A Generalized Full-to-Partial Registration Framework of 3D Point Sets for Computer-Aided Orthopedic Surgery

Area of Science:

• Orthopedic Surgery
• Computer-Aided Surgery
• Medical Imaging
• Machine Learning

Background:

• Precise 3D point set registration is vital for computer-aided orthopedic surgery.
• Challenges include partial overlap, noise, and poor initialization of 3D data.
• Existing registration techniques struggle with these inherent complexities.

Purpose of the Study:

• To develop a generalized and robust full-to-partial registration framework for orthopedic surgery.
• To overcome limitations of current methods in handling noisy and partially overlapping 3D point sets.
• To improve the accuracy and reliability of bone alignment in surgical planning and execution.

Main Methods:

• A novel registration framework utilizing reinforcement learning (RL) is proposed.
• The RL framework is designed to be robust against noise and poor initialization.
• The method handles partial overlap between full and partial 3D point sets effectively.

Main Results:

• The proposed method achieved a superior C.D. error of 8.211 e-05 on bone datasets.
• Demonstrated exceptional generalization across various bone types (pelvis, femur, tibia).
• Consistently outperformed state-of-the-art registration techniques in experiments.

Conclusions:

• The developed reinforcement learning framework enables precise bone alignments for computer-aided orthopedic surgery.
• The method offers a robust solution for challenging registration scenarios.
• This advancement has significant implications for improving surgical accuracy and patient outcomes.