Dental Age Estimation in Children Using Convolution Neural Network Algorithm: A Pilot Study
10.14476/jomp.2024.49.4.118
- Author:
Byung-Yoon ROH
1
;
Hyun-Jeong PARK
;
Kyung-Ryoul KIM
;
In-Soo SEO
;
Yeon-Ho OH
;
Ju-Heon LEE
;
Chang-Un CHOI
;
Yo-Seob SEO
;
Ji-Won RYU
;
Jong-Mo AHN
Author Information
1. Forensic Medicine Division, National Forensic Service Gwangju Institute, Jangseong, Korea
- Publication Type:Original Article
- From:
Journal of Oral Medicine and Pain
2024;49(4):118-123
- CountryRepublic of Korea
- Language:English
-
Abstract:
Purpose:Recently, deep learning techniques have been introduced for age estimation, with automated methods based on radiographic analysis demonstrating high accuracy. In this study, we applied convolutional neural network (CNN) techniques to the lower dentition area on orthopantomograms (OPGs) of children to develop an automated age estimation model and evaluate its accuracy for use in forensic dentistry.
Methods:In this study, OPGs of 2,856 subjects aged 3-14 years were analyzed. The You Only Look Once (YOLO) V8 object detection technique was applied to extract the mandibular dentition area on OPGs, designating it as the region of interest (ROI). First, 200 radiographs were randomly selected, and were used to train a model for extracting the ROI. The trained model was then applied to the entire dataset. For the CNN image classification task, 80% of OPGs were allocated to the training set, while the remaining 20% were used as the test set. A transfer learning approach was employed using the ResNet50 and VGG19 backbone models, with an ensemble technique combining these models to improve performance. The mean absolute error (MAE) on the test set was used as the validation metric, and the model with the lowest MAE was selected.
Results:In this study, the age estimation model developed using mandibular dentition region from OPGs achieved MAE and root mean squared error (RMSE) values of 0.501 and 0.742, respectively, on the test set, and MAE and RMSE values of 0.273 and 0.354, respectively, on the training set.
Conclusions:The automated age estimation model developed in this study demonstrated accuracy comparable to that of previous research and shows potential for applications in forensic investigations. Increasing the sample size and incorporating diverse deep learning techniques are expected to further enhance the accuracy of future age estimation models.
