RL-MG: Reinforcement Learning Framework for Magnetic Guidewire Path Following

摘要

Magnetic guidewires (MG), featuring active steering capabilities, show great potential for enhancing interventional surgery due to their flexibility and controllability. However, achieving real-time autonomous navigation of a MG within complex vascular networks remains a major challenge. This study presents a reinforcement learning (RL)-based control framework for MG path following in vascular navigation tasks. The tip of the MG is equipped with an embedded small permanent magnet, enabling magnetic actuation via a movable external magnet (EM). The proposed RL framework is trained in a Simulation Open Framework Architecture (SOFA)-based simulation, where it learns to control the motion of the EM and the translation of the MG to accurately follow a predefined reference path. To validate the effectiveness of the trained RL policy, an experimental setup is established, comprising an MG prototype, an EM mounted on a robotic arm, a vascular phantom model, and a depth camera for visual feedback. Experimental results demonstrate that the RL policy achieves reliable path following performance in physical environments, with an overall mean tracking error of 3.39 ± 1.87 mm (mean ± SD) across all four tested paths.