World Model for AI Autonomous Navigation in Mechanical Thrombectomy

TL;DR

This work addresses the challenge of generalizing autonomous navigation for mechanical thrombectomy (MT) across diverse patient vasculatures by deploying a world-model reinforcement learning framework. The authors implement TD-MPC2, a model-based RL algorithm, and compare it to a state-of-the-art SAC baseline across ten real patient vasculatures in a multi-task setting using an in silico stEVE-SOFA simulation with realistic devices. Results show that TD-MPC2 significantly improves mean success rate to $65\%$ versus $37\%$ for SAC ($p<0.001$), with notable gains in several navigation tasks, though at the cost of longer procedure times and lower exploration demands. These findings support the potential of world-models for generalizable AI-driven autonomous endovascular interventions and set a foundation for future work integrating differentiable physics and richer input modalities to move toward clinical deployment.

Abstract

Autonomous navigation for mechanical thrombectomy (MT) remains a critical challenge due to the complexity of vascular anatomy and the need for precise, real-time decision-making. Reinforcement learning (RL)-based approaches have demonstrated potential in automating endovascular navigation, but current methods often struggle with generalization across multiple patient vasculatures and long-horizon tasks. We propose a world model for autonomous endovascular navigation using TD-MPC2, a model-based RL algorithm. We trained a single RL agent across multiple endovascular navigation tasks in ten real patient vasculatures, comparing performance against the state-of-the-art Soft Actor-Critic (SAC) method. Results indicate that TD-MPC2 significantly outperforms SAC in multi-task learning, achieving a 65% mean success rate compared to SAC's 37%, with notable improvements in path ratio. TD-MPC2 exhibited increased procedure times, suggesting a trade-off between success rate and execution speed. These findings highlight the potential of world models for improving autonomous endovascular navigation and lay the foundation for future research in generalizable AI-driven robotic interventions.

Figures (1)

• Figure 1: Full MT vasculature with each navigation task labeled. $A_1$: Common iliac artery to top of descending aorta, $A_{2L}$: top of descending aorta to left common carotid artery (CCA), $A_{2R}$: top of descending aorta to right CCA, $A_{3L}$: left CCA to left ICA, $A_{3R}$: right CCA to right ICA.