Benchmarking reinforcement learning algorithms for autonomous mechanical thrombectomy.

Area of Science:

• Robotics and Artificial Intelligence in Medicine
• Endovascular Interventions
• Neurological Surgery

Background:

• Mechanical thrombectomy (MT) is the standard treatment for acute ischemic stroke.
• Current MT faces challenges including radiation exposure, operator dependency, and limited access.
• Autonomous robotics offer potential solutions to these limitations.

Purpose of the Study:

• To benchmark reinforcement learning (RL) algorithms for autonomous mechanical thrombectomy (MT).
• To evaluate Deep Deterministic Policy Gradient, Twin Delayed Deep Deterministic Policy Gradient, Soft Actor-Critic, and Proximal Policy Optimization.
• To establish a benchmark for autonomous endovascular navigation in stroke treatment.

Main Methods:

• Utilized the stEVE open-source platform for simulated endovascular interventions.
• Trained and evaluated RL algorithms for guidewire navigation in simulated MT procedures.
• Explored the impact of hyperparameter tuning on algorithm performance and assessed them on an MT benchmark.

Main Results:

• Before tuning, Deep Deterministic Policy Gradient achieved 80% success.
• After tuning, Proximal Policy Optimization reached 84% success with reduced procedure time.
• On the MT benchmark, Twin Delayed Deep Deterministic Policy Gradient showed a 68% success rate.

Conclusions:

• This study establishes a benchmark for autonomous endovascular navigation in MT.
• Hyperparameter tuning significantly impacts RL algorithm performance in this context.
• Further research is needed to identify the optimal RL algorithm for autonomous MT.