Background

Age estimation is extremely important in radiographical, clinical and forensic practice. Accurate age estimation is essential for multiple purposes, as it can be applied to determine the precise time and treatment strategy based on clinical findings [1, 2] and it can serve as important forensic evidence. In children and adolescents, despite several limitations, the development of dentition is one of the most stable and important markers for age estimation [3, 4]. Compared to other skeletal age evaluations, tooth growth and development are less affected by environmental circumstances [5, 6]. This may be related to the precise genetic control of tooth development and eruption [7].

There are many methods for estimating age based on tooth development, eruption, and mineralization stages [8,9,10]. However, theses usually provide slightly less accurate estimations. Many researchers have created modified methods to improve the accuracy of age estimations, adjusting the numbers for particular races and populations or constructing more complex methods of analysis [11, 12]. Even if there have been various improvements, learning the complicated methods that differ depending on the observer and require the intensive efforts of professionals for estimation analysis can still be challenging. However, with the recent advancements in deep learning technology, such as neural networks, multiple layers of interconnected nodes can process vast amounts of data. These networks adjust the weights and biases of the nodes to minimize the error between the predicted output and the actual output [13,14,15,16].

However, most previous machine learning studies have been based on the simple application of existing age estimation methods that are limited to using specific teeth or parts of dental panoramic images for analysis. This study evaluated the application of entire panoramic image data in the deep learning for the age estimation. The outcome performance of age estimation of two supervised learning models, WideResNet,(WRN) and DenseNet (DS) was analyzed.

Materials and methods

Ethical approval

This study was conducted in accordance with the guidelines of the World Medical Association Helsinki Declaration for biomedical research involving human subjects. This study was approved by the Institutional Review Board (IRB) and Clinical Data Warehouse (CDW) data review board of The Catholic University of Korea, Catholic Medical Center (XC21WADI0064). Needs for informed consent were waived by the IRB. Data were collected and administered by CDW and the images were exported under the supervision of Enterprise Data Platform (EDP) of The Catholic University of Korea Information Convergence Institute.

Data collection and classification

After IRB and Data review board’s approval, the CDW system searched for a list of subjects who visited Eunpyeong St. Mary’s Hospital, St. Vincent Hospital, or Seoul St. Mary’s Hospital of the College of Medicine of The Catholic University of Korea from 2016 to 2020 and underwent panoramic imaging obtained using a ProMax (Planmeca, Helsingki, Finland) or Kodak 8000 Digital Panoramic System (Carestream Health Inc., NY, USA) according to the user manual. The patient data list was undergone to an automatic de-identification process by the CDW system. The panoramic images of listed patients were provided by EDP system after the information had been de-identified and the privacy was ensured. From the collected list, a total of 121,469 qualified panoramic images were downloaded by the EDP system in JPEG format. (Fig. 1) The panorama radiographs with low resolution or pathologic lesion such as cyst and tumors were excluded. Of these radiographs, 27,877 images were randomly selected and labeled from 5 to 90 years of age and gender by two experienced dentists. Each image was resized to 256 × 256 pixels. Since the numbers of instances among classes were unbalanced, a re-sampling technique was utilized to uniformly match the amount of data (Tables 1 and 2).

Fig. 1
figure 1

Image data acquisition process through CDW & EDP system

Full size image
Table 1 Type 1 classification: Numbers of data classified by each age
Full size table
Table 2 Type 2 classification Number of images by age and gender in heuristics grouping where the age over 20 years were classified by every 5 years
Full size table

Modeling and learning

Total of 27,877 dental panorama images labeled from 5 to 90 years of age were classified by 2 types of grouping. In type 1, they were classified by each age and in type 2, using heuristic grouping, the age over 20 years was classified by every 5 years. In addition, the application of ± 3 years of deviation in both types was also analyzed. Dataset was split into three disjoint sets, including a training set, a validation set and a test set consisting of 13,220, 1,653 and 1,653 images, respectively. (Tables 1 and 2)

DN and WRN models were applied for supervised learning. Stochastic gradient descent was used as an optimizer with a learning rate of 0.005, a mini-batch size of 8, a resize of 256 and a momentum of 0.9.

Performance analysis

The accuracy, sensitivity, precision, and f1 scores were calculated to evaluate the performance of each model. Python programming language (v. 3.7.11), Pytorch (v.1.8.2) and a graphics card (Nvidia Quadro 6000 8GB *2) were used for analysis.

Results

Tables 3 and 4 show the model performances of DN and WRN. After a total of 13,220 classified panorama images were trained, 1,653 images were used for validation in each model. The same number of images used for validation was utilized for the test. The best performance was obtained using 40 epochs.

Table 3 Performance of DenseNet model
Full size table
Table 4 Performance of WideResNet model
Full size table

In DN model, the accuracy and F1 score for type 1 grouping were 0.1016 and 0.058, respectively, with a ± 3years of deviation, 0.2813 and 0.1768. For the type 2 grouping, the accuracy and F1 score were 0.3146 and 0.2027, respectively, with a ± 3years of deviation, 0.7641 and 0.6583. The precision and recall score of type 1 grouping were 0.0579 and 0.0583, respectively, with a ± 3years of deviation, 0.1776 and 0.1768. For the type 2 grouping, precision and recall score were 0.2115 and 0.2117, respectively, with a ± 3years of deviation, 0.6632 and 0.6658 respectively.

In WRN model, the accuracy and F1 score of type 1 grouping were 0.1041 and 0.0599, respectively, with a ± 3years of deviation, 0.2716 and 0.1709. For the type 2 grouping, the accuracy and F1 score were 0.3182 and 0.2071, respectively, with a ± 3years of deviation. 0.7323 and 0.6437 respectively. The precision and recall score of type 1 grouping were 0.0598 and 0.0608, respectively, with a ± 3years of deviation, 0.1707 and 0.1718. For the type 2 grouping, precision and recall score were 0.2098 and 0.2147, respectively, with a ± 3years of deviation, 0.7623 and 0.6476 respectively.

Figures 2 and 3 show the results of both DN and WRN models as a confusion matrix. Considering that a higher the diagonal value of the confusion matrix indicates a more accurate predictive model, the figure present a significant accurate diagnosis in type 2 grouping with a ± 3years of deviation in both DN and WRN models.

Figs. 2
figure 2

a and b. Confusion matrix of the results by DenseNet. 2a results before heuristic grouping (type1gourping). 2b, results after heuristic grouping (type 2 grouping)

Full size image
Figs. 3
figure 3

a and b. Confusion matrix of the results by WideResNet. 3a results before heuristic grouping (type1gourping). 3b, results after heuristic grouping (type 2 grouping)

Full size image

Discussion

Over the years age estimation through imaging has been a well-established method within the field of forensic dentistry, garnering widespread recognition for its inherent utility. Panchbhai discussed various radiological methods used for human age identification. The literature survey identified 46 relevant articles that highlighted the significance of radiography in assessing the extent of dental tissue calcification, crown and root formation, eruption stages, and their correlation with age [17]. Radiographic and tomographic techniques are cost-effective and important tools in forensic dentistry for human identification, especially when combined with information technology resources. Imaging, clinical, and forensic dentists should consider the available methods and legal requirements to ensure accurate age estimation.

Most available age estimation methods are statistical methods that require effort and time during preprocessing measurement. For example, age can be predicted using a regression formula with tooth-coronal index (TCI) [18,19,20]. In comparison, the present study estimated age based on the overall appearance of a panoramic image rather than the tooth shape, such as measuring the TCI of a specific tooth. The method used in this study differed from previous papers. Simply classifying the images by age reduced the effort of preprocessing step that traditionally required labelling of specific structure of tooth by professionals. And the application of deep learning allowed the process of the data from full panorama images for the analysis of the age estimation not limiting in only from specific teeth data. However, Due to their complexity, AI systems have been often regarded as black boxes, which do not provide any feedback on why and how they arrive at their predictions. In future, efficient application of “explainable AI” is expected to visualize, interpret, and explain the logic behind AI solutions and provide clear prediction strategies [21].

Several other methods for age estimation have been devised. In a machine learning study using Cone beam computed tomography (CBCT) images, the buccal alveolar bone levels of 150 images were utilized by dividing ages of 20–69 years old into 5-year units. In Saric’s CBCT based study, the Random Forest classifier achieved a correlation coefficient of 0.803 and a mean absolute error of 6.022 [22]. However, since the CBCT study used a small number of samples, additional research is needed to determine whether it can be widely applied. In addition, it is more difficult to obtain a CBCT image than a dental panorama for age estimation, and there is a risk of radiation exposure. The present method achieved relatively precise age estimations through heuristic grouping with of supervised classification learning models with 13,220 whole panoramic images.

An AI-based age estimation study using 1,922 panoramic images of patients 15–23 years old was conducted in Malaysia [23]. The study used a hybrid model of convolutional neural networks (CNN) and K nearest neighbors (KNN). Although the method age range was narrow, it successfully estimated age in one-year, six-month, three-months and one-month range with accuracies of 99.98%, 99.96%, 99.87% and 98.78%, respectively. The hybrid (HCNN-KNN) model made good predictions but is based on relatively certain eruption and developmental stages in adolescents and young adults except for those receiving orthodontic treatment, those with dysplasia or those who experienced trauma. The present study was analyzed not only young age patient, but also adult and older patients were included. The machine learning covered the images of the living patient of the age from 5 to 90.

In a CNN study using panoramic photos of 4,035 patients aged 19–85 years in Croatia, age estimation studies were conducted in four groups: 0–15 years old, 16–30 years old, 31–60 years old and over 61 years old with the VGG16 AI learning method [24] through whole orthopantomographic images of archaeological skull. The study demonstrated 73% accuracy. In Korea, a study was conducted on artificial intelligence learning using CNN on 1,586 dental panoramic X-rays [25]. The image of the first molar was exported and the age was estimated by CNN learning. Based on the data from the 10-year-old group, the patients were reclassified into three groups of 0–19 years old, 20–49 years old and 50 years old or older with an estimated accuracy ranging from 89.05 to 90.27%. In both studies, the use of CNN with graphics was attempted rather than simple AI learning and the Korean study also presented the results of heatmap and Grad-CAM. In the present study, grouping was conducted through artificial intelligence learning and the accuracy and f1 score were improved after heuristic grouping. While previous studies have focused on improving accuracy using a wide age range of patients, In the present study, heuristics grouping for over 20 years of age dividing by every 5 years with ± 3 years of deviation for the analysis was applied for provide improve accuracy of age estimation in narrower age range.

It is a known fact that, the external validation using panoramic radiograph datasets from other institutions is necessary to obtain reliable results [26]. However, since each medical imaging data contains private personal information, such data are primarily protected and locked. and not easily accessible and shareable between different institutions due to medical ethical issue [27]. Nevertheless, this study is characterized by the utilization of data from three hospitals of our university located in different districts and with different panorama equipment system. The collection and de-identification of the data were performed using CDW system. And the panorama image files were downloaded and protected by the EDP system of our institution. It would contributed to diminish the overfitting.

The supervised machine learning model used in this study, were WRN and DN. The WRN model is a type of SL using a novel network with decreased depth and increased width of residual networks compared to the previous ResNet model [28]. In addition to the effect of dropout in the residual block, WRN provides better performance and faster training compared to previous deep learning networks, achieving new state-of-the-art and significant improvements compared to ImageNet [28]. While WRN focused on the width of the network, DN focused on the shortcut connections of ResNet [29]. In previous SL involving ResNet, the Highway network, and ResDrop, only the output of the previous layer was sent to the next layer. In comparison, DN receives the output of many previous layers at once and combines the inputs by concatenation rather than addition [29]. Compared with WRN showing the same performance and similar error rates, DN reported an improvement with approximately two times fewer parameters, suggesting deep supervision as the reason for the improved performance [29]. Both SL models exhibited significantly improved results compared to the previous generation, with similar results between them. Based on this performance, both models are being applied in a wide range of medical research fields, with the possibility of more extensive use in the future [30, 31]. Another study compared age estimation on panoramic radiography using the Kvaal method and machine learning. The study found that machine learning techniques, specifically the XG Boosting Reg classifier, showed higher precision in age estimation (MAE: 4.77) compared to the Kvaal method (MAE: 5.68), indicating that ML can enhance age estimation on panoramic radiographs [32]. The reason for the superiority of various machine learning age estimation methods is that the range/quantity of features or patterns that a human can find in a panoramic image is smaller than the features/patterns that a deep neural network can find. It is also difficult to explain the results of age estimation because it is difficult to know which part of the image the deep neural network looked at to identify the features or patterns. However, if advances in this field continue in the future, more convenient and faster age estimation will provide an opportunity to better understand the principles of analysis using deep neural networks.

Artificial intelligence learning could be a useful solution in forensics fields such as age estimation because it can perform complex tasks that were previously difficult to complete in a faster and more accurate manner. In order to achieve this goal, research should continue to utilize and develop various machine learning methods. In the future, it is essential to conduct research on the application and evaluation of various new methods, including semi-supervised learning or SL using artificial intelligence.

Conclusion

This preliminary study attempts to utilize entire dental panoramic image data in a deep learning model for age estimation. Instead of traditionally requiring professionals to label specific tooth structures, simply classifying the images by age reduced the effort of the preprocessing step. The application of deep learning enabled the analysis of age estimation using data from full panoramic images, rather than being limited to specific teeth data. The performances of both DN and WRN models, with heuristics grouping (where ages over 20 years were classified in 5-year intervals) and a deviation of ± 3 years, yielded satisfactory results in accuracy, recall, precision, and F1 scores. These results are comparable to previous studies on age estimation using traditional methods that require intensive professional effort for analysis and utilize partial data from images, such as teeth. Further clinical and transdisciplinary studies in the medical and advanced technological fields are needed to enhance the quality and simplify the process of age estimation through AI. In the future, the application of AI is expected to assist humans in clinical and dentomaxillofacial radiology fields.