Building Factual Medical Large Vision-Language Models for Radiology

About This Grant

Project Summary This proposal aims to tackle the critical issue of hallucinations in medical large vision language models (Med- LVLMs) in radiology to significantly enhance the reliability of using these models to analyze medical images and aid in clinical diagnosis. While generative AI has revolutionized medical imaging, concerns about hallucinations generating factually incorrect medical analyses persist. Most research on reducing hallucinations in LVLMs has concentrated on natural images, leaving the unique challenges of medical images largely unexplored. Two primary challenges impede progress: (1) the intrinsic disparity between natural and medical images, where medical analyses must prioritize abnormal findings rather than treat all visual information equally; and (2) the inadequacy of single-model approaches for accurately handling complex and nuanced radiological tasks. To overcome these challenges, the project proposes two specific aims: (1) developing a modality-aligned Med- LVLM that leverages medical-aware preference learning and knowledge-augmented retrieval. This approach enhances alignment between clinical descriptions and medical image features, prioritizes abnormal findings, and systematically evaluates hallucinations through a novel medical hallucination benchmark; and (2) implementing a multi-agent reinforcement learning (MARL) framework to collaboratively mitigate hallucinations in complex radiology scenarios. This framework employs specialized Med-LVLM agents engaging in structured debate and verification, thereby enhancing reliability, interpretability, and efficiency in complex diagnostic tasks. The project team, with extensive expertise in LVLMs, medical imaging analysis, deep learning, and clinical applications, is committed to openly releasing developed models, source code, and evaluation benchmarks, fostering broader adoption and advancement of reliable Med-LVLMs.