AI navigation in stroke treatment hindered by complex vascular geometry, study finds

Mechanical thrombectomy (MT) is a time-critical intervention for acute ischemic stroke, but access is limited by a shortage of specialists. Researchers from King's College London (Wu et al.) propose using reinforcement learning (RL) to automate endovascular navigation. Their study, published on arXiv (July 2026), introduces an automated pipeline to quantify vascular geometry features that influence navigation difficulty. Using CT angiograms from 61 patients, they extracted metrics such as aortic arch type, bovine arch, vessel length, tortuosity, take-off angle, and number of reverse curves. A Soft Actor-Critic RL algorithm was then tested for 120-second autonomous navigation episodes to correlate these features with performance.

The results show statistically significant impacts: on the left side, a bovine arch increased navigation time by 30.19 seconds, and a type II/III aortic arch added 37.92 seconds. Greater tortuosity (beta coefficient 118.20) prolonged the procedure and reduced success probability. On the right side, type II/III arches added 45.94 seconds, and each additional reverse curve added 3.96 seconds while lowering success odds. This demonstrates for the first time that vascular geometry strongly affects RL-based navigation difficulty. The proposed pipeline provides objective, quantitative characterization, laying the groundwork for standardized complexity grading and more robust RL model evaluation—though the study does not claim clinically generalizable autonomous navigation yet.