LLMs Demonstrate Potential for Fracture Classification in CT Radiology Reports
July 9, 2025
LLMs like ChatGPT-4o and AmbossGPT can accurately classify bone fractures in CT radiology reports, aiding radiologists.
Key Details
- Study assessed four LLMs (ChatGPT-4o, AmbossGPT, Claude 3.5 Sonnet, Gemini 2.0 Flash) on 292 artificial CT reports representing 310 fractures.
- ChatGPT-4o and AmbossGPT showed highest overall classification accuracy (74.6% and 74.3%).
- Bone recognition rates were high for all models (90%-99%), but fracture subtype classification was lower (71%-77%).
- Statistically significant accuracy differences were noted between LLMs by fracture type and anatomical location.
- Validation with real-world reports (145 fractures) using LLaMA 3.3-70B yielded similar results to artificial datasets (~70% performance).
- Authors note need for further validation on large, multi-center real-world datasets.
Why It Matters
Radiology practices rely heavily on textual reporting, and automating fracture classification could streamline radiological workflows, reduce variability, and improve efficiency. While current LLMs show promise, further validation is necessary before widespread adoption.