Research On Medical Image Registration Algorithm Based On Multi-feature And Dual Constraints

Medical image diagnosis plays an important role in the fields of computer-assisted diagnosis and clinical surgery.At the same time,it is also a challenging task.As an important tool to assist doctors in medical image diagnosis,doctors’ requirements for medical image registration algorithms are increasing.In this paper,We take medical images(mainly retinal images)and propose a medical image registration algorithm based on multi-feature and dual constraints.In this paper,We use point-by-point Euclidean distance as the global geometric structure feature,and EOH-SIFT and vector-based geometric feature as local geometric structure feature.Among them,the EOH-SIFT feature descriptor can describe the contour information of the image,which avoids the problem of insufficient feature description due to the lack of contour information in the image.Vector-based geometric structure feature can solve problems such as the inability to distinguish or incorrect recognition of similar geometric structures to a certain extent.By introducing local geometric structure feature,combining them with global geometric structure feature and EOH-SIFT feature to form multiple features,the hybrid features proposed in this paper make the feature description more sufficient.It is conducive to subsequent correspondence estimation and spatial transformation updating,and also makes the process of image registration robust and accurate.This paper uses the popular EM algorithm optimization framework in the process of the correspondence estimation and spatial transformation updating to estimate the parameters involved in the above two steps.In the step of space transformation updating,we minimize the expectation of the energy function in the Reproducing Kernel Hilbert Space(RKHS).This process is achieved by derivation.In order to ensure and improve the stability and accuracy during the process of spatial transformation updating,we propose a Local Geometric Structure Constraint(LGSC)to constrain the local geometric structure between point sets,while using the global geometric constraint term which is based on Motion Coherence Theory(MCT)to jointly accurately complete the process of spatial transformation updating,and then realize the medical image registration.We mainly tested our algorithm on a variety of different retinal images and compared it with five popular algorithms.The algorithm in this paper shows the best performance in most cases.