Research On Multimodal Medical Image Registration Algorithm And Its Application

Non-rigid registration of multimodality medical images has important research value in medical diagnosis, operation navigation, evaluation of postoperative effects and so on, receiving more and more attention.This thesis mainly introduces the medical image registration techniques based on Bspline Transform. Image registration based on B-spline is a registration method which is very representative in non-rigid image registration techniques, and the B-spline based registration technique’s computational efficiency and the ability for local control is relatively good. However, this method also has some defects, trying to improve these defects can further improve the registration accuracy.In the single level fixed grid B-spline transform registration, controlling grid’s density directly affects the efficiency and the accuracy of registration. It is often difficult to satisfy both the requirements of efficiency and precision in the single level fixed grid transformation registration. Multi-hierarchy B-spline registration method is used for solving this problem. Firstly set the control grid density relatively sparse which could be used for global rough registration, and then refine the sparse grid to denser grid which could be used in finer registration, so as to improve registration efficiency and accuracy level by level. As the entire control gird of traditional single level B-Spline is uniform, to meet the requirements of local deformation is not easy. The registration results and time consuming mostly depend on the density of the control grid. The regional multilevel BSpline has been proposed to deal with these problems. The overall B-Spline control gird is used to make the results sparse, then the dense control gird is used to refine the results locally. The regional multilevel can ensure the accuracy and the speed by using nonuniform control grid.The similarity or metric of registration is to evaluate the effect of registration, how to choose and define specific measure of registration is significantly important. When there exist some isolated regions in image registration(Isolated regions are those regions which do not correspond to any regions in the other image), the original information about this isolated region should be remained as far as possible in the registration result. Aiming at this problem, the newly registration metric combining the normalized cross correlation with mutual information is proposed in which different weights are assigned to different regions, the regions which don’t match each other is assigned with smaller weights so as to reduce the regional contribution to the whole registration metric value. Through comparing the experimental results obtained with the newly proposed metric and other metrics as the registration metric respectively, the better effect of the newly proposed metric has been validated. Besides, through contrast experiments it is validated that when the Mean Square Difference Metric(MSD) is combined with the definition of spatial distance, higher registration accuracy can be obtained.In order to improve the registration efficiency further, multi-resolution strategy is used in registration process, the principles and ideas about this strategy are discussed in detail. Basing on the B-spline transform, multi-resolution strategy registration experiments were done with the same mode and different mode of images respectively, meanwhile analysis about this validation experiment results are given.This thesis presents the design and implementation of a medical image processing system which mainly includes the user interface module, segmentation, registration module, fusing module and 3D modeling module. The registration module is developed with the general medical image registration framework mentioned in this thesis based on the medical image algorithm library ITK, the realization of the algorithms mentioned in this thesis is also done with ITK.