OBJECTIVE: To compare diagnostic performance of four radiomics-based machine learning models for detecting Modic type 1-changes of the lumbar spine in photon-counting detector (PCD)-CT images, using MRI as the reference standard. MATERIALS AND METHODS: In this retrospective single-center study, 60 patients who underwent lumbar spine PCD-CT and MRI within a one-week interval showing Modic type 1-changes were analyzed. A total of 105 radiomic features were extracted from 360 segmented vertebrae, of which 348 were included in the final analysis after quality control. Least Absolute Shrinkage and Selection Operator (LASSO), Random Forest, Extreme Gradient Boosting (XGBoost), and support vector machines (SVM) were trained and evaluated using nested cross-validation. Discriminatory performance of the models was evaluated by area under the receiver operating characteristic curve (AUC). AUC values were compared using the DeLong Test with Benjamini-Hochberg correction to adjust for multiple testing. Diagnostic accuracy was assessed by calculating sensitivity, specificity and F1-score for each model. RESULTS: LASSO achieved the highest AUC (0.842, 95% CI 0.793-0.891), pairwise comparisons did not show significant differences across the models (p ≥ 0.337). Sensitivity was highest for LASSO (0.756, 95% CI 0.662-0.846), whereas specificity was highest for SVM (0.929, 95% CI 0.896-0.958). The highest F1-score was observed for LASSO (0.605, 95% CI 0.521-0.679). CONCLUSION: Four radiomics-based machine learning models demonstrated similar high discriminatory performance but differing diagnostic accuracy for detecting Modic type 1-changes on PCD-CT images. These results support the feasibility of radiomics for evaluation of pathologies beyond visual inspection, although further validation is required to determine clinical applicability.
Marth et al. (2026) studied this question.