Machine learning for the prediction of osteopenia/osteoporosis using the CT attenuation of multiple osseous sites from chest CT.

Purpose: To use machine learning and the CT attenuation of all bones visible on chest CT scans to predict osteopenia/osteoporosis.

Method: We retrospectively evaluated 364 patients with CT scans of the chest, and Dual-energy X-ray absorptiometry (DXA) scans within 6 months of each other. Studies were performed between 01/01/2015 and 08/01/2021. Volumetric segmentation of the ribs, thoracic vertebrae, sternum, and clavicle was performed using 3D Slicer to obtain the mean CT attenuation of each bone. The study sample was randomly split into training/validation (80 %, n = 291 patients) and test (20 %, n = 73 patients) datasets. Univariate analyses were used to identify the optimal CT attenuation thresholds to diagnose osteopenia/osteoporosis. We used penalized multivariable logistic regression models including Least Absolute Shrinkage and Selection Operator (LASSO), Elastic Net, and Ridge regression, and Support Vector Machines (SVM) with radial basis functions (RBF) to predict osteopenia/osteoporosis and compared these results to the CT attenuation threshold at T12.

Results: There were positive correlations between the CT attenuation between all bones (r > 0.6, P < 0.001 for all). There were positive correlations between CT attenuation of the bones and the L1-L4 BMD T-score, total hip T-score, and femoral neck T-scores (r > 0.4, P < 0.001 for all). A CT attenuation threshold of 170.2 Hounsfield units (HU) at T12 had an AUC of 0.702, while a threshold of 192.1 HU at T4 had an AUC of 0.757. The SVM with RBF had the highest AUC (AUC = 0.864) and was better than the LASSO (P = 0.011), Elastic Net (P = 0.011), Ridge regression (P = 0.011) but was not better than using the CT attenuation at T12 (P = 0.060).

Conclusions: The CT attenuation of the ribs, thoracic vertebra, sternum, and clavicle can be used individually and collectively to predict BMD and to predict osteopenia/osteoporosis.