Federal Grant Award Summary

The National Institute of Biomedical Imaging and Bioengineering (NIBIB) awarded the University of North Carolina at Chapel Hill a $414,079 Project Grant (CFDA 93.286: Discovery and Applied Research for Technological Innovations to Improve Human Health) effective June 1, 2026 through May 31, 2028. The award supports development of factual medical large vision-language models (Med-LVLMs) designed to reduce hallucinations in artificial intelligence systems used for radiology image analysis and clinical diagnosis. The project addresses the critical gap between hallucination reduction research on natural images and the unique requirements of medical imaging, where accurate identification of abnormal findings is essential for diagnostic reliability.

The project will deliver two primary products: (1) a modality-aligned Med-LVLM incorporating medical-aware preference learning and knowledge-augmented retrieval mechanisms, accompanied by a novel medical hallucination benchmark for systematic evaluation; and (2) a multi-agent reinforcement learning (MARL) framework enabling specialized Med-LVLM agents to collaborate through structured debate and verification protocols for complex radiological tasks. The research team commits to open-source release of developed models, source code, and evaluation benchmarks to facilitate broader adoption of reliable medical vision-language models across the research and clinical communities.