Report on developing and externally validating a deep learning chest radiograph model that screens for early thoracic ligamentum flavum ossification using multicenter cohorts.
Key Takeaways
- ResNet101 lateral-view model achieved 97.0% accuracy and 94.0% sensitivity in internal testing
- External validation showed AUCs of 0.995 for lateral and 0.954 for frontal radiographs with high specificity
- DL model outperformed experienced spine surgeons and imaging physicians on sensitivity and overall accuracy