Research On Anterior Spinal Surgery Planning Based On Medical Imaging And Machine Learning

Spinal decompression surgery is one of the most dangerous operations in surgery.The surgical approach planning is the core of preoperative preparation,and the results of the planning will have a direct impact on the surgical outcome.Compared with posterior approach,the anterior surgery has a more complete decompression effect on the spine and has less impact on the aesthetics of the lateral approach,making it an important therapeutic procedure for many specific spinal diseases.It is significant to perform preoperative path planning for anterior spinal surgery based on surgical navigation technology,because the current orthopedic surgery is based on the surgeon’s rich operational experience and skilled operation skills to determine the intraoperative path in general.During the operation of the operation,the patient is required to continuously be scanned by means of the instrument.And then surgeon continuously adjusts the posture of the surgical instrument by observing the intraoperative image data and artificially plans the surgical path,and finally performs the operation on the patient’s lesion location.Due to the accumulation of radiation and the long stay time of the surgical instrument in the patient’s body,such surgical method has a greater damage to the human body.Therefore,researching new ways to plan an optimal and safe surgical path to reduce the degree of human injury and improve the accuracy of surgery is of great significance for the development of spinal surgery.Based on the reinforcement learning technology,this paper proposes an anterior surgical path planning system for spinal surgery robots.Since the current orthopedic surgery is on the basis of the surgical navigation system,the CT images are first processed.In this paper,the multi-task segmentation of thoracic enhanced CT images is performed by region growing algorithm and level set method.An improved adaptive level set method is proposed to realize fast and accurate batch segmentation of CT images.The obtained true value map is constructed dataset’s label.In order to realize the automatic segmentation of medical images,this paper uses the optimized U-Net network to train the CT images of the chest and the corresponding label map.For the disadvantage of fewer data sets in medical images,the data sets are expanded to improve the performance of the network model.Also,the verification of verification sets and the prediction of test samples are carried out.The finally calculated value of IOU(Intersection over Union)reached about 0.924,which proves that the optimized U-Net network can achieve rapid and accurate automatic segmentation of CT images of the chest.According to the segmentation results,three-dimensional reconstruction of important organs such as bones and blood vessels is carried out.The preoperative three-dimensional path planning is implemented based on the Q-learning algorithm that is extended to 3D space on the environmental model,and its convergence is guaranteed by the ε-greedy strategy.In order to solve the problem that the planning process is easy to fall into the curse of dimensionality,this paper proposes two kinds of surface constrained optimization schemes-conical optimization and cylindrical optimization,which limit the search scope of the Agent and improve the search efficiency.Through simulation experiments,the system can plan an optimal surgical path from the surgical entrance to the lesion position and avoiding important organs of the human body for the end effector of the spinal surgery robot without prior knowledge of the environment.