This review focuses on foundation models in radiology and describes their key characteristics and methods for training, adapting, and evaluating; in addition, cautions and future directions associated with their responsible deployment in radiologic practices are discussed.

A fine-tuned, open-source large language model (Mistral-7B; Mistral AI) effectively extracted clinical history elements from imaging orders with substantial agreement with radiologists and rivaled a closed-source model (GPT-4 Turbo; OpenAI).