Rationale and Objectives: Osteoporotic vertebral compression fractures (OVFs), particularly at the thoracolumbar spinal junction spanning levels T11 to L2, represent a significant and debilitating global health challenge. The objective of the present research was to establish and verify a robust radiomics model using routine non-contrast computed tomography (CT) to predict thoracolumbar OVFs risk. Materials and Methods: In this retrospective cohort study, 80 patients with new thoracolumbar OVFs were propensity-matched 1:2 with 160 controls. A 3D U-Net automatically segmented T11-L2 cortical and cancellous bone. Following minimum redundancy maximum relevance (mRMR) and least absolute shrinkage and selection operator (LASSO) feature selection, 13 cortical and 16 cancellous radiomics features were extracted. Logistic regression models were developed using corresponding radscores. The best-performing model was compared to volumetric bone mineral density (vBMD) and evaluated via ROC curves, decision curve analysis (DCA), and calibration plots. Results: A total of 240 patients (147 females, 93 males) were enrolled, with no significant age or sex differences between groups. The combined cortical and cancellous radscore model (vertebral model) attained an AUC of 0.825 and 0.840 for OVF prediction, outperforming the vBMD model (AUC: 0.752/0.735). DCA and calibration plots verified its outstanding predictive performance. Notably, integrating vBMD with the vertebral model did not yield a statistically significant improvement ( p > 0.05). The selected features highlighted crucial microstructural insights, with cancellous features reflecting trabecular heterogeneity and cortical features indicating mineralization uniformity and integrity. Conclusion: Our findings demonstrate that radiomics derived from routine non-contrast CT, leveraging deep learning for automated bone compartment segmentation, offers a superior and practical tool for Early identification of high-risk individuals for thoracolumbar OVFs. This approach provides valuable, complementary information beyond vBMD, potentially enhancing clinical decision-making and reducing the burden of osteoporosis. Keywords: osteoporotic vertebral fractures, deep learning, radiomics, fracture prediction
Che et al. (Fri,) studied this question.