Large Language Models in radiology: A technical and clinical perspective
Jean Kao, Huan-Tang Kao
Abstract
Large Language Models (LLMs) are transformer-based deep learning models trained on vast text corpora, enabling advanced natural language understanding and generation. This technical note presents a comprehensive overview of LLMs in the field of radiology, following a style akin to high-impact journals. We discuss applications of LLMs in radiology, including image interpretation, automated report generation, and workflow efficiency improvements. We provide in-depth technical insight into transformer architectures and specific models such as GPT-4, PaLM, and Med-PaLM, including key mathematical formulations and algorithms that underpin LLM functionality. We examine evaluation metrics for radiology LLM tasks (accuracy, language metrics, etc.), along with critical considerations of bias, ethical issues, and regulatory guidelines. Finally, we address practical considerations for radiologists and AI researchers, offering perspectives on integrating LLMs into clinical practice and research. The aim is to combine technical rigor with clear explanations to serve a mixed audience of clinicians and AI scientists. Figures and tables are included as placeholders for clarity, and references are provided in a numbered format for further reading.