layer image registration as the initial value of thesearch of the higher-resolution layer image registration. Because Powell algorithm is a local searchmethod, the paper introduced SA (simulated annealing) algorithm to ensure global search, and thelayer of in the top uses the simulated annealing method, other layers use modified Powell algorithm.This scheme can overcome the initial value effects on Powell algorithm, thus accelerating theregistration process, avoiding the local extrema problem and improving accuracy and robustness ofregistration. Experiments show that the proposed scheme can effectively reduce the search time andreduce registration error, and it is an effective and feasible method.