Imitation Learning for Path Planning in Cardiac Percutaneous Interventions

Area of Science:

• Cardiovascular Surgery
• Medical Robotics
• Artificial Intelligence

Background:

• Minimally invasive transcatheter procedures for mitral regurgitation present a steep learning curve due to complex hand-eye coordination requirements.
• Automation of these procedures is crucial for enhancing accessibility and standardizing outcomes in cardiac interventions.

Purpose of the Study:

• To develop and evaluate a learning-based path planning framework for robotic catheter navigation in cardiac percutaneous interventions.
• To adapt automated navigation strategies to the dynamic and safety-critical environment of mitral valve repair.

Main Methods:

• Comparison of generative adversarial imitation learning and behavioral cloning against traditional path planning algorithms (e.g., rapidly-exploring random trees).
• Creation of patient-specific digital twins with dynamic cardiac motion simulation for realistic procedural replication.

Main Results:

• Learning-based approaches significantly reduced target position errors compared to traditional methods.
• Improved path smoothness and increased clearance from anatomical obstacles were observed with learning methodologies.
• Consistent and repeatable navigation routes were achieved in both static and dynamic cardiac scenarios.

Conclusions:

• Learning methodologies offer a robust solution for automated catheter navigation in complex cardiac anatomy.
• Task demonstration embedding in learning processes can optimize and standardize minimally invasive cardiac procedures, enhancing patient safety and procedural efficiency.