Enhanced opportunistic CT screening for osteoporosis using Machine learning derived volumetric vertebral and complementary body composition information

Abstract

Objectives: To assess whether integrating volumetric vertebral and body composition features obtained from deep learning segmentation of CT images enhances the prediction of bone mineral density (BMD) and the classification of osteoporosis compared to single-slice lumbar vertebral attenuation. Methods: This retrospective study included 383 adults (mean age 59.8 years; 50.1 % women) undergoing routine health check-ups, with same-day abdomen CT scans and dual-energy X-ray absorptiometry (DXA). A two-stage 3DnnU-Net was developed using 475 CT scans from clinical and public datasets to segment individual thoracolumbar vertebrae. Muscle and fat were segmented using a predeveloped 3D U-Net (DeepCatch). Using these segmentations, prediction models were built to estimate DXA-derived lumbar spine, femoral neck, and total hip BMD based on vertebral features alone, combined vertebral and body composition features, and these features plus clinical data (age, sex, body mass index). Model performance was compared against conventional linear regression using single-slice lumbar (L1) attenuation. Results: Compared with lumbar vertebral attenuation alone, the model using volumetric vertebral features significantly improved BMD prediction (lumbar spine correlation coefficient: 0.92 vs. 0.56; P < 0.001) and osteoporosis classification (AUROC = 0.95 vs. 0.87, P = 0.004). Adding body composition metrics further enhanced hip BMD predictions and significantly increased sensitivity in osteoporosis classification (86 % vs. 76 %; P = 0.046), maintaining high specificity (95 %). Incorporating clinical variables provided no additional benefit. Conclusion: DL segmentation-based integration of volumetric vertebral and body composition features enables accurate prediction of lumbar and femoral BMD and improves sensitivity for osteoporosis detection.