Research On Medical Image Registration Based On Mutual Information

Medical image registration is an interdisciplinary research field between Medicine and Information Technology, which is also the basis of information fusion of medical images.The information held by various modalities of medical images, which can by classified into two categories as functional information and anatomical information, may vary as a result of the diversity of imaging methodologies. The purpose of medical image registration is to make the two or multiple images geometrically aligned via space transformation. Image registration methods fall into two categories in general: interior characteristics-based and exterior characteristics-based methods. Mutual Information (MI), which is one of the former kinds, has been recognized as the most powerful registration measure for its distinct characteristics-it requires neither assumption regarding the nature of the relationship between the image intensities in both modalities nor any preprocessing of the data to be registered.Two key technologies-interpolation and sub-sampling of the image data during the registration process is studied in this paper. Local maximum of the registration curve is a serious problem which may deteriorate the accuracy of registration. Based on research on the cause of these local maximum brought about by traditional interpolation methods, a new method based on window function is proposed. A gradient-information based method for sub-sampling is proposed, with invariance on the trend of probability distribution before and after the sub-sampling process. Besides, elementary research on multiple image registration is also carried out-a new MI Matrix-based measure for multiple medical image registration is defined and applied to the registration between two images and among three images. Experiments show that the newly defined measure is robust to noise, shortening of gray scales, intensity in homogeneity and geometric distortion. In order to further improve the efficiency of multiple image registration process, Arithmetic-Geometric Mean Divergence (AGM) is applied to this work. Experiments showed that with acceptable decrease in accuracy, AGM can efficiently improve the efficiency of the registration task.