Dental age estimation has been used to obtain biological profiles for unidentified bodies, to address issues related to welfare benefits such as school enrollment and pension eligibility, and to determine the legal age of majority or minority for criminal prosecution [1,2].

Dental age estimation in children is based on tooth development, which is typically categorized into several stages to estimate age. Demirjian et al. [3] introduced a method that classifies tooth development into eight stages for the seven permanent mandibular left teeth and provided age estimation method based on these stages. Demirjian’s criteria remain the most widely used approach for evaluating the developmental stages of permanent teeth [4]. Other classifications, such as those proposed by Gustafson and Koch [5] and Moorrees et al. [6], have also been used to assess the degree of tooth development [7]. And various statistical techniques have been used to estimate age based on tooth development [8-10]. There are also methods for estimating age by comparing tooth development to a standardized dental atlas [11,12].

In recent years, artificial intelligence (AI) technologies, including machine learning and deep learning, have been introduced across various fields, leading to transformative changes. In particular, deep learning based computer vision techniques have been integrated into the medical field, where they are being utilized for tasks such as automated focal diagnosis, interpreting radiographs, and analyzing pathological tissue images. And many studies have explored age estimation using these technologies. Among them, age and sex estimation from facial images is one of the most actively researched areas in AI-driven image processing and computer vision [13]. More recently, studies using orthopantomograms (OPGs) to estimate age have reported very high accuracy, particularly in adults, where accurate results are achieved without the need for invasive procedures such as tooth extraction [14-16].

This study aimed to evaluate the applicability of a convolutional neural network (CNN) technique for dental age estimation in children by selecting the mandibular dentition as the region of interest (ROI) in OPGs of Korean children.

This study was conducted using OPGs of patients aged 3.00 to 14.99 years who visited the Chosun University Dental Hospital. A total of 2,856 OPGs were collected, 238 for each single-year age group to ensure equal representation, with an equal ratio of males and females. This study was conducted retrospectively to utilize existing clinical data efficiently, avoiding the need for additional radiographs and ensuring adherence to ethical guidelines. All data were anonymized to protect patient privacy. There was no risk of harm to the subjects in this study. The study was approved by the Institutional Review Board (IRB) of Chosun University Dental Hospital, with a waiver of informed consent (approval No. CUDHIRB 2206001).

We derived and validated a CNN computational model to estimate age using the mandibular dentition of the entire dentition. This is because, similar to the Demirjian method, age estimation in children is typically performed using the mandibular teeth, which have a simpler anatomy compared to maxillary teeth and lack overlapping structures such as the maxillary sinus. The research process of this study is shown in Fig. 1. To extract the lower dentition region as an ROI from the entire OPG image, we used You Only Look Once (YOLO) V8, a deep learning object detection technique [17]. We randomly selected 200 OPGs stratified by age, designated the lower dentition region as an ROI with the annotation tool LabelImg (https://github.com/tzutalin/labelImg), and trained a mandibular dentition detection model using YOLOv8 with a learning rate of 0.01 and 200 epochs, and validation was conducted using the training set (precision: 0.999, recall: 1.000, mAP50: 0.995, mAP50-95: 0.850) (Fig. 2). And the mandibular dentition regions were then extracted from the entire OPGs using this trained object detection model and regression analysis was performed on the image and age using CNN image classification algorithm. Due to the relatively small number of images, the 2,856 images were divided into 80% training set and 20% test set (stratified by age) to perform image classification, and validation was performed on the test set to derive a classification model. For regression analysis, we used an ensemble (stacking technique) of two backbone models, ResNet50 [18] and VGG19 [19], which are widely used in CNN analysis. Epochs were repeated 200 times, and the mean absolute error (MAE) and root mean squared error (RMSE) of the derived model were calculated to obtain the error. MAE was chosen as the primary evaluation metric due to its ability to directly quantify prediction accuracy by averaging the absolute differences between predicted and actual ages. Models with the lowest MAE on the test set were considered optimal for age estimation.

For this study, we used Google Colab Pro (L4 Tensor Core GPU, NVIDIA) based on python, and CNN image classification was performed using TensorFlow v. 2.5.0.

The age estimation model developed in this study, utilizing an ensemble backbone of ResNet50 and VGG19, demonstrated relatively high accuracy despite the small sample size. The results of this study are presented in Table 1. The model achieved MAE of 0.501 and RMSE of 0.742 on the test set. Naturally, the training set exhibited lower error values, with the MAE and RMSE being 0.273 and 0.354, respectively. Fig. 3 illustrates the progression of MAE and RMSE in the training and test sets over the epochs (the first 10 epochs are excluded due to high error and significant fluctuations). The errors in both sets decrease steadily before stabilizing, suggesting effective learning, although minor fluctuations in the test set indicate potential variability in generalization. The distribution of actual and predicted ages is shown in Fig. 4. The coefficient of determination (R²) was found to be 0.951, indicating that the predicted values are relatively well aligned with the actual values.

This study aimed to automate age estimation using a CNN by designating the lower dentition region as the ROI and achieved relatively high accuracy despite the limited sample size. The proposed method can be applied to estimate the age of children in practical cases. Notably, the ability to reduce human error through automation is a significant advantage in forensic practice.

Several studies have previously analyzed dental development in Korean children. For instance, Lee et al. [20]’s study used all 28 permanent teeth to develop an age estimation model based on tooth development stages using Demirjian’s criteria, achieving standard errors of 0.63 for males and 0.62 for females, indicating very high accuracy. In 2011, Lee et al. [21] analyzed the age correlation of seven permanent teeth on the left side of the mandible using the Demirjian method, resulting in MAE values of 0.481 for males and 0.621 for females, with RMSE values of 0.617 and 0.795, respectively. In 2023, Roh et al. [9]’s study assessed the developmental stages of seven permanent teeth on the left side of the mandible using the Demirjian method, and employed neural network techniques to derive a regression equation for age estimation. This resulted in MAE values of 0.589 for males and 0.529 for females, with RMSE values of 0.783 and 0.760, respectively. These studies derived age estimation models separately for males and females, and did not include validation using a separate test set, making direct comparisons challenging. Nonetheless, the results of this study, with an MAE of 0.501 and an RMSE of 0.742, are considered to show similar errors compared to these studies [9,20,21].

This study has several limitations. First, the study was conducted with a relatively small dataset. A sufficient amount of data is necessary to derive generalized results and to prevent overfitting. In machine learning and deep learning, data are typically divided into training, validation, and test sets to mitigate overfitting. In this study, the training error was lower than the test error, which may result from insufficient generalization or overfitting. Due to the relatively small number of images compared to other dental age estimation methods using CNN techniques, it was not possible to divide the data into separate validation and test sets, which limited the rigor of the model verification. It is anticipated that using a larger dataset could improve the model’s accuracy.

Additionally, the absence of separate analyses for males and females is another limitation and an area for future improvement. Generally, it is recommended that age estimation models be developed separately for each country, race, and sex, as variations may exist.

Furthermore, the inclusion and exclusion criteria in this study were not strictly applied; certain cases, such as those with braces, missing teeth, and dental anomalies, were not excluded. This may have contributed to increased error rates. It is believed that a more accurate model could be derived by applying stricter criteria when selecting subjects.

This study developed an age estimation algorithm using a backbone consisting of an ensemble of ResNet50 and VGG19, as it is suggested that combining multiple models can achieve higher accuracy. However, while it is important to compare various backbone architectures when developing CNN image classification models using transfer learning, this study employed only a single ensemble model. In the future, it would be beneficial to compare various CNN architectures to identify a model with higher accuracy.

For an accurate estimation model, it is crucial to establish an ROI that effectively captures age-related changes. In this study, the mandibular dentition region was selected as the ROI due to its relatively clear imaging. Manually setting ROIs for all images is a time-consuming process. To address this, deep learning-based object detection techniques were employed. Specifically, the widely used YOLO technique was applied for object detection [17]. In OPG images, the mandible region exhibits a consistent shape and appears in relatively predictable positions, allowing us to develop a model capable of accurate detection with relatively little data. After applying the object detection model, the researchers visually verified the ROIs for any inaccuracies before using them for training. For more accurate results, it may be necessary to define ROIs for the entire OPG, the upper and lower dentition, as well as the upper jaw and lower jaw separately, and compare their outcomes.

In this study, we developed an automated age estimation model using the mandibular dentition region in OPGs of Korean children, applying CNN ensemble techniques. The model achieved accuracy comparable to previous studies where experts assessed the developmental stages of teeth. It is anticipated that more accurate age estimation models can be developed in the future by utilizing a larger sample size, selecting more refined samples, and applying diverse deep learning image analysis methods.

No potential conflict of interest relevant to this article was reported.

Due to privacy protection concerns, the data underlying this study are not publicly accessible. Data sharing requests must comply with institutional guidelines and may require ethical approval.

This study was supported by National Forensic Service (NFS2024CLI12), Ministry of the Interior and Safety, Republic of Korea.

Conceptualization: JMA. Data curation: YSS, KRK, ISS, YHO, JHL, CUC. Formal analysis: BYR, HJP. Funding acquisition: BYR. Methodology: HJP. Project administration: JMA. Resources: YSS, KRK, ISS, YHO, JHL, CUC. Writing - original draft: BYR, HJP. Writing - review & editing: JWR.