A multi-task learning-based deep learning model for precise estimation of rib fracture age on chest CT.
Authors
Affiliations (5)
Affiliations (5)
- Academy of Forensic Science, Shanghai Key Laboratory of Forensic Medicine (21DZ2270800), Shanghai Forensic Service Platform, Key Laboratory of Forensic Science, Ministry of Justice, 1347 GuangFu West Road, Shanghai, 200063, China.
- School of Basic Medical Sciences, Jiamusi University, Jiamusi, 154007, Heilongjiang, China.
- Department of Radiology, Shanghai Public Health Clinical Center, Fudan University, Shanghai, 201508, China.
- Shanghai Shuzhiwei Information Technology Co., LTD, 333 WenHai Road, Shanghai, 200444, China.
- Academy of Forensic Science, Shanghai Key Laboratory of Forensic Medicine (21DZ2270800), Shanghai Forensic Service Platform, Key Laboratory of Forensic Science, Ministry of Justice, 1347 GuangFu West Road, Shanghai, 200063, China. [email protected].
Abstract
Rib fractures are a common type of chest injury, and the estimation of the fracture age mainly relies on clinical or forensic imaging experts making rough judgments based on CT scans, which is highly subjective. In recent years, artificial intelligence (AI) models have performed exceptionally well in rib fracture detection tasks, providing a potential technical foundation for inferring the time of fracture formation. Therefore, it is necessary to develop deep learning model tools to assist human judgment. In this study, a multi-task deep learning model (based on the 3D-ResNet18) was developed to predict the rib fracture age, classify their healing stages, and identify fracture types concurrently. The model was trained on a multicenter dataset comprising 1,848 rib fractures derived from chest CT scans and demonstrated generalizability in external testing. Specifically, the model achieved a mean absolute error (MAE) of 7.94 days and an accuracy (ACC) of 71.18% in the fracture age prediction and healing stage classification tasks, respectively, outperforming manual evaluation (MAE: 10.65 days). Additionally, the model achieved an ACC of 94.71% in the fracture type classification task.