Prediction of the Number of Root Canals in Mandibular Premolars Using Artificial Intelligence

Accurate determination of root canal count is essential for successful endodontic therapy, especially in mandibular premolars, which often present substantial anatomical variability. Although periapical (PA) and panoramic (OPG) radiographs are widely used because they are accessible and low‑risk, their 2D nature limits the ability to distinguish overlapping structures. Cone‑beam computed tomography (CBCT) offers superior 3D visualization but is not routinely employed due to higher radiation dose and cost. This study therefore aimed to evaluate how effectively an artificial intelligence (AI) model could predict the number of root canals in mandibular premolars using PA and OPG images.

PA and OPG radiographs were paired with corresponding CBCT scans for analysis. All images were processed, and CBCT interpretations were performed by two calibrated examiners. After preprocessing and labeling, the dataset was divided into training (80%), validation (10%), and test (10%) subsets. The YOLOv8 model was trained for 30 epochs with a batch size of 16. Performance metrics included accuracy, sensitivity, specificity, F1‑score, AUC with 95% confidence intervals, and concordance with CBCT findings.

A total of 463 radiographs (248 OPG, 215 PA) were assessed. Based on CBCT, teeth were categorized as single‑canal (76.2%), two‑canal (14.9%), or three‑canal (8.9%). YOLOv8 achieved 92.3% accuracy (AUC 0.95) on PA images and 88.7% accuracy (AUC 0.92) on OPG images, with particularly strong performance in identifying three‑canal cases (87.1%). Inter‑examiner agreement for CBCT evaluation was excellent (κ = 0.98).

In conclusion, YOLOv8 showed strong capability in predicting root canal anatomy from routine 2D radiographs. Nonetheless, broader multicenter studies with larger datasets are required to confirm the model’s generalizability and long‑term reliability.