An Agentic Multimodal Framework for Reliable Medical Reasoning under Privacy Constraints

As artificial intelligence evolves, medical question-answering systems are changing to generative artificial intelligence (AIGC). Nevertheless, the general large-scale language models cannot find clinical applications yet because of the lack of multimodal perception, hallucinations, no clear reasoning chains, and the rigid data privacy requirements, which limit the deployment to clouds. This paper proposes a Multimodal Medical Agent (MMA) framework to address these challenges. It is based on a visual encoder that is connected to the LLaVA-Med backbone and introduces Retrieval-Augmented Generation (Self-RAG) and ReAct mechanisms. Based on the 4-bit quantization technology, this research is able to realize a low-resource, offline implementation of the end-to-end system. Experimental results demonstrate that the system scores an accuracy of 74.0 per cent on the Slake multimodal medical question-answering dataset, representing a notable increase of 29 percentage points over the baseline model. Such findings confirm that the suggested framework could be successfully used to improve the multimodal medical reasoning performance and ensure the security of data, which proves that this technology could be applied to the computer-aided diagnosis context in the real-world setting.