Abstract
Background/Objectives: Accurate identification of deciduous and permanent teeth on panoramic radiographs (PRs) during mixed dentition is fundamental for early detection of eruption disturbances, yet relies heavily on clinician experience due to developmental variability. This study aimed to develop a deep learning model for automated tooth detection and segmentation in pediatric PRs during mixed dentition. Methods: A retrospective dataset of 250 panoramic radiographs from patients aged 6-13 years was analyzed. A customized YOLOv11-based model was developed using a novel hybrid pre-annotation strategy leveraging transfer learning from 650 publicly available adult radiographs, followed by expert manual refinement. Performance evaluation utilized mean average precision (mAP), F1-score, precision, and recall metrics. Results: The model demonstrated robust performance with mAP(0.5) = 0.963 [95%CI: 0.944-0.983] and macro-averaged F1-score = 0.953 [95%CI: 0.922-0.965] for detection. Segmentation achieved mAP(0.5) = 0.890 [95%CI: 0.857-0.923]. Stratified analysis revealed excellent performance for permanent teeth (F1 = 0.977) and clinically acceptable accuracy for deciduous teeth (F1 = 0.884). Conclusions: The automated system achieved near-expert accuracy in detecting and segmenting teeth during mixed dentition using an innovative transfer learning approach. This framework establishes reliable infrastructure for AI-assisted diagnostic applications targeting eruption or developmental anomalies, potentially facilitating earlier detection while reducing clinician-dependent variability in mixed dentition evaluation.