Investigating robot control policy learning for autonomous x-ray-guided spine procedures

Area of Science:

• Robotics and Artificial Intelligence
• Medical Imaging and Interventional Procedures
• Surgical Navigation Systems

Background:

• Imitation learning (IL) is gaining traction in robotics, particularly for video-based control.
• Its application to sparse-input scenarios like X-ray-guided procedures, such as spine instrumentation, is not well-established.
• This study investigates IL for bi-plane X-ray-guided cannula insertion.

Purpose of the Study:

• To assess the feasibility, opportunities, and challenges of imitation policy learning for X-ray-guided spine procedures.
• To evaluate the effectiveness of IL in a simulated environment for cannula insertion.
• To explore the potential for IL in developing robotic assistance for spinal navigation.

Main Methods:

• Developed a realistic, in silico simulation sandbox for X-ray-guided spine procedures.
• Curated a dataset of correct trajectories and corresponding bi-planar X-ray sequences.
• Trained IL policies for planning and open-loop control of cannula alignment using only visual information.

Main Results:

• The IL policy achieved 68.5% success on the first attempt in simulated cannula insertion.
• Safe intra-pedicular trajectories were maintained across diverse vertebral levels and complex anatomies (e.g., fractures).
• Partial sim-to-real transfer was observed, with plausible trajectories on real X-ray data.

Conclusions:

• Imitation learning shows promise for X-ray-guided robotic cannula insertion, achieving high success rates in simulation.
• Limitations exist, particularly in entry-point precision, highlighting areas for future improvement.
• Further development with robust priors and domain knowledge could enable CT-free robotic spinal navigation.