Automatic rib fracture detection on postmortem CT data using deep learning
Manel Lopez-Melia1,2, Virginie Magnin3,4, Sami Schranz3
1University Centre of Legal Medicine Lausanne-Geneva, University of Geneva, Rue Michel-Servet 1, 1206, Geneva, Switzerland. manel.lopezmelia@unige.ch.
International Journal of Legal Medicine
|December 3, 2025
Summary
A deep learning model (nnDetPM) showed notable performance in detecting rib fractures on postmortem CT scans, comparable to radiologists. However, factors like arm position and medical ware significantly impact performance due to domain shift.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Forensic Science
Background:
- Rib fracture detection is crucial but time-consuming in forensic investigations.
- Deep learning (DL) models can enhance radiologist performance in detecting rib fractures on clinical CT scans.
Purpose of the Study:
- Assess the performance of a DL model (nnDetection) for rib fracture detection on postmortem (PM) CT scans.
- Identify domain shift factors between clinical and PM CT imaging.
Main Methods:
- Trained two instances of the nnDetection model: nnDetPM on 50 PMCT scans and nnDetClin on 660 clinical CT scans.
- Evaluated model performance on a PMCT testing set.
Main Results:
- nnDetPM achieved 70.2% sensitivity and 78.1% precision on PMCT scans.
- nnDetClin performed poorly (19.8% sensitivity, 25.5% precision) on PMCT scans, highlighting domain shift.
- Arm position and medical ware were identified as key domain shift factors.
Conclusions:
- The nnDetPM model demonstrated notable performance in PMCT rib fracture detection, comparable to radiologists.
- Further research is needed to explore advanced techniques to overcome domain shift challenges for DL models in this application.
