This paper reviews the application of LLMs in radiology for structured reporting, highlighting their potential to enhance efficiency and accuracy in diverse radiological contexts while addressing multilingual challenges and limitations in clinical deployment.
1️⃣ Ten studies, focusing primarily on GPT-3.5 and GPT-4, explored SR automation in imaging modalities such as whole-body CT, CEUS, thyroid ultrasound, and chest X-rays.
2️⃣ GPT-4 demonstrated 100% accuracy in matching MRI/CT reports to structured templates, transforming free-text reports into JSON formats without errors.
3️⃣ Multilingual capabilities were validated in Italian, Chinese, German, and Japanese, showcasing LLMs’ adaptability across linguistic and cultural settings.
4️⃣ Commercial tools leverage LLMs for voice-guided and structured reporting, but lack peer-reviewed validation.
5️⃣ Limitations include hallucinations in 27% of cases, misinterpretations in 19%, and performance variability across tasks like management recommendations.
6️⃣ GPT-4 outperformed GPT-3.5 in critical areas like nodular categorization and management recommendations, emphasizing its role in nuanced SR tasks.
7️⃣ LLMs have been proposed for integrating external knowledge bases, prompt engineering, and human-in-the-loop workflows to address biases and errors.
8️⃣ Safety and transparency remain critical, with proprietary models like GPT-4 not meeting EU AI Act regulatory requirements due to opaque training data and algorithms.
9️⃣ Adoption barriers include the lack of clinical software integration and uncertainties regarding acceptance of LLM-generated reports among clinicians.
🔟 The review underscores a need for comparative studies and integration trials to establish LLMs' real-world feasibility in radiology.
✍🏻 Felix Busch, Lena Hoffmann, Daniel Pinto dos Santos, Marcus R. Makowski, Luca Saba, Philipp Prucker, Martin Hadamitzky, Nassir Navab, Jakob Nikolas Kather, Daniel Truhn, Renato Cuocolo, Lisa Adams, Keno Bressem. Large language models for structured reporting in radiology: past, present, and future. European Radiology. 2024. DOI: 10.1007/s00330-024-11107-6
