Pre-trained VGG16 Model for Forensic Dental Age Estimation

Background A practical utilization of Machine Learning in Forensic Odontology is yet underexplored, especially in the field of Age Estimation. Age Estimation is essential in legal proceedings to protect the rights of individuals without proper documentation, whether for seeking asylum or when caring for a found child. This study aimed to utilize VGG16 model to read, analyze, and provide classification of tooth development stages of the third molars. Specifically, the third molars 38 and 48 were used to classify individuals into age groups based on thresholds of 16, 18, and 21 years old. The goal was to compare the accuracy of the traditional age estimation methods established by Demirjian, Moorrees, Funning, and Hunt, with a CNN-based approach. Method A total sample of 876 orthopantomograms (OPGs) from the Portuguese population was collected from the ULS Hospital Santa Maria, University of Lisbon. The sample comprised 447 OPGs from male patients and 429 from female patients, aged 10 to 25 years old. Age Estimation was calculated manually using the methods established by Demirjian and Moorrees. Furthermore, we trained the VGG16 model to read, analyze, and provide classification of the development stages, and afterwards we evaluated the VGG16 model through overall accuracy, recall, precision, and F-Score. Results The VGG16 model provided excellent results for the cropped images with only third molars (38 and 48) and captured very well the patterns and the features of development stages, so the overall accuracy obtained was more than 90%. However, to analyze and provide classification of the development stages established by Demirjian and Moorrees, Funning, and Hunt the VGG16 model faced some limitations due to the insufficient sample of OPGs. Conclusion The age groups classification based on the development of teeth 38 and 48 demonstrated promising results with a high degree of accuracy. However, the limited sample size constrained the VGG16 model's ability to effectively differentiate between the numerous stages of tooth development. To enhance the model's accuracy and reliability, a larger and more diverse dataset is necessary to better capture the nuances of each developmental stage.