Automated forensic human identification on dental panoramic radiographs using transformer-based detection with...

Area of Science:

• Forensic Science
• Computer Vision
• Biometrics

Background:

• Dental radiographs (DPRs) are crucial for human identification in mass disasters when other methods fail.
• Large antemortem (AM) databases pose challenges for manual comparison with postmortem (PM) data.
• An automated pre-screening tool is needed to manage large-scale identification efforts.

Purpose of the Study:

• To develop and evaluate an automated framework for efficient human identification using dental radiographs.
• To pre-screen large AM databases, generating a ranked shortlist for forensic expert verification.

Main Methods:

• A Detection Transformer (DETR) model with Circle Loss was employed to detect teeth and generate embeddings.
• The framework was evaluated on 1,029 individuals, comparing its performance against CNN-based detectors.
• Key performance metrics included Rank-K accuracy, mean average precision (mAP), and normalized discounted cumulative gain (nDCG).

Main Results:

• The detection-based framework achieved 65.4% Rank-1 and 87.4% Rank-10 accuracy.
• It outperformed CNN-based methods, with a 44.4% improvement in Rank-1 accuracy over image-level approaches.
• Embeddings captured detailed dental anatomy, enhancing identification accuracy.

Conclusions:

• The developed framework serves as a practical tool for pre-screening dental records in disaster victim identification.
• It efficiently narrows down potential matches, supporting forensic experts rather than replacing them.
• This automated approach addresses the scalability bottleneck in comparing large AM and PM dental databases.