LSTM-Based Modeling and Reinforcement Learning Control of a Magnetically Actuated Catheter

TL;DR

This paper introduces an LSTM-based model for magnetically actuated catheters and demonstrates reinforcement learning controllers, showing improved accuracy and smoothness in path following tasks for minimally invasive procedures.

Contribution

It presents a novel LSTM model for nonlinear hysteretic catheter dynamics and compares RL controllers, highlighting the superior performance of actor-critic methods for navigation tasks.

Findings

LSTM model achieves RMSE of 0.42 mm and 99.8% coverage within 3 mm.

Actor-critic RL outperforms DQN in accuracy and smoothness.

RL controllers enable precise and smooth catheter navigation.

Abstract

Autonomous magnetic catheter systems are emerging as a promising approach for the future of minimally invasive interventions. This study presents a novel approach that begins by modeling the nonlinear and hysteretic dynamics of a magnetically actuated catheter system, consists of a magnetic catheter manipulated by servo-controlled magnetic fields generated by two external permanent magnets, and its complex behavior is captured using a Long Short-Term Memory (LSTM) neural network. This model validated against experimental setup's data with a root mean square error (RMSE) of 0.42 mm and 99.8% coverage within 3 mm, establishing it as a reliable surrogate model. This LSTM enables the training of Reinforcement Learning (RL) agents for controlling the system and avoiding damage to the real setup, with the potential for subsequent fine-tuning on the physical system. We implemented Deep Q-Network (DQN) and actor-critic RL controllers, comparing these two agents first for regulation and subsequently for path following along linear and half-sinusoidal paths for the catheter tip. The actor-critic outperforms DQN, offering greater accuracy and faster performance with less error, along with smoother trajectories at a 10 Hz sampling rate, in both regulation and path following compared to the DQN controller. This performance, due to the continuous action space, suits dynamic navigation tasks like navigating curved vascular structures for practical applications.