Differentiation of cervical caries and cervical burnout on orthopantomographic images: a radiomics-based machine learning approach

To evaluate the diagnostic performance of machine learning (ML) algorithms based on radiomic features extracted from orthopantomographic (OPG) images in differentiating cervical caries from cervical burnout. A total of 170 OPG images were retrospectively analyzed, including 85 cervical caries cases and 85 age- and sex-matched cervical burnout cases. Following intra- and interobserver reproducibility filtering, dimensionality reduction was performed using Variance Threshold, SelectKBest, and least absolute shrinkage and selection operator methods. Four ML classifiers, Random Forest (RF), K-Nearest Neighbors (KNN), Logistic Regression (LR), and Decision Tree (DT), were compared using repeated stratified five-fold cross-validation in the training cohort. The classifier with the highest mean AUC was selected and evaluated on the held-out test set. Additional internal validation included bootstrap optimism correction, nested cross-validation, and repeated train-test split analysis. Feature selection yielded eight radiomic features from first-order and texture feature classes. RF achieved the highest mean cross-validated AUC (0.68 ± 0.09) and was selected as the final classifier. The locked RF model yielded an AUC of 0.57 on the held-out test set. Bootstrap optimism correction yielded an AUC of 0.64 (95% CI: 0.51–0.75), while nested cross-validation yielded a mean AUC of 0.65 ± 0.13. Across 10 repeated train-test splits, the mean AUC was 0.65 ± 0.09 (range: 0.47–0.79). Radiomics-based ML may provide quantitative imaging information relevant to differentiating cervical caries from cervical burnout on OPG images. Given the modest sample size and variability across internal validation analyses, these findings should be considered proof-of-concept and require external validation before clinical application.