With the continual development of medical imaging and computer science, medicalimage registration technique plays an increasingly important role. Among variousregistration methods, registration method based on mutual information (MI) does not needto preprocess images and is of high accuracy and robustness. Hence it has been used inmedical widely.We study deep in medical image registration based on MI in this paper, describe themethod of MI in detail and make effective improvements.The main works in this paper areas follows:Firstly, we describe the background, implication and development of medical imageregistration. And then the theory and detailed steps of MI method also are analyzed,including the transformation model, the interpolation method and the optimizationalgorithm.Secondly, it is unreasonable because conventional MI method only considers statisticsinformation of gray and neglects spatial information of images. We think pixels in differentlocations have different contributions in registration. So we developed the spatial weightedmutual information (SWMI) metric by combining the saliency value (SV) of each pixel,and use this metric to align the images. Experiments show SWMI-SV method is morerobust and accuracy than conventional MI method.Finally, Researchers show that partial volume (PV) interpolation in MI method caninduce local maximain mutual Information function. Firstly, we analyze the reason indetail and then develop an improved PV interpolation method. Our interpolation uses anew trigonometric function as its interpolation kernel function and counts nineneighbothood pixels. Experiments show this interpolation can effective to avoid localmaxima in mutual information function. |