Research On Medical Image Registration Methods Based On Mutual Information

Modern medical imaging has provided many medical images with different modality (such as CT, MR, PET, and so on.), which provide different information on the human body. Through the image fusion technique, we can obtain one synthetical image which combines the useful information of different medical images together to reflect the information of human body comprehensively. However, image registration must be done before image fusion. The techniques of medical image registration originated in the ninety's of twentieth century, has become one of the most important research fields of medical image processing. In this thesis, medical image registration by mutual information is investigated, and it has been accepted as one of the most accurate, highly-automated, robust, popular registration methods. The main work and contributions in the thesis are as follows:(1) Because the distance and angle between pixels do not change in rigid transformation, we decompose the two-dimensional and three-dimensional rigid transformation formula, calculate the rigid body transformation step by step. The improved formula can effectively reduce the time-consuming of transformation and not affect the registration results.(2) We continue to study the reasons of local extremums of mutual information based on previous studies and find that the local extremums are not only affected by the interpolation algorithm, but also affected by the nature of entropy itself. Combining the PV interpolation with entropy nature, we carried out a detailed analysis about the reasons of local extremums.(3) The concept of weighted entropy is introduced and one kind weight selection is given. This kind of weighted entropy is refered to image registration and medical image registration by mutual information based on weighted entropy being put forward. Also we describe the registration principle of this method.(4) In the field of two-dimensional image registration, we compare registration results of normalized mutual information, Tsallis entropy mutual information and entropy-weighted mutual information in many options. Experimental results show that entropy-weighted mutual information can be better to avoid the impact of local extremums and obtain the correct registration results when using some techniques to reduce the computing time.(5) In the field of three-dimensional image registration, we combined the principal axis method with the mutual information method to achieve three-dimensional image registration, It is not only effectively avoid the local extremums, but also greatly reduce the registration time. First of all, we get the edge of images by Canny edge detect operator, then use mathematical morphology method to connect the edge and pick-up the figures of brain image. Secondly we use the principal axis method to compute the principal axis and the center of mass of the figures, and obtain coarse registration results. Finally, we make use of this coarse registration results as the initial search points of the optimization of the image registration by mutual information method and obtain accurate registration results.