Research On Multiresolution Nonrigid Registration Based On Improved Mutual Information

Research has proved that early detection plays a very important role in cancer treatment and can obiviously reduce the death rate of breast cancer. Computer-Aided detection (CAD) based on mammograms is a popular and hard work. In the current, the majority of research into CAD has focused on analyzing a single image. In the clinical practice however, mammograms are often interpreted by analyzing corresponding mammographic studies of the same patient. This technique is termed"comparative analysis"when used in CAD.The comparative analysis based on multiple mammograms is appealing, but it also required a number of difficult preprocessing steps and registering two mammograms that are not exactly symmetric in shape and size is the key process. There are two main motheds, control points-based method and similarity-based method are used in this application, but they all face difficulties. In mammogram, except nipple which is not easily detected automaticly, there are no other obvious control points which can be directly used to compute nonrigid spatial transformations; The similarity functional may have many local opitima when is used to in a nonrigid transformation whose number of parameters is large. Thus, registering the mammogram exactly and automaticly is a challenging work.In this paper, a CAD systerm based on mammogram is presented and the research that has done in this field is summarized systemicly. Base on previous work, this paper presented a automatic nonrigid registration method combined with improved mutual information and wavelet-multiresolution analysis (named multiresolution nonrigid registration in this paper). This method computer control points from local registration which utilizes similarity-based method and don't need select control points in advance. Base on these points, it computes global nonrigid transform parameters.Registration based on general mutual information model only takes into account only the relationships between corresponding individual pixels. As a result, much of the spatial information inherent in images is not utilized. In the paper, four inproved mutual information models are summarized. They are normalized mutual information (named MI in this paper), mutual information combined with gradient information (named GMI in this paper), mutual information of regions (named RMI in this paper) and mutual information based on grey-level co-occurrence (named COMI in this paper). Then a novel mutual information based on wavelet is presented and three kinds of wavelets are used for comparision. Compared with the forenamed mutual information models, it is proved that our method is superior than the GMI and is robust. This new method is ultilized in the local registration of forenamed method.