Secure On-Premise Deployment of Open-Weights Large Language Models in Radiology (arxiv.org)

A recent study introduced an isolation-first architecture for the secure on-premise deployment of open-weights large language models (LLMs) in radiology, specifically utilizing the DeepSeek-R1 model. This innovative approach emphasizes strict network segmentation and monitoring to ensure compliance with regulatory requirements while handling sensitive patient data. During a one-week pilot involving 22 residents and radiologists, participants evaluated the clinical utility and stability of the system, finding it user-friendly and capable of effectively supporting text-anchored tasks like report corrections while noting that open-ended tasks led to critical errors such as hallucinations in output. This development is significant for the AI/ML community as it demonstrates a viable framework for integrating LLMs in clinical settings, overcoming hurdles posed by data privacy laws. By enabling the safe processing of unanonymized patient health information (PHI), this architecture not only enhances the technical feasibility of LLM deployments in healthcare but also sets a precedent for similar infrastructure in other sectors requiring heavy data protection, making the deployment package publicly available to further encourage innovation in secure AI applications.