Feature Based 2D-3D Spine Registration Method

With the development of minimally invasive surgery technology and the continuous development of surgical robots,the accurate determination of surgical targets is a key factor in the success of surgery.Therefore,in order to ensure surgical accuracy,it is necessary to register preoperative 3D CT image data with intraoperative 2D X-ray image data to determine the patient’s posture.This paper studies two registration methods for spinal image data of two modalities in the spinal surgery scenario,aiming to improve the accuracy and speed of existing methods.In 2D-3D registration,the diversity and complexity of data and the time-consuming nature of multiple iterations of traditional algorithms are the two aspects that most affect registration accuracy and speed.In response to these two problems,this paper has carried out the following work:(1)Two solutions are proposed for the complexity of human data taken in real scenes and the particularity of medical image data.In response to complex situations such as low quality human XR image data,low contrast,and submerged skeletal structure,in order to improve robustness of feature point detection and matching,CT values of XR image are segmented(uniformly divided into three segments in experiments)for processing to improve contrast for each segment image data.Finally all feature points are merged and collected resulting in a 100% success rate for feature point detection and matching.For intraoperative XR images with metal objects(surgical instruments)while preoperative CT data has no metal objects a solution is proposed to extract metal object masks and superimpose them on CT images to reduce impact on similarity measurement resulting in successful improvement in stability.(2)Combining iterative optimization algorithm(particle swarm algorithm)with deep learning feature detection and matching algorithm to complement each other and improve the overall speed and accuracy of the algorithm.Specifically including three aspects:Firstly,using feature point detection and matching algorithm for coarse registration,outputting the initial solution of particle swarm algorithm.Coarse registration does not have a complex iterative process and does not require an accurate solution.It can fully utilize the advantages of deep learning networks.In experiments,the time for coarse registration was shortened by about 88.9%.Secondly,optimizing the position and speed update module of particle swarm algorithm with feature point detection and matching algorithm.According to the coordinate rules of feature points output by the network,calculate the translation deviation between XR image and DRR image,thereby directly optimizing displacement parameters,greatly improving the convergence speed of displacement parameters and reducing the overall calculation amount of the algorithm.Thirdly,involving feature point detection and matching algorithm in the design of similarity measurement to improve the stability of similarity measurement between different modalities.