Medical Images Registration Method Integrating Features And Multiscale Information

Medical image registration was the key technology of modern medical image processing.And it's the basis and precondition for the fusion of medical information. The current researchesmainly focus on the speed and accuracy. This thesis has studied two types of technologies ofmedical images registration, and based on these technologies, a new method was proposed formedical images registration.The method based on images' features started with the features, and they were matched byemploying the similarity measure and one-dimensional search algorithm, then the imageregistration's parameters could be obtained. It's an effective method of medical image coarseregistration for that the number of feature points was far less than the number of image pixelpoints. The registration method based on wavelet transform was achieved by using images'feature information in different levels, in addition, the similarity measure and the optimizationalgorithm were employed at the same time. Then the registration parameters were obtained at last.The medical image registration method integrating features and multiscale informationincluded two parts. They were the coarse registration based on images' features and the fineregistration based on wavelet transform. In the coarse registration procedure, Harris algorithmwas employed to detect corners of the couple images, and Canny operator was used to extract theedge information at the same time. Then the centers and the translation data could be obtained byusing edge information. The rotation parameter could be calculated by using the principal axes.Thus, the first coarse registration was completed. In the following, the corner points detected byHarris algorithm were used to accomplish the second coarse registration. In the fine registrationprocedure, the couple images were decomposed by wavelet transform at first. Thenlow-frequency sub-images were selected as the registration objects. The parameters obtained bythe coarse registration were use to initialize the initial searching direction of Powell algorithm,and mutual information was selected as the similarity measure. The registration of the coupleimages was started from low-frequency sub-images of the last level of wavelet transform. Thenlow-frequency sub-images of the first level of wavelet transform were registered. At last theinitial images were selected as registration objects. Thus, the whole registration procedure wasaccomplished, and the registered image could be obtained.Experiment results showed that accuracy registration results could be obtained by using theproposed method for that features and multiscale information of the couple images were usedeffectively. And the couple images were registered from the coarse registration to the fineregistration. So the error of registration results of the proposed method could achieve the sub-pixel level. And the accurate registration results could also be obtained under noisyenvironment.