Kim Inseok, Yang Sujin, Choi Yiseul, Kwon Hyeokhyeon, Lee Changmin, Park Wonse
Department of Advanced General Dentistry, Yonsei University College of Dentistry, Seoul, Republic of Korea.
Department of Advanced General Dentistry, Yonsei University College of Dentistry, Seoul, Republic of Korea; Institute for Innovation in Digital Healthcare, Yonsei University, Seoul, Republic of Korea.
Forensic Sci Int. 2025 Sep;374:112531. doi: 10.1016/j.forsciint.2025.112531. Epub 2025 Jun 17.
INTRODUCTION/OBJECTIVES: Estimating sex and chronological age is crucial in forensic dentistry and forensic identification. Traditional manual methods for sex and age estimation are labor-intensive, time-consuming, and prone to errors. This study aimed to develop an automatic and robust method for estimating sex and chronological age from orthopantomograms using a multi-task deep learning network.
A deep learning model was developed using a multi-task learning approach with a backbone network and separate attention branches for sex and age estimation. The dataset comprised 2067 orthopantomograms, evenly distributed across sex and age groups ranging from 3 to 89 years. The model was trained using the VGG backbone, optimizing for both sex classification and age regression tasks. Performance was evaluated using mean absolute error (MAE), coefficient of determination (R²), and classification accuracy.
The developed model demonstrated outstanding performance in chronological age estimation, achieving a mean absolute error (MAE) of 3.43 years and a coefficient of determination (R²) of 0.941. For sex estimation, the model achieved an accuracy of 90.2 %, significantly outperforming human observers, whose accuracy ranged from 46.3 % to 63 % for sex prediction and from 16.4 % to 91.3 % for age estimation.
The proposed multi-task deep learning model provides a highly accurate and automated method for estimating sex and chronological age from orthopantomograms. Compared to human predictions, the model exhibited superior accuracy and consistency, highlighting its potential for forensic applications.
引言/目的：在法医牙科学和法医鉴定中，估计性别和实足年龄至关重要。传统的手动估计性别和年龄的方法劳动强度大、耗时且容易出错。本研究旨在开发一种使用多任务深度学习网络从全景曲面断层片中自动且可靠地估计性别和实足年龄的方法。
使用多任务学习方法开发了一个深度学习模型，该模型具有主干网络以及用于性别和年龄估计的单独注意力分支。数据集包含2067张全景曲面断层片，均匀分布在3至89岁的性别和年龄组中。该模型使用VGG主干进行训练，针对性别分类和年龄回归任务进行优化。使用平均绝对误差（MAE）、决定系数（R²）和分类准确率来评估性能。
所开发的模型在实足年龄估计方面表现出色，平均绝对误差（MAE）为3.43岁，决定系数（R²）为0.941。对于性别估计，该模型的准确率达到90.2%，显著优于人类观察者，人类观察者在性别预测方面的准确率为46.3%至63%，在年龄估计方面的准确率为16.4%至91.3%。
所提出的多任务深度学习模型提供了一种从全景曲面断层片中高度准确且自动化地估计性别和实足年龄的方法。与人类预测相比，该模型表现出更高的准确性和一致性，突出了其在法医应用中的潜力。
