Automated Rib Fracture Detection on Chest X-Ray Using Contrastive Learning

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Rib fractures are common injuries requiring accurate detection.
• Current diagnostic methods for rib fractures can be challenging.
• Deep learning offers potential for automated fracture detection.

Purpose of the Study:

• To develop and evaluate a deep learning model for detecting rib fractures on chest X-rays.
• To assess the impact of pretraining on model performance.
• To compare detection accuracy across different fracture types.

Main Methods:

• A deep learning model was developed using a contrastive learning framework (simCLR) and a fully convolutional one-stage (FCOS) detection network.
• The model was trained and validated on a large dataset of 18,631 chest digital radiography (DR) images.
• Performance was evaluated on multicenter test sets (Data-CZ and Data-CH) comparing different pretraining strategies.

Main Results:

• Pretraining with DR images significantly improved sensitivity compared to no pretraining or ImageNet pretraining (e.g., Data-CZ: 0.822 vs. 0.465/0.735).
• The model achieved higher sensitivity for displaced fractures (0.873 in Data-CZ) compared to other types.
• False positive rates remained consistent at approximately five per scan across methods.

Conclusions:

• Pretrained deep learning models substantially enhance the accuracy of rib fracture detection from X-ray images.
• This technology has the potential to reduce missed diagnoses and improve overall diagnostic efficacy in radiology.
• Further validation across diverse datasets is warranted to confirm generalizability.