Study Of Automatic Guidewire Policy For Vascular Interventional Robot

Cardiovascular interventional surgery is a medical treatment for cardiovascular disease,which has the advantages of less trauma,fewer postoperative complications and better targeting.In recent years vascular interventional robots can assist doctors to realize telemedicine,but the intelligence of the robots is insufficient.In this paper,we investigate automatic guidewire navigation method by combining artificial intelligence algorithms.We propose the algorithmic framework based on reinforcement learning and imitation learning.First,the framework of the guidewire tip tracking algorithm was built in this paper.The images of the guidewire tip were collected on the simulated vascular model to construct the dataset.The SiamRPN algorithm was used to track the guidewire tip.The results show that the tracking success rate reaches 97.14%when the IOU threshold is set to 0.5.Secondly,this paper implements guidewire navigation based on reinforcement learning algorithm.Markov model is established for vascular interventional operations.Sample splitting method is proposed to improve the Soft Actor-Critic algorithm and expand the experience pool by splitting a single action into multiple actions.This method increases explored state space of robot and reduces training time.Then,this paper combines expert demonstration data to build a generative adversarial imitation learning framework.On this basis,selfimitation learning is added to solve the problem of limited demonstration data.We propose to implement pre-training in an offline environment based on the demo data,and loading the pre-trained model during online training can further reduce the training time.Finally,this paper builds a simulation experiment platform and validates guidewire navigation algorithm.The reinforcement learning algorithm is validated on the platform,and result show that improved reinforcement learning algorithm by sample splitting can successfully train robot to complete task.Training time is effectively reduced compared to unimproved algorithm.The result combined with expert demonstration data show that imitation learning has shorter training time.The task can be completed with small sample of experts by combining self-imitation learning.Loading pre-trained model can further reduce training time.The automatic guidewire navigation policy proposed in this paper can assist physicians in interventional procedures.It has the advantages of reducing physician workload,improving average treatment outcomes,and integrating contextual information faster.It can also be used in physician training to help rookies learn excellent guidewire navigation policies more quickly.