Research On Automatic Spinal Segmentation And Registration Technology For Intraoperative CT Navigation

Minimally invasive technology guided by computer images has a wide range of applications in spinal surgery.The navigation system of minimally invasive surgery combines medical images with computer science,and the surgical images are processed by computers,which can assist doctors in preoperative planning,intraoperative reference and postoperative evaluation.The spinal minimally invasive surgery has a small incision,and the doctor cannot directly observe the bone structure of the patient.In order to enrich the operation vision of doctors,the intraoperative navigation system needs to provide two functions:one is to segment the spinal region in the CT data obtained before surgery,and combine the three-dimensional reconstruction technology to restore the patient’s complete and three-dimensional spine shape,so as to provide reference for doctors;the second is the registration and positioning of the lesion point with the two-dimensional image and the three-dimensional to guide the surgical puncture.Based on the two functions of the navigation system,the main content of this paper are as follows:Firstly,it analyzes the generation principles and image characteristics of CT images,X-ray plain film images and registration intermediate images DRR involved in minimally invasive spinal surgery image navigation,and at the same time studies the basic principles of image segmentation technology and image registration technology.Research on CT image segmentation technology: In order to solve the problem that the accuracy and efficiency of segmentation methods based on CT threshold and boundary cannot meet the needs of clinical applications,this paper introduces deep learning technology into the segmentation task of surgical image navigation system,and proposes a CT image spine segmentation Network structure.The network uses the U-Net convolution-deconvolution structure as the basic framework,and introduces a convolution module connected by layered residuals and a channel attention mechanism to strengthen the propagation of features and enhance the segmentation effect of image detail information.Experiments on real data sets show that the network structure can alleviate the effects of low contrast of various tissue structures in CT images and blurred boundary contours on segmentation accuracy.Research on the registration technology of two-dimensional X-ray images and three-dimensional CT images: Due to the long registration time and low success rate of the existing registration algorithms based on global iterative optimization.This paper introduces the image classification function of convolutional neural network into registration,optimizes the similarity measurement method in iterative optimization algorithm of registration parameters,and improves the optimization algorithm,and proposes a step-by-step 2D-3D registration method of coarse registration combined with fine registration.This method transforms the projection parameters of the CT images collected before the operation to generate a labeled DRR data set and inputs it into the Inception V3 network for training.During the operation,the DRR projection parameter classification model is used to return the X-ray image to be tested.A parameter value is used as the initial search point for iterative optimization,and the multi-parameter local optimization algorithm is combined to iteratively search for precise registration parameters from the initial search point to achieve registration.Experimental results show that compared with traditional registration algorithms,the algorithm proposed in this paper improves the success rate and accuracy of registration and shortens the registration time.