https://atlas.rsna.org/schemas/2025-11/model.json

SpineTK: Automated Cobb angle measurement on AP radiographs

https://pubs.rsna.org/doi/10.1148/ryai.220158

1.0

Supported by National Institutes of Health National Institute of Arthritis and Musculoskeletal and Skin Diseases (R01AR068382).

Institutional review board approval was obtained; imaging studies were analyzed in a HIPAA-compliant manner.

Region-based convolutional neural network using a modified Mask R-CNN backbone (SpineTK). Detects vertebral bodies and their four corner points; fits a 6th-degree polynomial to vertebral centroids to identify curve inflection points and compute the Cobb angle from endplates of vertebrae nearest inflection points.

Open-source code and documentation: https://github.com/abhisuri97/SpineTK

Automated, rapid, and accurate Cobb angle measurement to aid diagnosis and longitudinal monitoring of scoliosis, including in presence of surgical hardware.

Clinical decision support systems; workflow optimization for radiographic measurement and follow-up monitoring.

Fully automated Cobb angle measurement from input AP radiographic images without manual input.

Automated measurement of the major Cobb angle on standing anteroposterior full-spine radiographs (including EOS low-dose stereoradiographic images and stitched digital radiographs) in patients evaluated for scoliosis, including those with spinal hardware present.

AP EOS DICOM images and stitched AP digital radiographs resized to 800×800 pixels.

Provide standing AP full-spine radiograph. The network detects vertebral bodies and corner points, fits a curve through centroids, and outputs the major Cobb angle. Authors note a practical flag may be warranted for radiologist review when only a few vertebral bodies are detected due to artifact/occlusion.

Performance can degrade when extensive hardware/artefact occludes much of the spine, potentially reducing detected vertebrae; however, errors did not exceed 5° in analyzed cases. Endplates selected by the network differed from annotators in 7.5% of endplates, which could impact surgical planning. Training EOS dataset under-represented younger patients (≤20 years), potentially biasing features; racial demographics unavailable. Scanning parameters for EOS and external radiographs not available, which may affect replicability.

Can be used to monitor scoliosis progression with rapid, accurate, hardware-invariant measurements; consider flagging cases with few detected vertebrae for radiologist review.

Fivefold cross-validation used to select best fold; early stopping; hyperparameters tuned by grid search. Code repository provided for reproducibility.

Inference time under 0.5 second per image on reported setup; training performed on Google Colab.