The Impact of Image Resolution on Biomedical Multimodal Large Language Models

2025-10-21

AI Summary

Biomedical multimodal large language models (MLLMs) suffer from critical detail loss due to reliance on low-resolution general-domain data for pretraining. We systematically investigate the impact of image resolution on model performance and identify resolution mismatch between training and inference as a key factor degrading biomedical understanding. Method: We propose a native-resolution end-to-end training and inference paradigm, coupled with a hybrid-resolution training strategy: high-resolution samples are preserved at full fidelity while downsampled variants are introduced to balance information completeness and computational efficiency. Contribution/Results: Our native-resolution approach achieves an average 12.7% performance gain across diverse biomedical understanding tasks. The hybrid strategy attains 92% of native-resolution performance with only a 5% increase in training cost. These results provide a reproducible methodological foundation for high-fidelity biomedical MLLM development and clinical deployment.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Computer Vision: Multi-modal Vision
• Natural Language Processing: Language Grounding & Multi-modal NLP

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation

Abstract

Imaging technologies are fundamental to biomedical research and modern medicine, requiring analysis of high-resolution images across various modalities. While multimodal large language models (MLLMs) show promise for biomedical image analysis, most are designed for low-resolution images from general-purpose datasets, risking critical information loss. We investigate how image resolution affects MLLM performance in biomedical applications and demonstrate that: (1) native-resolution training and inference significantly improve performance across multiple tasks, (2) misalignment between training and inference resolutions severely degrades performance, and (3) mixed-resolution training effectively mitigates misalignment and balances computational constraints with performance requirements. Based on these findings, we recommend prioritizing native-resolution inference and mixed-resolution datasets to optimize biomedical MLLMs for transformative impact in scientific research and clinical applications.

Problem

• Investigates image resolution impact on biomedical multimodal models
• Addresses information loss from low-resolution training in medical imaging
• Proposes native-resolution inference and mixed-resolution training solutions

Innovation

• Native-resolution training and inference enhance performance
• Mixed-resolution training mitigates resolution misalignment issues
• Prioritizing native-resolution inference optimizes biomedical MLLMs