Lightweight Joint Optimization of General-Purpose Vision-Language Models and Retrievers for Medical Diagnosis

Clinical decision-making often involves interpreting images (e.g., radiology) for making diagnoses. Retrieving relevant visual information from medical literature and hospital records could enhance diagnostic accuracy. In this paper, we develop a model in which a multimodal retriever is jointly optimized with an LVLM for medical diagnosis, unlike standard RAG where LVLM error signal is not propagated down to the retriever. We show that using only general-purpose backbones, with only lightweight fine-tuning, our model is able to achieve competitive results with medically-pretrained models across clinical multi-label classification and visual question answering tasks. In a novel analysis, we additionally find that in many cases different top retrieved images each lead to different predictions for a given target, and that these cases are empirically challenging for all models, even for non-retrieval models. Our joint retrieval optimization significantly improves these challenging cases over standard RAG. However, oracle analysis reveals that while the correct diagnosis is frequently achievable using one of the top retrieved images, in practice there is a large performance gap from the oracle, and rerankers using frontier LVLMs do not close this gap -- leaving ample room for improvement by future methods. Code will be made publicly available.