A zero-shot reinforcement learning strategy for autonomous guidewire navigation.

Area of Science:

• Medical Robotics
• Artificial Intelligence in Medicine
• Cardiovascular Interventions

Background:

• Cardiovascular disease treatment involves complex guidewire and catheter navigation.
• Current methods lead to prolonged procedures and patient/clinician X-ray radiation exposure.
• Deep reinforcement learning (DRL) shows potential for automating catheter navigation but struggles with generalization.

Purpose of the Study:

• To develop a zero-shot learning strategy for autonomous endovascular navigation.
• To enable DRL algorithms to generalize to unseen vascular anatomies without retraining.
• To improve the efficiency and safety of robotic interventions for cardiovascular diseases.

Main Methods:

• Proposed a zero-shot learning strategy for 3D autonomous endovascular navigation.
• Utilized a small training set of branching patterns for reinforcement learning.
• Developed a control algorithm applicable to novel vascular geometries.

Main Results:

• Demonstrated the method on 4 diverse vascular systems with a 95% success rate in reaching random targets.
• Achieved computationally efficient training, completing in just 2 hours.
• Validated the algorithm's ability to navigate unseen geometries.

Conclusions:

• The proposed training method enables navigation of unseen vascular geometries.
• The strategy leverages a nearly shape-invariant observation space for generalization.
• This approach offers a promising solution for automated catheter navigation in complex interventions.