Related Experiment Videos
Development of a deep learning-based model for personal identification using postmortem CT and antemortem chest
Ryusei Inamori1, Tomoya Kobayashi2, Takaya Kawasumi3
1Department of Diagnostic Imaging, Tohoku University Graduate School of Medicine, 2-1 Seiryo-machi, Aoba-ku, Sendai, Miyagi 980-8575, Japan.
Forensic Science International
|August 2, 2026
Summary
A deep learning method using postmortem CT-derived RaySum images effectively matches victims' chest X-rays for identification. This AI approach aids forensic examinations by improving accuracy in mass disaster victim identification.
Area of Science:
- Forensic radiology
- Medical imaging analysis
- Artificial intelligence in medicine
Background:
- Accurate personal identification of mass disaster victims is crucial.
- Traditional methods can be time-consuming and challenging.
- Postmortem computed tomography (PMCT) provides detailed anatomical data.
Purpose of the Study:
- To develop and evaluate a deep learning model for matching PMCT-derived RaySum images with antemortem chest X-rays (CXRs).
- To facilitate personal identification of mass disaster victims.
Main Methods:
- Retrospective study of 1385 deceased individuals' PMCT and antemortem CXR data.
- Generation of RaySum images from PMCT trunk segments.
- Application of a pretrained EfficientNet-B3 model with AdaCos metric learning.
- Matching RaySum images against CXR galleries categorized by examination date.
Main Results:
- The "all examinations" category achieved high identification rates (83.0% top-1, 95.2% top-20).
- "Nearest-date" examinations significantly outperformed "oldest-date" examinations (78.9% vs. 54.2% top-1).
- The model demonstrated high retrieval performance in one-to-many CXR galleries.
Conclusions:
- PMCT-derived RaySum images offer high retrieval performance for CXR matching.
- This deep learning approach can narrow down identification candidates in forensic investigations.
- The method assists in prioritizing potential matches for detailed forensic examination.