ALICE - Artificial Intelligence Catheter. Towards Autonomous Closed Loop Control of Passive Endovascular Catheters Based on Deep Reinforcement Learning

Endovascular catheters are used for state of the art therapies of many widespread diseases. Navigating them can be very laborious and so far no robotic assistance exists for passive catheters. Steer-able catheters exist, but due to their large diameter they are not suitable for many interventions. We propose a closed loop control system where a deep reinforcement learning based control algorithm steers the catheter. The algorithm is provided with live data by a tracking system. Prior to the intervention the control algorithm is trained on the simulation model and by expert demonstration. Here we present the results of our experiments, where a control algorithm learns to steer a guidewire through a simplified vascular tree. Learning is performed in the simulation model and the result transferred to the test bench. Our results show that the algorithm is able to learn catheter steering. However the simulation results cannot be transferred to the test bench directly without facing a reduced accuracy due to the test bench not having perfect states like the simulation.