Toward AI Autonomous Navigation for Mechanical Thrombectomy using Hierarchical Modular Multi-agent Reinforcement Learning (HM-MARL)

A research team from King's College London and University College London has published a breakthrough in IEEE Robotics and Automation Letters demonstrating the first successful in vitro autonomous navigation for mechanical thrombectomy (MT), a life-saving stroke procedure. Their system, called Hierarchical Modular Multi-Agent Reinforcement Learning (HM-MARL), uses multiple AI agents working together to navigate surgical tools through complex vasculature.

The technical approach decomposes the challenging navigation task into specialized subtasks, each trained using Soft Actor-Critic reinforcement learning. In simulation tests, the single-vasculature model achieved remarkable 92-100% success rates on individual anatomies, while the more challenging multi-vasculature model achieved 56-80% success across multiple patient anatomies. In physical lab tests using realistic vasculature models, HM-MARL successfully navigated 100% of trials from the femoral artery to the right common carotid artery and 80% to the right internal carotid artery.

The research addresses a critical healthcare access problem: mechanical thrombectomy is the optimal treatment for acute ischemic stroke but remains limited by geographic and logistical barriers requiring specialized neurointerventionalists. This AI-assisted navigation system could eventually enable remote procedures or assist less experienced operators. While the system failed on a particularly challenging left-side anatomy due to catheter limitations, the study represents significant progress toward clinical translation of autonomous surgical navigation, with future work focusing on refining RL strategies using world models and validation on unseen data.