Application of entire dental panorama image data in artificial intelligence model for age estimation

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) [18‐20]. 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.