Performance evaluation of deep learning-based osteoporosis diagnostic models with conventional chest X-ray in a clinical cohort

Bona Koo1,Yelin Roh1,Gyubin Shin1,Yujin Yang1,Rena Lee2,3,Sungho Cho3,So Hyun Ahn4,5,Kwan-Chang Kim6

1School of Medicine, Ewha Womans University, Seoul, South Korea; 2Department of Biomedical Engineering, College of Medicine, Ewha Womans University, Seoul, South Korea; 3REMEDI Research and Development Center, Seoul, South Korea; 4Ewha Medical Research Institute, School of Medicine, Ewha Womans University, Seoul, South Korea; 5Ewha Medical Artificial Intelligence Research Institute, Ewha Womans University College of Medicine, Seoul, South Korea; 6Department of Thoracic and Cardiovascular Surgery, School of Medicine, Ewha Womans University, Seoul, South Korea

Contributions: (I) Conception and design: B Koo, Y Roh; (II) Administrative support: SH Ahn, KC Kim; (III) Provision of study materials or patients: None; (IV) Collection and assembly of data: All authors; (V) Data analysis and interpretation: G Shin, Y Yang; (VI) Manuscript writing: All authors; (VII) Final approval of manuscript: All authors.

Correspondence to: So Hyun Ahn, PhD. Ewha Medical Research Institute, School of Medicine, Ewha Womans University, 25 Magokdong-ro 2-gil, Gangseo-gu, Seoul 07804, South Korea; Ewha Medical Artificial Intelligence Research Institute, Ewha Womans University College of Medicine, Seoul, South Korea. Email: mpsohyun@ewha.ac.kr; Kwan-Chang Kim, MD, PhD. Department of Thoracic and Cardiovascular Surgery, School of Medicine, Ewha Womans University, 25 Magokdong-ro 2-gil, Gangseo-gu, Seoul 07804, South Korea. Email: mdkkchang@ewha.ac.kr.

Background: Dual-energy X-ray absorptiometry (DXA) is the gold standard for diagnosing osteoporosis; however, its limited accessibility often hinders routine screening in primary care settings. To address this gap, we developed and evaluated a deep learning-based model, PROS® CXR: OSTEO (Promedius, Inc., Seoul, South Korea), which predicts osteoporosis from conventional chest radiographs.

Methods: This retrospective study included 80 adult patients who underwent both DXA and chest radiography within a 3-month interval. The deep learning model, based on convolutional neural networks and trained via transfer learning, generated osteoporosis predictions from chest X-rays. Model performance was assessed against DXA-derived T-scores of the femur and lumbar spine, using either the minimum or average T-score per site as the reference standard.

Results: The proposed model achieved an area under the curve (AUC) of 0.94 for femur and 0.93 for lumbar spine predictions. For osteoporosis screening, the sensitivity and specificity were 90% and 81%, respectively. Subgroup analysis demonstrated higher predictive performance in female patients, whereas false-positives (FPs) occurred more frequently in males.

Conclusions: The PROS® CXR: OSTEO model enables opportunistic and low-cost osteoporosis screening using routine chest radiographs. This approach holds promise for early detection in aging populations and resource-limited settings. Further optimization is required to improve specificity and minimize FPs before clinical implementation.

More