Daniel Truhn

Text-only models can rival multimodal counterparts in chest radiography accuracy, questioning the necessity of image input for clinical AI applications.

Specialization in vision-language models dramatically increases the risk of re-linking de-identified medical images to their original reports, revealing critical privacy vulnerabilities in clinical AI applications.

Scaling clinical LLMs doesn't guarantee safety: high-risk errors persist even with advanced RAG and max-context prompting, highlighting the critical role of evidence quality and deployment strategy.

Even with multi-agent adjudication, MLLMs still struggle to reliably differentiate visually confounded diseases in a zero-shot setting, highlighting limitations in current agent-based diagnostic approaches.