Deep learning-based detection of root numbers in maxillary premolars

Official URL: https://dx.doi.org/10.1111/iej.70091

Aim: This study aimed to detect the root number of maxillary premolars on panoramic radiographs using deep learning models. Methodology: This retrospective study included 925 maxillary premolars from 350 patients with panoramic radiographs and CBCT scans, which served as the reference standard to determine root numbers. Panoramic images were cropped to isolate the premolar root region, preprocessed, resized, and used to train three convolutional neural network (CNN) models (AlexNet, DenseNet-121, EfficientNet-B0) equipped via transfer learning. Data augmentation was applied to address class imbalance. Five-fold cross-validation was performed, with each fold allocating 86% of the data (n = 797) for training, 7% (n = 64) for validation and 7% (n = 64) for testing, without patient-level overlap. An independent external validation set was also constructed to assess generalizability. An experienced endodontist evaluated the same external validation set for comparison. Accuracy, sensitivity, specificity, F1-score, and the area under the ROC curve (AUC) were calculated as mean ± standard deviation (SD) with 95% confidence intervals (CI). The ensemble model was used to improve robustness. Paired DeLong tests were performed to compare AUCs for both cross-validation predictions (n = 322) and external validation predictions (n = 148). Results: In cross-validation, the ensemble achieved the strongest performance (accuracy 0.90, F1-score 0.90, sensitivity 0.89, specificity 0.91, AUC 0.94). Among individual models, EfficientNet-B0 and AlexNet performed similarly (accuracy 0.85), while DenseNet-121 performed lower (0.81). DeLong analyses confirmed significantly higher AUCs for the ensemble compared with all individual models (p < 0.05). On the external validation set, the ensemble again performed best (accuracy 0.87), followed by AlexNet (0.85), EfficientNet-B0 (0.84) and DenseNet-121 (0.84). The expert clinician achieved an accuracy of 0.82. DeLong comparisons on external validation predictions revealed no significant AUC differences among models (all p > 0.05), except for AlexNet vs. the ensemble. Confidence interval plots confirmed the ensemble's reduced variability and narrowest CIs. Conclusions: Deep learning models showed reliable performance in predicting root numbers of maxillary premolars from panoramic radiographs, with the ensemble model achieving the most stable and accurate results. These findings indicate that deep learning systems may serve as a supportive tool in clinical decision-making.