Deep Rib Fracture Instance Segmentation and Classification From CT on the RibFrac Challenge.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Rib fractures are common, severe injuries difficult to detect in CT scans.
• Lack of large datasets and benchmarks hinders deep learning development for rib fracture analysis.
• The RibFrac Challenge was created to address these limitations.

Purpose of the Study:

• Introduce the RibFrac Challenge dataset and evaluation framework.
• Benchmark deep learning algorithms for rib fracture detection and classification.
• Analyze the performance of AI models against human experts.

Main Methods:

• Developed a benchmark dataset with over 5,000 annotated rib fractures from 660 CT scans.
• Established two challenge tracks: detection (instance segmentation) and classification.
• Utilized voxel-level instance masks and diagnosis labels for four fracture types.

Main Results:

• Several top detection models achieved performance comparable to or exceeding human experts.
• AI-based classification of rib fractures currently lacks clinical applicability.
• Post-challenge analysis explored advancements like large-scale pretraining and rib segmentation.

Conclusions:

• The RibFrac Challenge provides a valuable resource for AI in medical imaging.
• AI demonstrates strong potential for rib fracture detection, with classification as a future research direction.
• Findings support further development of AI-assisted diagnosis for rib fractures.