Personal identification with orthopantomography using simple convolutional neural networks: a preliminary study.

Area of Science:

• Forensic Odontology
• Computer Science
• Radiology

Background:

• Personal identification (PI) in forensic dentistry traditionally relies on visual comparison of dental records and radiographs.
• There is a lack of globally accepted digital methods for PI in forensic odontology.
• Effective image recognition models, such as CNNs, are underutilized in this field.

Purpose of the Study:

• To evaluate the efficacy of convolutional neural network (CNN) technologies for personal identification (PI) using paired orthopantomographs.
• To assess the accuracy of various CNN architectures in dental PI.
• To compare pretraining and fine-tuning transfer learning methods for CNN-based dental PI.

Main Methods:

• Thirty pairs of orthopantomographs, taken on different days, were analyzed.
• Six established CNN architectures (VGG16, ResNet50, Inception-v3, InceptionResNet-v2, Xception, MobileNet-v2) were employed.
• Both pretraining and fine-tuning transfer learning techniques were validated for accuracy.

Main Results:

• All tested CNN architectures achieved a detection accuracy of 80.0% or higher.
• Fine-tuning transfer learning yielded higher validation accuracy compared to pretraining.
• The VGG16 model demonstrated the highest accuracy, reaching 100.0% with both pretraining and fine-tuning.

Conclusions:

• CNN technology is useful for personal identification (PI) in forensic odontology, even with limited orthopantomographic data.
• The VGG16 architecture proved to be the most effective among the six CNN models tested for this application.