Benchmarking reinforcement learning algorithms for autonomous mechanical thrombectomy

Mechanical thrombectomy (MT) is the gold standard for treating acute ischemic stroke. However, challenges such as operator radiation exposure, reliance on operator experience, and limited treatment access remain. Although autonomous robotics could mitigate some of these limitations, current research lacks benchmarking of reinforcement learning (RL) algorithms for MT. This study aims to evaluate the performance of Deep Deterministic Policy Gradient, Twin Delayed Deep Deterministic Policy Gradient, Soft Actor-Critic, and Proximal Policy Optimization for MT.

Simulated endovascular interventions based on the open-source stEVE platform were employed to train and evaluate RL algorithms. We simulated navigation of a guidewire from the descending aorta to the supra-aortic arteries, a key phase in MT. The impact of tuning hyperparameters, such as learning rate and network size, was explored. Optimized hyperparameters were used for assessment on an MT benchmark.

Before tuning, Deep Deterministic Policy Gradient had the highest success rate at 80% with a procedure time of 6.87 s when navigating to the supra-aortic arteries. After tuning, Proximal Policy Optimization achieved the highest success rate at 84% with a procedure time of 5.08 s. On the MT benchmark, Twin Delayed Deep Deterministic Policy Gradient recorded the highest success rate at 68% with a procedure time of 214.05 s.

This work advances autonomous endovascular navigation by establishing a benchmark for MT. The results emphasize the importance of hyperparameter tuning on the performance of RL algorithms. Future research should extend this benchmark to identify the most effective RL algorithm.