Radiological imaging is globally prevalent, but radiology reports' free text is rarely used secondarily. Natural Language Processing (NLP) offers structured data retrieval from these reports. This scoping review, following PRISMA-SCR, summarizes research on Large Language Model (LLM)-based information extraction (IE) from radiology reports. Of 34 studies meeting inclusion criteria (post-August 1st, 2023 search), only pre-transformer and encoder-based models are described. External validation shows decreased performance, though LLMs may improve IE generalizability. CT and MRI reports, especially thoracic, are prevalent. Common challenges include missing external validation and method augmentation. Varying reporting granularities affect comparability and transparency.

Daniel Reichenpfader, Henning Müller, Kerstin Denecke