Objective: Medical Visual Question Answering (VQA) is a quintessential application scenario of biomedical Multimodal Large Language Models (MLIMs). Previous studies mainly focused on input image-question pairs, neglecting the rich medical knowledge of the relevant captions of the pretrained datasets. This limits the model's reasoning capability and causes overfitting. This paper aims to effectively utilize the captions of pretrained datasets to solve the above issues.

Methods: This paper proposes a Caption-Augmented Reasoning Model (CARM), which introduces three innovative components to leverage the captions during finetuning: (1) A Cross-Modal Visual Augmentation (CMVA) module that enriched image feature representations through semantic alignment with retrieved captions; (2) A Retrieval Cross-Modal Attention (RCMA) mechanism that established explicit connections between visual features and domain-specific medical knowledge; (3) A Hierarchical Rank Low-Rank Adaptation (HR-LoRA) module that optimized parameter-efficient finetuning through rank-adaptive decomposition in both unimodal encoders and multimodal fusion layers.

Results: The proposed CARM achieved state-of-the-art performance across three benchmark datasets, with accuracy scores of 0.798 on VQA-RAD, 0.867 on VQA-SLAKE, and 0.718 on VQA-Med-2019, respectively, outperforming existing medical VQA models. Qualitative evaluations revealed that our caption-based augmentation effectively directed model attention to the image regions related to a question.

Conclusions: The proposed CARM effectively improves visual grounding and reasoning accuracy with the systematic integration of medical captions, and the HR-LoRA alleviates overfitting and improves training efficiency.