Study on the Autonomous Guidewire Navigation Method Utilizing the Generalization of Reinforcement Learning

In vascular interventional surgery, robots are playing an increasingly important role in order to protect doctors and improve surgical efficiency. The guidewire is a crucial instrument in vascular interventional surgery. However, the nonlinear structure of the guidewire makes it difficult to perform dynamic modeling and control. Currently, guidewire navigation methods based on reinforcement learning are on the rise. But the existing methods at this stage all obtain agents in a single training scenario, and it is difficult for them to adapt to the ever-changing practical needs. In this paper, we propose an autonomous guidewire navigation method based on the generalization of reinforcement learning. By constructing an environmental vector space that includes the physical properties of the guidewire, hydrodynamic parameters, and imaging features, we design an environment encoder to extract its implicit representations. We also fuse visual and tactile information and combine it with an adaptive module to achieve policy adjustment. Experiments have shown that after changing the phantom or the guidewire, the success rate of the proposed method for the task of the left subclavian artery (LSA) remains above 60%.