Abstract
Osteoporosis is a common condition, and treatment can reduce the risk of fracture and extend healthy life expectancy, but most cases go undiagnosed and untreated. Dual-energy X-ray absorptiometry (DXA), the gold standard for diagnosing osteoporosis, is costly, time-consuming, and labor-intensive, with limited availability in low-resource settings and small clinics, so it is not suitable for screening for potential osteoporosis. To address this problem, in recent years, some studies have attempted to screen for osteoporosis by estimating DXA bone mineral density (BMD) from chest radiographs (CR), which are frequently used in daily clinical practice, by applying deep learning technology. Although these models have shown good screening performance, the performance of the external data from different facilities and equipments still requires further investigation. In this study, we developed deep learning models for osteoporosis screening and determined the performance of internal and external data. The performance on internal data was good across all models, accurately predicting osteoporosis diagnosed by DXA. Performance on external data exceeded that of calcaneal quantitative ultrasound (QUS), which is widely used as a screening tool for osteoporosis. The screening performance for external data was poor compared to internal data, but by mixing at least 500 external data into the training data, the model could be calibrated and the performance improved. Our results demonstrate that the model can easily perform osteoporosis screening from CR, the most commonly performed imaging test worldwide, without additional invasiveness or cost.