Balel Yunus, Sağtaş Kaan, Bülbül Havva Nur
Department of Oral and Maxillofacial Surgery, Faculty of Dentistry, Sivas Cumhuriyet University, Atatürk Blv. 1B, Mehmetpaşa, Merkez, 58040, Sivas, Türkiye.
SEMRUK Technology Inc. Cumhuriyet Teknokent, Sivas, Türkiye.
BMC Oral Health. 2025 Jul 16;25(1):1172. doi: 10.1186/s12903-025-06420-5.
To develop and evaluate a deep learning-based model for automatic dental age estimation using the Demirjian method on panoramic radiographs, and to compare its performance with the traditional manual approach.
A total of 4,800 panoramic radiographs (mean age: 10.64 years) were used to train, validate, and test a YOLOv11-based deep learning model for tooth development staging. Model performance was evaluated using precision, recall, F1 score, and mAP metrics. In addition, a separate dataset of 650 individuals (325 females, 325 males) was used to compare chronological age, manual Demirjian assessments, and AI-assisted estimations through repeated-measures ANOVA and linear regression analysis.
The model achieved its highest performance in the 2nd Molar-H group (Precision: 0.99, Recall: 1.0, F1: 0.995), and its lowest in the 1st Molar-B group (Precision: 0.471, F1: 0.601). Both manual and AI-assisted Demirjian methods significantly overestimated chronological age (p < 0.001), but no significant difference was observed between them (p = 0.433). Regression analysis indicated a weak but statistically significant relationship between age and prediction error, more pronounced in the AI-assisted model (R² = 0.042).
The AI-assisted system demonstrated comparable accuracy to the manual Demirjian method and showed higher performance in later stages of tooth development. The developed Python script and graphical interface allow for rapid, scalable, and user-friendly application of the method. While the system shows promise for use in clinical and forensic settings, broader validation with diverse populations and alternative model architectures is recommended before clinical deployment.
开发并评估一种基于深度学习的模型，该模型使用全景X光片上的德米尔坚方法自动估计牙齿年龄，并将其性能与传统的手动方法进行比较。
总共4800张全景X光片（平均年龄：10.64岁）用于训练、验证和测试基于YOLOv11的深度学习模型，以进行牙齿发育分期。使用精确率、召回率、F1分数和平均精度均值指标评估模型性能。此外，一个由650人（325名女性，325名男性）组成的单独数据集用于通过重复测量方差分析和线性回归分析比较实际年龄、手动德米尔坚评估和人工智能辅助估计。
该模型在第二磨牙-H组中表现最佳（精确率：0.99，召回率：1.0，F1：0.995），在第一磨牙-B组中表现最差（精确率：0.471，F1：0.601）。手动和人工智能辅助的德米尔坚方法均显著高估了实际年龄（p < 0.001），但两者之间未观察到显著差异（p = 0.433）。回归分析表明年龄与预测误差之间存在微弱但具有统计学意义的关系，在人工智能辅助模型中更为明显（R² = 0.042）。
人工智能辅助系统显示出与手动德米尔坚方法相当的准确性，并且在牙齿发育后期表现出更高的性能。所开发的Python脚本和图形界面允许该方法快速、可扩展且用户友好地应用。虽然该系统在临床和法医环境中有应用前景，但在临床部署之前，建议使用不同人群和替代模型架构进行更广泛的验证。
