An international team of radiology and AI experts examined the transformative potential of generative AI and foundation models in medical imaging, addressing both promising applications and critical implementation challenges. Published in Radiology, the comprehensive review explores how large-scale AI models trained on vast unlabeled datasets demonstrate unprecedented versatility across multiple radiology tasks, from automated report generation to synthetic data creation and multimodal image interpretation. The authors analyzed various foundation model training paradigms including autoregressive techniques, autoencoders, masked autoencoders, and contrastive pretraining methods that enable these models to understand spatial context and linguistic patterns in radiologic data.
The paper outlines both opportunities and cautions across technical, clinical, and regulatory domains. Foundation models capable of generating draft reports directly from medical images mark a major advancement with potential to dramatically enhance efficiency and scale the net capacity of the current radiology workforce, addressing the global radiologist shortage. However, critical challenges include preventing inaccuracies and hallucinations in AI-generated content, ensuring models incorporate essential clinical domain expertise, and understanding hidden biases that models may synthesize beyond visible lesions. The authors emphasize that successful implementation requires radiologists on development teams, robust validation frameworks for safety, transparent explanation of AI reasoning, and regulatory oversight particularly for vision-language models entering clinical practice.