Research On 2D/3D Image Registration Technology For Minimally Invasive Spine Surgery Navigation

Surgical navigation uses a wealth of medical imaging information to achieve rapid and accurate positioning of the patient’s anatomical structure during surgery and lower the surgery risk.It has become a popular technology in the field of medical surgery today.However,due to the limitations of operating room space and imaging conditions,it is difficult to perform real-time 3D imaging during surgery.Therefore,to achieve intraoperative 3D navigation,it is necessary to register preoperative 3D images(such as CT)with intraoperative 2D images(such as X-rays).That is,2D/3D registration.The essence of 2D/3D registration is to use optimization algorithms to align the two images with the preoperative 3D floating image after dimensionality reduction and the intraoperative 2D reference image,so that the two images have the highest degree of similarity.This process involves projection algorithm,space transformation,interpolation algorithm and similarity measure.Based on the projection strategy,this paper conducts in-depth research on the registration of 2D/3D medical spine images.The main research content is three aspects:(1)A 2D/3D hierarchical registration method based on deep learning and parameter optimization is proposed.First,deep learning is used for coarse registration,and then the parameter optimization method is used for fine registration of the two vertebrae,which not only ensures the overall deformation of the rigid body,but also incorporates local deformation information.And based on the LIDC-IDRI data set,the method is experimentally verified.The experimental results show that the method has an improvement of 0.19 in the normalized cross-correlation index and an increase of 0.45 in the normalized mutual information index compared with the benchmark experiment based on Elastix.High precision and robustness.It proves the innovation,feasibility and effectiveness of the algorithm.(2)A 2D/3D registration method based on a generative confrontation network is proposed.First train a generator G that can generate a relative deformation parameter that meets the joint constraints between the vertebrae,and then optimize the generated deformation parameters.And based on the LIDC-IDRI data set,the method is experimentally verified.The results show that the method improves the normalized crosscorrelation index by 0.13 compared with the benchmark experiment based on Elastix,and increases the normalized mutual information index by 0.43.Higher registration accuracy.It proves the innovation,feasibility and effectiveness of the algorithm.(3)Designed DRR simulation software system.The DRR simulation software system not only has high-precision DRR projection effects,but also facilitates users to adjust real-time spatial transformation parameters,greatly reducing the time of DRR image generation,and is a set of high-precision,high-efficiency,and high-flexibility DRR generation software.