Design, Secure Deployment, and Clinical Evaluation of Medical Vision–Language Foundation Models
Date
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
The clinical integration of medical Vision–Language Models (VLMs) is currently impeded by two primary challenges: the critical need for Protected Health Information (PHI) security and the limitation for clinical deployment. We present MedFoundationHub: a containerized architecture tailored for the privacy preserving and clinician-in-the-loop evaluation of large-scale medical foundation models. On the architectural level, MedFoundationHub leverages a containerized, Docker-orchestrated framework to ensure privacy-preserving inference on local, offline workstations. Our system incorporates critical performance optimizations to ensure robustness under clinical environments. These features enable the execution of state-of-the-art foundation models on a single NVIDIA A100 GPU, bridging the gap between high-performance AI and real-world clinical environments. On the evaluation level, we conducted a large-scale study involving board-certified pathologists who executed 1,015 clinician–model scoring events across colon and renal pathology cases. Our findings identify the systemic limitations in pathology terminology and reasoning consistency in VLMs. By providing both a secure deployment pipeline and a rigorous evaluation interface, MedFoundationHub serves as a foundational framework for translating medical VLMs from theoretical research into verifiable clinical decision support tools.