Real-Time Reconstruction of 3D Bone Models via Very-Low-Dose Protocols

TL;DR

SSR-KD is introduced, 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 and significantly improves the practicality of bone models, offering transformative applications in orthopedics.

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.

Figures (8)

• Figure 1: Comparison of conventional reconstruction and our proposed solution.a: Conventional CT-based reconstruction. Traditionally, the CT scan of a patient should be collected first, and then the expert will operate commercial software like Mimics to delineate 3D bone models from the CT scan. b: Our proposed automatic reconstruction from biplanar X-rays. The approach requires only two X-rays of the patient; then, the deep reconstruction network SSR-KD will automatically estimate the occupancy field; finally, Marching Cubes lorensen1987marching is applied to extract underlying iso-surfaces and reconstruct 3D bone models. c: Comprehensive comparison of different reconstruction pipelines. Our SSR-KD uses low-cost, low-radiation biplanar X-rays to automatically create 3D bone models. The reconstruction is fast enough to be used during surgery and produces results comparable in quality to CT scans without needing expert assistance.
• Figure 2: Qualitative comparison of two reconstruction methods. CT-based annotated bone models are obtained manually from the corresponding CT, while 2-view reconstructed models are automatically generated by our proposed SSR-KD from biplanar X-rays. We also visualize the surface error (mm) between two reconstructed models. The position, pose, and contour shape of 2-view reconstructed models are exactly consistent with CT-based annotated models. The error is slightly higher in patella and some corner regions because these parts do not appear clearly in the input views due to occlusion.
• Figure 3: User study to compare the clinical difference.a: For each case, given 2-view X-ray images as the reference, the user was asked to evaluate bone models (CT-based annotated or 2-view reconstructed) in terms of shape, details, and clinical significance. b: Evaluation results (mean and standard deviation) were collected from 10 experts.
• Figure 4: HTO simulation to validate the clinical practicality.a: To simulate a HTO, patient-specific bone models were 3D-printed from CT-based annotations to replicate the patient's anatomy (1). Two different versions of surgical guides were then fabricated: one designed from CT-based annotated bone model (2) and another based on 2-view reconstructed bone model (3). During the simulation, surgical guides were used to direct the procedure. A surgeon would place the guide on the tibial region using a Vise Grip 360° Swivel Head, insert Kirschner-wires through the holes in the surgical guide and into the bone model for fixation purposes, and perform the osteotomy cut using a saw blade. b: Three scores (fixing, stability, and accuracy) were used to evaluate the surgical guides, and the operation time (minutes) was also recorded during the operation, demonstrating that 2-view reconstructed bone models have comparable practicality than CT-based annotated models.
• Figure 5: Experimental analysis.a: Sensitivity analysis of model performance with respect to the number of labeled and unlabeled training pairs. b: Experiments with different numbers of input views, showing that a 2-view setup provides the optimal trade-off for accurate knee bone reconstruction. c: Evaluation of model robustness under reduced angular separations, simulating clinical scenarios where gantry rotation is restricted to less than 90°. d: Assessment of generalization on an external dataset (20 cases), where low ASSD values (mm) and qualitative visualizations confirm the strong performance of SSR-KD on unseen data.