Abstract
OBJECTIVES: Undiagnosed osteoporosis before spinal surgery increases severe complication risks. This study develops the machine learning-based CT radiomics model to preoperatively screen lumbar osteoporosis. MATERIALS AND METHODS: This retrospective study enrolled 166 patients undergoing concurrent dual-energy X-ray absorptiometry (DEXA), spinal CT and MRI. Vertebral data from normal and osteoporotic cases were partitioned into training and validation cohorts (8:2 ratio). A total of 851 radiomics features were extracted from lumbar spine CT scans using the 3D slicer PyRadiomics module. Feature selection employed mRMR (minimum redundancy maximum relevance) for preliminary screening followed by LASSO regression for dimensionality reduction. Four machine learning classifiers were developed: logistic regression (LR), support vector machines (SVM), XGBoost, and random forest (RF). Model performance was assessed through receiver operating characteristic (ROC) analysis with DeLong test comparisons. Clinical utility was quantified via decision curve analysis (DCA). RESULTS: Nine radiomic features based on spine CT images were constructed to develop the model. The radiomic-XGBoost model with the highest area under the curve (AUC) of 0.89 of the training cohort and 0.91 of the test cohort among the machine learning algorithms. The DeLong test showed that the differences between the radiomic-XGBoost, vertebral bone quality (VBQ) and Hounsfield unit (HU) models were statistically significant (p < 0.05). DCA revealed that the radiomics-based model offers a superior net benefit compared to the other two models. CONCLUSION: CT-based machine learning radiomics significantly outperformed VBQ scoring and HU measurements in osteoporosis diagnostic accuracy.