Refining Early Detection of Low Bone Mineral Density: A Deep Learning Model for Osteopenia Screening Using Chest Radiographs

Objective(s): We developed and validated a deep learning model to classify osteoporosis, osteopenia, and normal bone mineral density (BMD) from standard frontal chest radiographs. This framework offers finer categorization, enabling earlier identification of osteopenic individuals and more timely DEXA referrals for those at moderate risk.

Materials and Methods: We expanded our PROS CXR:OSTEO product—which initially achieved high accuracy (AUC = 0.94 in internal validation) in classifying osteoporosis versus non-osteoporosis—to further differentiate osteopenia from normal within the non-osteoporosis category. Training data comprised 46,048 chest radiographs from Hospital A (South Korea), categorized by DXA T-scores, augmented with 16,265 unlabeled images from public datasets, totaling 62,313. For internal validation, 1,989 chest radiographs from Hospital A were used. For external validation, we collected datasets from three independent institutions spanning secondary healthcare (Hospital B), a facility serving veterans (Hospital C), and a global platform (Platform D).

Results: Our framework achieved an AUC of 0.93, sensitivity of 0.94, and specificity of 0.72 in classifying normal and osteopenia on External A (88.9% female; mean age 59.0 years), an AUC of 0.83, sensitivity of 0.78, and specificity of 0.82 on Hospital B (55.5% female; mean age 59.3 years), an AUC of 0.83, sensitivity of 0.73 and specificity of 0.67 on Hospital C (44.8% female; mean age 73.6 years), and an AUC of 0.78, sensitivity of 0.75, and specificity of 0.77 on Platform D (95.4% female; mean age 67.2 years).

Conclusion(s): Our model effectively classified osteopenia from normal BMD within the non-osteoporosis category, helping identify individuals at risk of reduced bone density and potentially enabling earlier intervention for osteopenia and osteoporosis.