Proprietary and open-source large language models (LLMs) generated radiologically relevant indications from clinical notes that were more comprehensive and factual than clinician indications, and when generated by the proprietary LLM, were ranked most useful in protocoling and imaging interpretation by radiologists.