Study: Large Language Models Outperform on Chest CT Report Analysis
July 30, 2025
Advanced large language models like GPT-4 accurately identify thoracic diseases in chest CT reports, enhancing pre-operative surgical planning.
Key Details
- Five LLMs (GPT-4, Claude-3.5, Qwen-Max, GPT-3.5-Turbo, Gemini-Pro) compared using 13,489 real-world chest CT reports.
- GPT-4 achieved up to 75% accuracy in identifying 13 common chest diseases with multiple-choice prompts.
- Multiple-choice prompts significantly improved model accuracy compared to open-ended questions.
- Fine-tuning GPT-3.5-Turbo increased its accuracy from 42% to 65% in challenging cases.
- No single LLM was best for all diseases, suggesting a tailored approach may be optimal.
- Future research will use explainable AI tools to increase transparency and reliability.
Why It Matters
The study demonstrates that modern LLMs can act as accurate 'second readers' for radiology reports, possibly reducing diagnostic errors and alleviating radiologist workload. Fine-tuning and prompt design further boost performance, potentially making AI support accessible even in resource-limited settings.