Patient-specific 3D bone models are used in preoperative planning and the design of surgical guides, especially when anatomy is complex. In current practice, these models are commonly derived from CT scans and then manually delineated by experts, a process that can take around four hours for a single leg. That workflow also carries a higher radiation burden and is difficult to apply flexibly during surgery. A new deep-learning framework, Semi-Supervised Reconstruction with Knowledge Distillation, or SSR-KD, reconstructs four knee bones from biplanar X-ray images instead of CT. Using anterior-posterior and right-left views, the approach generates femur, tibia, fibula and patella models in about 25.2 seconds. Across the four bones, it achieved a Dice similarity coefficient of 90.9%, a Hausdorff distance of 2.76 mm and an average symmetric surface distance of 0.94 mm. The framework also underwent qualitative, external and clinical evaluation, including simulated high tibial osteotomy, where reconstructed models performed comparably to CT-based annotated models.
An Occupancy-Field Approach Replaces CT-Based Modelling
SSR-KD reconstructs bone anatomy by estimating a 3D occupancy field rather than predicting meshes or volumes directly. Each point in 3D space is assigned a probability of belonging to patella, femur, fibula, tibia or background. Bone surfaces are then extracted using Marching Cubes. The framework takes two orthogonal X-ray views as input and uses deep neural networks to map spatial coordinates to occupancy values.
The reconstruction network first extracts semantic features from each X-ray image. For a given 3D point, pixel-aligned features are queried from the two feature maps using projected coordinates and bilinear interpolation. Depth values from both views are added before the final occupancy prediction. The resulting representation supports reconstruction at a resolution of at least 256³.
Must Read: AI Support Lifts Resident Performance in Knee MRI
The total reconstruction time was about 25.2 seconds, including 20.8 seconds for network inference, 1.5 seconds for Marching Cubes and 2.9 seconds for post-processing. That contrasts with the hours normally required for manual CT-based delineation. The framework also used an enhanced X-ray projection input that highlighted bone regions while preserving raw image information, reducing reconstruction error by 0.12 mm. Quantitative testing showed the strongest performance in femur and tibia reconstruction, with average symmetric surface distances of 0.80 mm and 0.77 mm respectively, while patella and fibula remained more difficult at 1.12 mm and 1.05 mm.
Semi-Supervised Learning Reduces Annotation Burden
The training strategy addresses the burden of manual annotation. The dataset comprised 605 one-leg CT scans covering the knee joint across varied ages, genders and scanning ranges. Among them, 120 CT scans were manually annotated by three orthopaedic experts, while 485 were treated as unlabelled data. For development, 70 labelled cases and 485 unlabelled cases were used for training, 10 labelled cases for validation and 40 labelled cases for testing.
SSR-KD uses a two-stage process. A CT-based reconstruction network is first trained on labelled CT data. That network then supports training of the X-ray-based model through pseudo labels and cross-modal knowledge distillation. As a result, the X-ray model learns from both labelled examples and paired unlabelled CT and X-ray data. This allows the framework to capture broader anatomical variability without relying on large numbers of manually annotated models.
Sensitivity analysis showed that the framework remained effective with as few as 28 labelled cases when combined with 485 unlabelled cases. Performance dropped markedly when labelled data were reduced to 14 cases or fewer, or when too few unlabelled data were available. The number of projection views was also tested. A single view did not capture sufficient geometry because of occlusion. Increasing the number of views beyond two produced only limited improvement compared with the gain from one view to two. The two-view configuration therefore offered the best balance between reconstruction accuracy and efficiency.
Clinical Assessment Supports Planning Use
The framework was assessed beyond numerical metrics through a blinded user study involving 10 experts, including orthopaedists, medical school professors and medical engineers. Forty test cases were reviewed. Each case included biplanar X-rays and two model types: CT-based annotated models and two-view reconstructed models. Experts scored shape, detail and clinical significance using four categories ranging from poor to perfect.
Across all three measures, reconstructed models achieved scores comparable to CT-based annotated models. Differences between the two model types were difficult to identify when judged against the biplanar X-rays. Mean clinical significance scores for the reconstructed models were above 3, indicating agreement that the models were useful for planning procedures such as high tibial osteotomy.
A further evaluation examined patient-specific surgical guide design in simulated high tibial osteotomy. For each patient, two guides were created: one based on CT-annotated models and one based on reconstructed models. In five patient cases assessed in a blinded setting, surgeons rated the guides for fitting, stability and accuracy. Procedure time from guide placement to completion of the cut was also recorded. Guides based on reconstructed models received comparable ratings to CT-based guides and were associated with shorter operating time. Additional testing under reduced angular separation showed that even at 30°, femur and tibia reconstruction remained below 1.0 mm average surface error, while patella and fibula were more affected by occlusion and information loss. External validation on 20 cases from another institution also produced femur and tibia reconstruction below 1.0 mm average surface error.
SSR-KD replaces CT-dependent bone reconstruction with a two-view X-ray workflow that is faster and less dependent on manual annotation. The framework reconstructs four knee bones in about half a minute and achieves submillimetre average surface error overall. Clinical review and simulated high tibial osteotomy both support the use of reconstructed models in planning and guide design. Femur and tibia showed the strongest and most consistent results, while patella and fibula remained more challenging because of occlusion. Testing with metal implants also showed that smaller implants causing limited structural disruption could still permit successful reconstruction. For hospitals with limited medical resources, the approach offers a lower-cost, lower-radiation route to patient-specific 3D bone modelling.
Source: npj Digital Medicine
Image Credit: iStock
References:
Lin Y, Sun H, Li Y et al. (2026) Real-time reconstruction of 3D bone models via very-low-dose protocols. npj Digit Med: In Press.
