Warm-started reinforcement learning for iterative 3D/2D liver registration

Area of Science:

• Medical Imaging
• Computer-Assisted Surgery
• Machine Learning

Background:

• Accurate registration between preoperative CT scans and intraoperative laparoscopic video is essential for augmented reality (AR) guided minimally invasive surgery.
• Learning-based methods offer faster inference than traditional optimization-based approaches for surgical registration.
• However, supervised methods often require computationally expensive optimization-based refinement for precise alignment.

Purpose of the Study:

• To develop a reinforcement learning (RL) framework for automated and efficient CT-to-video registration.
• To improve the speed and accuracy of registration for augmented reality (AR) surgical guidance.
• To eliminate the need for manual tuning of parameters in the registration process.

Main Methods:

• A discrete-action reinforcement learning (RL) framework was designed, treating CT-to-video registration as a sequential decision-making problem.
• A shared feature encoder, initialized with a supervised pose estimation network, processed CT and laparoscopic video data.
• The RL policy head determined rigid transformations and the optimal stopping point for the registration iteration.

Main Results:

• The RL-based method achieved an average target registration error (TRE) of 15.70 ± 8.18 mm on a public laparoscopic dataset.
• The performance was comparable to supervised methods that included optimization-based refinement.
• The proposed framework demonstrated faster convergence compared to existing approaches.

Conclusions:

• The developed RL formulation enables automated and efficient iterative registration for surgical AR applications.
• This discrete-action framework removes the need for manual step size and stopping criteria adjustments.
• It lays the groundwork for future advancements in continuous-action and deformable registration models for surgical AR.