Research On Algorithm For Brain Magnetic Resonance Image Registration Based On Maximization Of Mutual Information

In order to obtain the different functional information and anatomic information, manifold imaging modes are always used in the existing clinical diagnoses and surgery plan. To get the complementary information of the multimodality images, the registration and fusion of the images is needed. To the inner modality images from the same case in different time, the change of images information should be obtained by the image registration and fusion. In image registration, at first, the two images should be put the same coordinate system, so that the homologous anatomical points are mapped onto the same spatial position. Because the human brain can be treated as the rigid body, so, in this thesis, the spatial transformations of brain are restricted to 2D rigid transforms. The spatial position can be achieved through one rotation parameter and two translation parameters.Usually, the procedure of the algorithm of the medical image registration is shown as follows. First, you should choose the feature space and search space, a similarity measure and a search strategy should be determine in this step. In second step, the initial spatial transformation should be obtained and the float image should be transformed using this transformation in the coordinate system of the reference image. Using the interpolation algorithm to the float image and getting the value of the similarity measure through the formula are the tasks in the third step. The final step is to use some optimization techniques optimizing the similarity measure to the best global values. In these steps, the choice of the similarity measure and the search policy is the most important. In this thesis , the mutual information is chosen as the similarity measure, many search strategies is used to the registration and we compare and discuss the results of the different search strategies.Among the search strategies , the Powell algorithm is used in many cases. Through the deeply study to the Powell algorithm, we find that this algorithm always get the local best value not the global best value. Given a random initial spatial transform, the registration results is always not very accurate in this algorithm. So the Cross-Weighted Moments Algorithm is introduced to the image coarse registration. The Cross-Weighted Moments Algorithm borrow the conception of the quality distribution of the rigid body in the classic mechanics. First ,the center of mass and principal axis of the two images should be calculated, then align the center of mass and principal axis of the two images through the translation and rotation, so we can register the two images in this way.With the Cross-Weighted Moments algorithm, the results of the registration always reach the sub-pixel level accuracy. But the registration time is still too long. To improve the registration efficiency , the local search algorithm is proposed to replace the Powell algorithm. The local search algorithm reduce the calculate time remarkable because it avoid the linear search which consume time most in Powell algorithm and avoid plentiful repeated calculation. The experiment show that the mixed algorithm of the Cross-Weighted Moments Algorithm and local search algorithm base the maximization of mutual information can reach the sub-pixel level accuracy and enhance the registration rapidity obviously.