Generative Vision-Language Models Are Reshaping Medical Visual Question Answering, Review Finds

A systematic review of 27 studies shows Med-VQA has shifted from text-only classifiers to generative vision-language models. Frameworks using chain-of-thought reasoning and retrieval-augmented generation improve interpretability, but lack standardized clinical benchmarks and rea…

• Systematic review following PRISMA guidelines, covering 27 representative Med-VQA studies from 2023–2026.
• Identifying the shift toward generative models using chain-of-thought, multi-agent frameworks, and retrieval-augmented generation (RAG) for free-form answering.
• Major limitations include a lack of standardized evaluation, limited multi-view and multi-lingual capability, higher computational cost, and no real-world clinical validation.

RadPigeon summaries are original and for information only. They are not clinical advice.