Aureliana Caraiane, Cristina Gabriela Pușcașu, Gheorghe Raftu, Cristina Bartok-Nicolae, Claudia Elena Sin, Steliana Gabriela Buștiuc, Erdogan Elvis Șachir
ABSTRACT
Aim of the study To assess an AI model’s ability to detect the second mesiovestibular (MB2) canal in maxillary first molar periapical radiographs of adults and compare its performance with experienced human readers using intraoperative findings as ground truth. Materials and methods Fifty periapical radiographs (40 with clinically confirmed MB2 canal, 10 without) were evaluated. Two blinded endodontists provided consensus readings. A CNN-based AI model, trained on a separate annotated dataset with intraoperative confirmation, generated binary predictions and heatmap overlays. Performance metrics (sensitivity, specificity, accuracy, AUC-ROC) with 95% CIs were calculated; inter-reader agreement was assessed via Cohen’s kappa. McNemar’s and DeLong’s tests compared human versus AI results. Results Human readers achieved sensitivity 55.0% (95% CI: 38.5–70.7%), specificity 80.0% (95% CI: 49.7–95.6%), accuracy 60.0% (95% CI: 45.2–73.6%), and AUC-ROC ~0.68. Inter-observer kappa was ~0.64. The AI model showed sensitivity 85.0% (95% CI: 70.2–94.3%), specificity 90.0% (95% CI: 55.5–99.7%), accuracy 86.0% (95% CI: 73.3–94.2%), and AUC-ROC ~0.90. Differences in sensitivity, accuracy, and AUC-ROC were significant (p < 0.01). Pie charts indicated human readers correctly detected MB2 canals in 60.0% of cases (missed 36.0%, false positives 4.0%), whereas AI achieved 86.0% correct detection (missed 12.0%, false positives 2.0%). Conclusions AI-assisted analysis markedly enhances MB2 canal detection on periapical radiographs compared to human readers, reducing missed canals and potentially improving endodontic planning. Incorporating AI overlays into routine radiograph review is advised, with advanced imaging reserved for unclear cases.
Loading...
Reload document 
Open in new tab