RULE: Reliable Multimodal RAG for Factuality in Medical Vision Language Models

TL;DR

RULE introduces a method to improve factual accuracy in medical vision-language models by calibrating retrieval and fine-tuning to balance reliance on internal knowledge and external data, significantly enhancing performance.

Contribution

The paper presents a novel approach with calibrated retrieval and fine-tuning strategies to address factuality issues in Medical Large Vision Language Models.

Findings

Achieved an average of 47.4% improvement in factual accuracy across tasks.

Demonstrated effectiveness on medical VQA and report generation datasets.

Publicly released benchmark and code for reproducibility.

Abstract

The recent emergence of Medical Large Vision Language Models (Med-LVLMs) has enhanced medical diagnosis. However, current Med-LVLMs frequently encounter factual issues, often generating responses that do not align with established medical facts. Retrieval-Augmented Generation (RAG), which utilizes external knowledge, can improve the factual accuracy of these models but introduces two major challenges. First, limited retrieved contexts might not cover all necessary information, while excessive retrieval can introduce irrelevant and inaccurate references, interfering with the model's generation. Second, in cases where the model originally responds correctly, applying RAG can lead to an over-reliance on retrieved contexts, resulting in incorrect answers. To address these issues, we propose RULE, which consists of two components. First, we introduce a provably effective strategy for controlling factuality risk through the calibrated selection of the number of retrieved contexts. Second, based on samples where over-reliance on retrieved contexts led to errors, we curate a preference dataset to fine-tune the model, balancing its dependence on inherent knowledge and retrieved contexts for generation. We demonstrate the effectiveness of RULE on medical VQA and report generation tasks across three datasets, achieving an average improvement of 47.4% in factual accuracy. We publicly release our benchmark and code in https://github.com/richard-peng-xia/RULE.