Apr 10, 2025
WCO-IOF-ESCEO
Abstract
OSTEO
External validation of a deep-learning based osteoporosis screening method using chest frontal radiograph - A single center, retrospective study
Jeongmin Song¹, Minjee Kim¹, Saerom Park¹, Hyun-Jin Bae¹* Promedius Inc., Seoul, South Korea
Objective: This study aims to externally validate the performance of a deep-learning model for screening osteoporosis using frontal chest radiographs.
Materials and Methods: We retrospectively included individuals (age ≥ 50yrs) who visited Yangji Hospital in Seoul, South Korea, from June 2023 to June 2024, and underwent both a chest posteroanterior radiograph (CXR) and a DXA scan on the same date. For AI analysis, we used commercially available AI software (CXR:OSTEO, Promedius Inc.), which analyses CXR to produce a confidence score (0–1). Based on this score, the individuals are classified as either high or low risk for osteoporosis. We assessed the model’s performance by using the bone mineral density (BMD) category from DXA T-scores as the reference standard. Subgroup analysis was conducted based on age or sex. Also, the model’s AUC was compared with that of calcaneal quantitative ultrasound (QUS).
Results: Among the 500 subjects (76.4% female; mean age, 63.3 years), 100 individuals had osteoporosis, while 400 did not. The model demonstrated an AUC of 0.866 (95% CI: 0.827–0.905). Using the pre-defined threshold 0.5 for binary classification, the sensitivity, specificity, PPV, and NPV were 77.0%, 80.3%, 49.4%, and 93.3%, respectively.
In the sex-based subgroup analysis, the model achieved AUCs of 0.87 and 0.80 for female and male subjects, respectively. The AUC ranged from 0.82 to 0.91 among different age groups Compared with QUS, the model's AUC of 0.866 was higher than the previously reported QUS AUCs (0.66–0.766).
Conclusion(s): The deep-learning model demonstrated robust performance in identifying individuals with osteoporosis from CXRs during external validation. This highlights its potential as a screening tool for osteoporosis, utilizing CXRs, the most widely used imaging modality worldwide.
Disclosures
J.Song, S.Park, M.Kim - employee of Promedius Inc.
H.Bae - Board member and shareholder of Promedius Inc.
J.Kim - principal investigator of the study
