Automated detection of C-shaped canals in mandibular second molars from panoramic radiographs: comparing single and ensemble convolutional neural networks within a 2-stage pipeline

This study aimed to develop and validate a two-stage deep-learning workflow that automatically localizes mandibular second molars on panoramic radiographs and classifies C-shaped canal presence as an explainable, clinically deployable decision-support tool. A total of 1,252 panoramic radiographs (2015–2025) were retrieved from a digital archive. Stage 1 employed YOLOv8 to detect teeth #37 and #47. Stage 2 applied three CNNs (DenseNet-169, EfficientNet-B6, ConvNeXt-Base) and ensemble approaches to classify C-shaped canals on YOLO-derived crops. Data were split 70/15/15 for training/validation/testing. Detection was evaluated using IoU and mAP; classification via accuracy, sensitivity, specificity, F1-score, and AUC. Pairwise AUC comparisons used DeLong test with Bonferroni correction. Grad-CAM provided visual explanations. YOLOv8 achieved high localization performance, reliably supplying regions of interest to classifiers. ConvNeXt-Base demonstrated the most consistent discrimination with top-tier accuracy and F1-score. A binary EfficientNet-ConvNeXt ensemble performed comparably to the best individual model, while the tertiary ensemble showed no improvement. DeLong analysis revealed statistically significant AUC advantage of ConvNeXt-Base over DenseNet-169 for tooth #47 after correction. This two-stage pipeline enables accurate, automation-ready detection of C-shaped canals from routine panoramic radiographs, potentially informing patient triage and CBCT indication decisions. Multicenter validation is warranted before clinical deployment. This AI system offers a practical screening tool using existing panoramic radiographs to identify high-risk patients, potentially improving treatment planning efficiency, reducing unnecessary CBCT referrals, and alerting clinicians to complex anatomy before endodontic therapy.