Deep Learning-Based Volumetric Tooth Segmentation from Postmortem CT Scans for Forensic Applications
Organized by: International Forensic Scientist Awards
Website: forensicscientist.org
15th Edition of Forensic Scientist Awards 27-28 October 2025 | Paris, France
Revolutionizing Forensic Dental Identification with Artificial Intelligence
Forensic dental analysis plays a vital role in human identification, especially in postmortem investigations where other biological identifiers may be unavailable. However, the manual segmentation of teeth from CT scans is time-intensive and requires specialized expertise. Addressing this challenge, UTooth emerges as an innovative deep learning-based solution designed to automate the process with precision, speed, and consistency.
What is UTooth?
UTooth is a novel AI-driven framework developed for automatic dental segmentation from postmortem computed tomography (CT) scans. It harnesses the power of 3D deep learning to accurately detect and segment teeth, focusing initially on canine teeth—known for their forensic importance and distinctive structure.
How It Works
Using data from 52 CT scans sourced from the New Mexico Decedent Image Database (NMDID), the system employs a three-phase pipeline:
-
Radiodensity-Based Volumetric Preprocessing – Enhances bone contrast and prepares the CT volume for accurate segmentation.
-
Heuristic Jaw Isolation – Automatically identifies and extracts the jaw region from the CT volume.
-
3D U-Net Segmentation with Focal Tversky Loss – Applies a specialized deep neural network architecture that refines segmentation accuracy while handling class imbalance.
Key Results
UTooth achieved a mean Dice coefficient of 0.831 ± 0.061, with the best fold reaching 0.897, proving its reliability and high accuracy under challenging postmortem imaging conditions. These results underscore its potential as a foundational technology in forensic AI and automated dental analysis.
Why It Matters
The integration of AI and forensic imaging through UTooth represents a major step toward fully automated biological profiling and victim identification. This advancement can accelerate investigations, reduce human error, and provide valuable support in disaster victim identification and forensic anthropology.
Future Directions
The current model serves as a cornerstone for expanding into full dental segmentation, enabling automated feature extraction and digital comparison for forensic identification databases. As the technology evolves, UTooth may become an essential component in next-generation forensic workflows.
🔗 Learn more and apply at:
https://forensicscientist.org/
Nominations Open Now: Click here
–––––––––––––––––––––––––––––––––––––
Get Connected Here:
Comments
Post a Comment