Abstract
Patient-specific bone models are essential for designing surgical guides and preoperative planning, as they enable the visualization of intricate anatomical structures. However, traditional CT-based approaches for creating bone models are limited to preoperative use due to the low flexibility and high radiation exposure of CT and time-consuming manual delineation. Here, we introduce Semi-Supervised Reconstruction with Knowledge Distillation (SSR-KD), a fast and accurate AI framework to reconstruct high-quality bone models from biplanar X-rays in 30 seconds, with an average error under 1.0 mm, eliminating the dependence on CT and manual work. Additionally, high tibial osteotomy simulation was performed by experts on reconstructed bone models, demonstrating that bone models reconstructed from biplanar X-rays have comparable clinical applicability to those annotated from CT. Overall, our approach accelerates the process, reduces radiation exposure, enables intraoperative guidance, and significantly improves the practicality of bone models, offering transformative applications in orthopedics.
Introduction
In modern orthopedics, patient-specific 3D bone models represent the "gold standard" for preoperative planning and the fabrication of surgical guides. These models allow surgeons to visualize complex anatomical structures and simulate procedures with high precision. Historically, these models have relied exclusively on Computed Tomography (CT) scans. However, CT-based workflows present three significant challenges: high radiation exposure, substantial time requirements for manual segmentation (often taking hours), and a lack of intraoperative flexibility due to the bulk of CT hardware.
To address these limitations, we present SSR-KD, a semi-supervised AI framework designed to reconstruct high-fidelity 3D bone models from just two biplanar X-ray views. By leveraging knowledge distillation from CT data, this system achieves sub-millimeter accuracy in under 30 seconds, potentially replacing high-radiation CT protocols for both surgical planning and real-time intraoperative guidance.
Methodology: The SSR-KD Framework
The core of this research is the Semi-Supervised Reconstruction with Knowledge Distillation (SSR-KD) framework. Unlike traditional methods that rely on fixed 3D grids (voxels) or meshes, SSR-KD utilizes Implicit Occupancy Fields.
Implicit Field Representation
Instead of predicting a rigid grid, the network learns a continuous function that determines whether any specific 3D coordinate $(x, y, z)$ resides inside or outside the bone structure. This approach allows for significantly higher resolution and smoother surfaces without the massive memory overhead typically associated with high-resolution 3D volumes.
Teacher-Student Knowledge Distillation
The framework employs a dual-network strategy to bridge the gap between 2D imaging and 3D reconstruction:
- The Teacher Network: A 3D network trained on a small subset of high-quality, manually labeled CT scans.
- The Student Network: A 2D network that accepts biplanar X-rays as input.
During the training phase, the Teacher provides "pseudo-labels" for a large volume of unlabeled CT data. This guides the Student network to understand complex 3D structures from 2D projections. Furthermore, Cross-Modal Knowledge Distillation aligns 2D pixel-aligned features from X-rays with 3D voxel-aligned features from CTs, ensuring the model learns underlying anatomical principles rather than simple image patterns.
Input Enhancement
To improve accuracy, a pre-processing U-Net segments the bones within the raw X-rays. This creates an "enhanced" input that allows the reconstruction network to focus on the target anatomy while ignoring soft tissue noise and other artifacts.
Experimental Setup and Datasets
The study utilized a comprehensive clinical dataset consisting of 605 one-leg CT scans. To overcome the rarity of paired X-ray and CT data for the same patient, the researchers used Digitally Reconstructed Radiographs (DRRs) to simulate biplanar X-rays from the CT volumes.
The dataset was partitioned as follows:
- Labeled Data: 120 scans manually annotated by three orthopedic experts.
- Unlabeled Data: 485 scans used to fuel the semi-supervised learning process.
- External Validation: 20 cases from an independent institution were used to verify the model's generalizability.
Key Results and Clinical Validation
The SSR-KD framework demonstrated exceptional performance across both quantitative and qualitative metrics:
- High Accuracy: The model achieved an average Symmetric Surface Distance (ASSD) of 0.94 mm. Performance on the femur and tibia was even higher, with errors dropping below 0.8 mm—approaching the 0.68 mm voxel spacing of the original CT scans.
- Rapid Processing: The end-to-end pipeline produces a 3D model in approximately 25.2 seconds. This is a transformative improvement over the ~4 hours required for manual expert annotation.
- Surgical Applicability: In a blinded clinical study, surgeons performed High Tibial Osteotomy (HTO) simulations. The results showed that surgical guides designed from AI-reconstructed models were indistinguishable in fit, stability, and accuracy from those designed using traditional CT-based models.
Discussion and Future Work
While SSR-KD marks a significant advancement, certain challenges remain. The current model struggles with large metal implants, which create significant artifacts and shadows in X-ray views. Additionally, while DRRs are effective for training, real-world clinical X-rays contain scatter and beam hardening artifacts that require further modeling.
Future research will focus on fine-tuning the framework using large-scale datasets of real clinical X-rays and extending the architecture to handle more complex, overlapping anatomical regions such as the spine or the hands. Despite these limitations, SSR-KD offers a path toward safer, faster, and more flexible orthopedic care by bringing CT-quality 3D visualization into the operating room via low-dose X-ray protocols.
Summary updated