Research On Medical Image Non-rigid Registration Method Based On Improved SIFT Feature

Medical image registration is to find an optimal transformation between two medical images, making an image spatially aligned to the other image through the given transformation. In the medical field, registration is mainly used in medical image fusion of CT, MR, PET and other medical images, comparison of actual medical images and the map, surgical navigation, cardiac motion estimation and many other aspects. In this paper, this hotspot is studied.If there are several similar regions in the whole image, feature descriptors, which is achieved by the traditional SIFT (Scale Invariant Feature Transformation) algorithm, will be very similar and cause a large number of mismatching points. In order to overcome the shortage which the traditional local feature matching algorithm caused because of the sensitive to noise and gray scale nonlinear transformation, this paper proposes an improved SIFT feature extraction and matching algorithm. First, Harris corner detection operator is used to extracte feature points, and then SIFT feature descriptors are obtained by the the gradient histogram of the pixels around feature points. Improved SIFT algorithm adds the global texture information, making SIFT feature descriptors contain the context of a larger neighborhood, and decreases the mismatching probability as a result of similar local information.In the paper, non-rigid registration algorithm is proposed, based on the hierarchical model which combines affine transformation and thin plate spline model, using affine transformation as the global coarse registration, and thin plate spline model as the local precise registration.On that basis, the paper gives an integrated non-rigid registration algorithm based on improved SIFT feature:First of all, feature extraction is carried out separately to extract features of the reference image and the float image, and improved SIFT feature descriptors are generated; the hierarchical model is chosen, combined affine transformation to realize global coarse registration with thin plate spline to realize the local precise registration; Euclidean distance and Arithmetic-geometical mean distance are used as the similarity measure, ultimately non-rigid medical image registration is implemented. Non-rigid registration software platform is developed based on OpenCV (Open Source Computer Vision Library), and a visual experimental interface is given. In the paper, the registration experiments of multiple CT, MR and PET images are realized by the developed image registration system. In feature extraction, contrast experiments are done between the traditional SIFT feature extraction and matching method and the improved method in this paper, the results show the superiority of the proposed algorithm; In the process of image registration, a few of objective evaluations, such as COEF, MSE, NMI, SNR, are used to evaluate the effects of the registration. The objective evaluation results show that this proposed method in this paper get better registration results.