Opportunistic Detection of Osteoporosis Using Routine CT Scans: Development and Validation of an Artificial Intelligence-Based Screening Model

Background: There is a significant underdiagnosis of osteoporosis despite its association with an increased fracture risk, the morbidity burden and increased healthcare costs. Opportunistic osteoporosis screening by routine computed tomography (CT) without imaging or radiation exposure provides an opportunity for routine screening for osteoporosis. The main purpose of this study was to create and externally validate an artificial intelligence (AI) model of automated osteoporosis detection using a routinely acquired computed tomography (CT) scan.

Materials and methods: A total of 1,862 adults (≥50 years old) underwent routine imaging with both computed tomography (CT) and dual-energy X-ray absorptiometry (DXA) within 6 months. Vertebral segmentation was performed automatically, and a variety of radiomic and morphometric measures and vertebral attenuations were extracted from the CT data. This data was split into training group (n=1304), validation group (n=279) and testing group (n=279) with external validation in a separate group of 512 patients. A DXA T score ≤ -2.5 was considered to represent osteoporosis. The performance of the models was evaluated with sensitivity, specificity, accuracy, area under the receiver operating characteristic curve (AUC), calibration analysis, and decision-curve analysis.

Results: Overall 28.4% of the study subjects were diagnosed as osteoporosis. Age, female sex, and vertebral attenuation, texture entropy, cortical thickness reduction, and vertebral compression ratio were all independent predictors (p<0.001). The AI model achieved an AUC of 0.944 (95% CI: 0.927–0.961), sensitivity of 91.3%, specificity of 88.6%, and overall accuracy of 89.4% in the test cohort. External validation showed excellent sustained accuracy (AUC 0.932 [0.909–0.955]) and accuracy (87.8%). Compared with a target-fail DXA service, the clinical utility analysis recorded a reduction in unnecessary referrals of 42.7% with a high percentage of DXA services provided prioritizing the detection of osteoporosis.

Conclusion: The design of an AI-driven opportunistic CT screening model was found to be very good with excellent diagnostic accuracy and strong external validity, and had the potential for significant clinical utility. Adoption of this approach into day-to-day radiological practice could improve the results of early detection of osteoporosis at the population level as well as contribute to fracture prevention.

Keywords: Osteoporosis, Artificial Intelligence, Opportunistic Screening, Bone Mineral Density, Radiomics, Diagnostic Accuracy, Vertebral Fracture Risk.