Opportunistic osteoporosis screening using CT-based Hounsfield units: a machine learning approach for preoperative spine surgery planning

Opportunistic assessment of vertebral bone quality using Hounsfield Units (HU) on preoperative computed tomography (CT) may complement dual-energy X-ray absorptiometry (DXA) for osteoporosis screening. We developed a CT-based framework to predict osteoporosis from HU and basic demographics, restricting model development to cases with consistent HU–DXA agreement (“non-deviant”; residual within ± 1 SD of the axial HU–lowest T-score regression). We retrospectively reviewed 162 patients (2018–2025) who underwent lumbar CT and DXA. HU was measured at L1–L4 using axial (three-slice) and sagittal (single-slice) methods; sagittal S1 HU was also obtained. Residual filtering retained 123 non-deviant cases. An XGBoost classifier used 26 features (21 HU measures plus age, sex, height, weight, BMI) and was evaluated in three stratified 80/20 splits (AUC, accuracy, F1, sensitivity, specificity). Feature importance was assessed. In the non-deviant cohort ( n = 123), both axial and sagittal HU values showed strong correlations with DXA lowest T-score (ρ = 0.820 and ρ = 0.766, respectively; both p < 0.001). The XGBoost model achieved an average AUC of 0.804 ± 0.067, sensitivity of 0.857 ± 0.143, specificity of 0.759 ± 0.128, accuracy of 0.787 ± 0.061, and F1 score of 0.694 ± 0.048. HU measurements at L3 and L4, sex, body weight, and height were among the top predictive features. In a carefully selected cohort demonstrating a stable HU–lowest T-score relationship, both axial and sagittal HU measurements demonstrated strong correlations with DXA lowest T-score, supporting their use as surrogate indicators of bone quality. The XGBoost model, trained on CT-derived HU values and demographic data, exhibited moderate classification performance within the non-deviant subset. These findings support the clinical utility of HU-based opportunistic screening as a triage approach to prompt confirmatory assessment and perioperative bone health optimization, particularly in settings where DXA is unavailable. These results represent proof-of-concept performance in HU–DXA–concordant cases and may not generalize to unfiltered populations.