Medical Image Registration Based On Quantitative-Qualitative Measure Of Mutual Information

A fundamental problem in medical image analysis is the integration of information from multiple images of the same subject, acquired using the same or different modalities and possibly at different time. In order to fuse the information from different images, an essential problem, which should be solved firstly, is to align one image to the other images. Medical image registration is to find the geometric relationship between corresponding points in different images. After image registration, all anatomical points and other points of interest in the images should be easily related. Various registration methods have been proposed over recent years. Among them, registration strategy based on maximization of mutual information (MI) has been proved to be a promising method and has been widely used in medical image registration.However, almost all mutual information based registration methods treat the voxels of the images equally, when calculating their mutual information. In fact, different voxels have different characteristic and utilities on image registration. Salient voxels should have higher utility, and hence contribute more to determine the transformation between two images. For example, when measuring the mutual information of two brain images, the white matter (WM) voxels near the cortex should contribute more than the WM voxels inside the large WM regions since it is more effective to match WM voxels near cortex than the inside regions.To incorporate utility information into the image registration procedure, we propose a new information measure, named quantative-qualtitative measure of mutual information (Q-MI), in the view of cybernetics. Then, we propose an image registration method based on Q-MI. To define the Q-MI of two images, we use salient values to represent voxels's significance in the image and also utility in image registration. Moreover, the joint utilities of intensity pairs are calculated from integrating the voxels utilities in the two images. In order to test the performance of the proposed method, we design lots of quantitative experiments using simulate brain images. Experimental results show that compared to MI-based registration method, Q-MI-based registration method has a higher successful rate and the increased rate can achieve more than 20 percent, which indicates the robustness of the proposed method.To assure that the registration method has high accuracy, we propose a hierarchical registration strategy based on Q-MI. In the hierarchical registration strategy, the utility values of voxels are not fixed, and they will be hierarchically updated during the registration procedure, with all voxels contributing equally in the final stage. In particular, the initial utility of each voxel will assigned according to its saliency value; with the progress of image registration, this utility will gradually move towards to one. Thus, by mainly focusing on the voxels (or the regions) with higher utilities in the initial registration procedure, the robustness of registration can be improved. Also, by changing each joint utility to one in the final stage, the sub-voxel accuracy of registration can be retained as that obtained by the conventional MI-based registration methods, because of using MI in the final registration procedure.In this paper, the proposed Q-MI has been validated and applied to the rigid registrations of clinical brain images, such as MR, CT and PET images. Experimental results demonstrate that the registration function generated by Q-MI is much smoother than that by MI, and it has a larger capture range due to the incorporation of the joint utilities of the two images into the Q-MI measurement. Moreover, experimental results also show that hierarchical registration strategy not only improves the robustness of registration method, but also makes it have sub-voxel accuracy.In many applications, a rigid transformation is not sufficient to describe the deformation between two images, thus nonrigid transformations are required. In this thesis, we studied a general nonrigid registration method based on quantitative-qualitative measure of mutual information. In addition, we also derive the analytic expression for the gradient of Q-MI w.r.t the transformation parameter when partial volume interpolation is used. Therefore, the registration strategy can hire gradient-based optimization method. We applied the proposed method to correct the motion between MR breast images.Experimental results show that the proposed method performed well and can reduce effectively the difference casued by the motion of breast.