Application of large language models in radiological diagnostics: a scoping review
摘要
BACKGROUND: Modern large language models show potential for application in radiological diagnostics across a wide range of routine tasks. AIM: The work aimed to conduct a scoping review of the application of large language models in radiological diagnostics by analyzing possible use-case scenarios and assessing the methodological quality of relevant studies. METHODS: Two search strategies were employed: a primary search (PubMed and eLibrary) targeting full-text publications with well-developed methodology, and a supplementary search (PubMed) aimed at broader coverage of large language model use cases in radiological diagnostics during 2023–2025. Extracted data included bibliometric characteristics, study objectives, use-case scenarios of large language models, nosological profiles, key methodological parameters, and both quantitative and qualitative indicators of diagnostic performance—for both the models and the specialists involved, including their number and experience. The quality was assessed using the modified QUADAS-CAD questionnaire. RESULTS: The primary search yielded 9 studies for analysis; the supplementary search yielded 216. A total of 9 major use-case scenarios for large language models in radiology were identified. The most common among them was the rephrasing of radiology reports in order to improve their accessibility for patient understanding. Models predominantly used were GPT-4 and BERT, along with GPT-3.5, Llama 2, Med42, GPT-4V, and Gemini Pro. The large language model GPT-4 demonstrated high diagnostic accuracy in identifying brain tumors (73.0%), myocarditis (83.0%), and in making decisions on invasive procedures for acute coronary syndrome (86.0%). In turn, it demonstrated low diagnostic accuracy for nervous system disorders of various etiologies (50.0%) and for musculoskeletal diseases (43.0%). The BERT model exhibited high diagnostic accuracy in detecting pulmonary nodules (99.0%) and signs of intracranial hemorrhage (sensitivity and speci