Real-time reconstruction of 3D bone models via very-low-dose protocols

Area of Science:

• Orthopedic surgery
• Medical imaging
• Artificial intelligence in medicine

Background:

• Patient-specific bone models are crucial for surgical planning and guide design.
• Current CT-based methods are time-consuming, involve high radiation, and are limited to preoperative use.
• Manual delineation of bone structures is labor-intensive and prone to error.

Purpose of the Study:

• To introduce a novel AI framework, Semi-Supervised Reconstruction with Knowledge Distillation (SSR-KD).
• To enable rapid and accurate reconstruction of high-quality bone models from biplanar X-rays.
• To reduce reliance on CT scans and manual segmentation for orthopedic applications.

Main Methods:

• Developed SSR-KD, an AI framework utilizing knowledge distillation for semi-supervised learning.
• Trained the model on biplanar X-ray data to reconstruct 3D bone models.
• Evaluated model accuracy with an average error under 1.0 mm.
• Conducted high tibial osteotomy simulations using reconstructed models.

Main Results:

• SSR-KD reconstructs bone models from biplanar X-rays in approximately 30 seconds.
• Achieved an average reconstruction error of less than 1.0 mm.
• Expert simulations showed comparable clinical applicability to CT-derived bone models.
• Demonstrated significant acceleration and reduced radiation exposure compared to traditional methods.

Conclusions:

• SSR-KD provides a fast, accurate, and low-radiation alternative for creating patient-specific bone models.
• The AI framework enhances the practicality of bone models for both preoperative planning and potential intraoperative guidance.
• This technology has transformative potential for orthopedic surgery and patient care.