Abstract
Automatic Cobb angle measurement from X-ray images is crucial for scoliosis screening and diagnosis. However, most existing regression-based methods and segmentation-based methods struggle with inaccurate spine representations or mask connectivity/fragmentation issues. Besides, landmark-based methods suffer from insufficient training data and annotations. To address these challenges, we propose a novel framework including Self-Generation pipeline and Low-Rank Approximation representation (SG-LRA) for automatic Cobb angle measurement. Specifically, we propose a parameterized spine contour representation based on LRA, which enables eigen-spine decomposition and spine contour reconstruction. We can directly obtain spine contour with only regressed LRA coefficients, which form a more accurate spine representation than rectangular boxes. Also, we combine LRA coefficient regression with anchor box classification to solve inaccurate predictions and mask connectivity issues. Moreover, we develop a data engine with automatic annotation and automatic selection in an iterative manner, which is trained on a private Spinal2023 dataset. With our data engine, we generate the largest scoliosis X-ray dataset named Spinal-AI2024 largely without privacy leaks. Extensive experiments on public AASCE2019, private Spinal2023, and generated Spinal-AI2024 datasets demonstrate that our method achieves state-of-the-art Cobb angle measurement performance.
Introduction
The Cobb angle is the clinical gold standard for diagnosing and monitoring scoliosis. However, manual measurement is labor-intensive and prone to significant inter-observer variability, with errors typically ranging from 2.8 to 8 degrees. Automating this process using deep learning faces several hurdles. Existing regression-based methods often rely on simple bounding boxes that fail to capture the complex geometry of wedge-shaped or irregular vertebrae. Segmentation-based approaches frequently suffer from mask fragmentation or connectivity issues in low-quality X-rays. Furthermore, landmark-based methods require vast amounts of precisely labeled data, which is difficult to obtain due to medical privacy regulations and the high cost of expert annotation.
To address these challenges, we propose SG-LRA, a framework that introduces a parameterized spine contour representation based on Low-Rank Approximation (LRA) and an iterative data engine for self-generation of training data.
Methodology
The SG-LRA framework consists of two primary innovations: a latent contour representation and an iterative data engine.
Latent Contour Representation (The "Eigen-Spine")
Rather than predicting raw coordinates for dozens of landmarks—which can lead to anatomically inconsistent shapes—SG-LRA utilizes Singular Value Decomposition (SVD) to learn a latent space of spinal shapes. By decomposing spinal contours into "eigen-spines" (principal components of spine geometry), the model only needs to regress a small set of latent contour coefficients. This approach acts as a powerful anatomical prior, ensuring that the predicted spine maintains a realistic structure even in the presence of image noise.
Iterative Data Engine
To overcome the scarcity of labeled medical data, we developed a self-sustaining loop for data generation and annotation.
- Generation: A Latent Diffusion Model (LDM) is trained on a private clinical dataset (Spinal2023) to generate realistic, synthetic scoliosis X-rays.
- Privacy Review: To prevent the leakage of real patient data, a privacy module compares synthetic images against the training set using metrics like SSIM to ensure the model is not merely memorizing images.
- Auto-Annotation: The SG-LRA model pseudo-labels the synthetic data. These labels are then filtered and refined through a manual-assisted process before being fed back into the training set.
Datasets
The performance of SG-LRA was evaluated across three datasets:
- AASCE2019: A public benchmark containing 609 frontal X-rays.
- Spinal2023: A private clinical dataset of 904 images used to initialize the generative model.
- Spinal-AI2024: A new, large-scale dataset of 20,000 synthetic X-ray images with automatic annotations, released to the research community to facilitate further study without privacy risks.
Results and Discussion
SG-LRA achieves state-of-the-art performance on the AASCE2019 benchmark, reaching a Symmetric Mean Absolute Percentage Error (SMAPE) of 5.05%. This significantly outperforms previous landmark-based models like Linformer (7.91%) and segmentation-based models like Augmented U-Net (9.2%).
The model is also highly efficient, operating at 16.7 FPS on an RTX 4090, making it suitable for real-time clinical decision support. A key insight from this work is that Low-Rank Approximation serves as an effective regularizer, embedding anatomical knowledge directly into the output format and preventing the fragmentation issues common in pixel-level segmentation.
Conclusion
SG-LRA demonstrates that combining latent space representations with generative data engines can effectively bypass the data bottlenecks prevalent in medical imaging. While the model shows high accuracy for most spinal curves, challenges remain in the Proximal Thoracic (PT) region, where the model occasionally misidentifies cervical vertebrae as thoracic. Future work will focus on introducing stricter anatomical constraints and region-specific filtering to further refine PT Cobb angle measurements.
Summary updated