A Patch-Based Deep Learning Approach for Detecting Rib Fractures on Frontal Radiographs in Young Children
Adarsh Ghosh1,2,3, Daniella Patton4, Saurav Bose4
1Department of Radiology, Children's Hospital of Philadelphia, Philadelphia, PA, USA. adarsh.ghosh@cchmc.org.
Journal of Digital Imaging
|March 10, 2023
Summary
A new patch-based deep learning algorithm effectively detects rib fractures in children under two years old using chest radiographs. This computer-aided detection shows promise for improving diagnostic accuracy in pediatric radiology.
Area of Science:
- Pediatric Radiology
- Medical Imaging Analysis
- Deep Learning in Healthcare
Background:
- Chest radiography is the primary imaging method for identifying rib fractures in young children.
- Automated detection of rib fractures is challenging due to high spatial resolution requirements in deep learning models.
Purpose of the Study:
- To develop and evaluate a patch-based deep learning algorithm for automatic rib fracture detection on frontal chest radiographs in children under two years old.
Main Methods:
- A dataset of 845 pediatric chest radiographs was used, with manual segmentation of rib fractures by radiologists serving as ground truth.
- A patch-based sliding-window technique was employed to meet high-resolution needs for fracture detection.
- ResNet-50 and ResNet-18 architectures were utilized with transfer learning.
Main Results:
- The ResNet-50 model achieved an AUC-ROC of 0.74 with 88% sensitivity and 43% specificity on a whole-radiograph level.
- The ResNet-18 model achieved an AUC-ROC of 0.75 with 75% sensitivity and 60% specificity on a whole-radiograph level.
- Patch-level analysis showed varying AUC-PR and AUC-ROC values for both models.
Conclusions:
- Patch-based analysis is a viable approach for detecting rib fractures in children under two years old.
- Further research with larger, multi-institutional datasets is needed to enhance generalizability, particularly for cases involving suspected child abuse.
