TL;DR
This paper presents a novel deep-learning pipeline to reconstruct 3D models of the proximal tibia from low-dose, standing biplanar X-rays (EOS system). This approach aims to eliminate the need for CT scans in preoperative planning for high tibial osteotomies, thereby reducing radiation exposure and costs, while enabling weight-bearing assessments.
What Problem Does This Paper Address?
Preoperative planning for orthopedic surgeries, particularly for procedures like high tibial osteotomies (HTOs), traditionally relies on 3D models reconstructed from CT scans. While CT provides detailed 3D information, it comes with significant drawbacks: high radiation doses, increased costs, and the fact that scans are typically performed in a non-weight-bearing (supine) position. This non-weight-bearing aspect is a critical limitation, as many orthopedic conditions, such as osteoarthritis, are exacerbated by weight-bearing. Surgeons often prefer standing radiographs for their ability to capture the lower limb in a weight-bearing state, but these 2D images are less informative for detailed 3D planning. The challenge, therefore, is to leverage the benefits of standing radiographs (low dose, weight-bearing) to create accurate 3D models suitable for clinical-grade preoperative planning, thus avoiding the need for CT.
Methods
The authors developed a deep-learning pipeline consisting of three main stages:
- Leg Separation Network: A key innovation is a Convolutional Neural Network (CNN) designed to separate the left and right legs in sagittal biplanar X-rays. This is crucial because standard standing leg X-rays often show superimposed legs, making individual leg reconstruction difficult. The network takes a biplanar EOS scan as input and outputs two separate sagittal images, one for each leg. This separation is achieved by first applying a contrast enhancement to the input image to better distinguish bone from soft tissue.
- Landmark Localization Network: Another CNN is employed to identify key joint center landmarks (hip, knee, and ankle) in both the frontal and the separated sagittal X-ray images. These landmarks are essential for subsequent orthopedic measurements and planning.
- 3D Reconstruction Network: A U-Net architecture is used to reconstruct the 3D surface model of the proximal tibia. This network takes the frontal X-ray and the separated sagittal X-ray of the target leg as input. The ground truth for training comes from segmented CT scans. The network outputs a 3D model, which is then refined using the Marching Cubes algorithm.
These reconstructed 3D models and localized 3D landmarks are then fed into a previously validated preoperative planning framework that uses a genetic algorithm for multi-objective optimization. This framework aims to determine optimal osteotomy axis, angle, and fixation plate placement.
Datasets
The study utilized a dataset of 175 HTO patients who underwent CT scans of both legs. Digitally Reconstructed Radiographs (DRRs) were generated from these CT scans to create training data that mimics the geometry of the EOS imaging system. For training the separation network, the authors generated DRRs of both legs and then created separate sagittal images for each leg as targets. For landmark localization, 2D landmarks from ground truth CT data were projected onto the DRRs.
For evaluation, a separate set of 52 patients was used, which was the same set as in a previous publication by the authors.
Key Insights and Ideas
- Addressing Sagittal Superimposition: The leg separation network is a significant contribution, directly tackling the challenge of overlapping anatomy in standard standing leg X-rays, which is a common hurdle in 2D-3D reconstruction from biplanar views.
- Weight-Bearing 3D Models: By using standing EOS radiographs, the pipeline generates 3D models that represent the anatomy in a weight-bearing state, which is more clinically relevant for assessing deformities and planning surgeries like HTOs.
- CT Replacement Potential: The study demonstrates that these 3D models derived from low-dose X-rays can be used for clinical-grade preoperative planning, suggesting a potential to replace CT scans in specific cases, thereby reducing patient radiation exposure and healthcare costs.
- Automated Planning Integration: The seamless integration of the 3D reconstruction pipeline with an existing automated preoperative planning framework highlights the practical clinical utility of the developed method.
Key Results
- 3D Reconstruction Accuracy: The reconstructed proximal tibiae achieved high Dice coefficients: 0.92 ± 0.02 for the right tibia and 0.89 ± 0.06 for the left tibia. Mean Euclidean distances between predicted and ground truth meshes were 1.21 ± 0.38 mm (right) and 1.63 ± 0.74 mm (left).
- Landmark Localization Accuracy: Mean localization errors for joint centers were within acceptable ranges: frontal plane (1.64-1.78 mm) and sagittal plane (1.64-3.72 mm). The mean 3D Euclidean distance to ground truth landmarks was 2.87 ± 1.37 mm.
- Deformity Measurement Accuracy: The mean differences in measured anatomical parameters were: Mechanical Axis (MA) of 0.52° ± 0.47° and Tibial Slope (TS) of 4.33° ± 3.92°.
- Preoperative Planning Feasibility: The pipeline successfully generated clinically feasible surgical plans for all 52 patients. The planning solutions derived from the reconstructed models showed only slight deviations compared to ground truth solutions based on CT scans. For instance, the mean absolute difference for the target MA was 0.06° ± 0.11° (ground truth) vs. 0.11° ± 0.33° (reconstruction), and for TS, it was 0.04° ± 0.13° (ground truth) vs. 0.80° ± 2.13° (reconstruction).
Limitations and Future Work
- Tibial Slope Reconstruction: The authors acknowledge that achieving precise reconstruction and automated measurement of the tibial slope remains a significant challenge. This is attributed to the inherent difficulty in accurately discerning the tibial plateau's complex geometry from sagittal projections alone, especially in cases of osteoarthritis where the plateau may not be even.
- DRR vs. Real EOS Images: The study used DRRs generated from CT scans for training. While this approach mimics the EOS geometry, the appearance of DRRs can differ from real EOS images. Evaluating the framework's performance on actual EOS radiographs is a crucial next step.
- Domain Adaptation: Future work could explore domain adaptation techniques (e.g., CycleGAN) to bridge the gap between DRRs and real X-ray images.
- Patient Positioning Variability: The standing position of patients during image acquisition can vary due to factors like pain or radiologist instructions. Investigating the impact of this variability on reconstruction accuracy is suggested.
- Patient-Specific Instruments (PSIs): The authors propose exploring the use of PSIs created from these 2D-3D reconstructed models for surgical navigation, as PSIs rely heavily on the accuracy of the underlying bone model.
Summary updated