[论文解读] BiomedCLIP: a multimodal biomedical foundation model pretrained from fifteen million scientific image-text pairs

Sheng Zhang, Yanbo Xu+22 more|arXiv (Cornell University)|Mar 2, 2023

Topic Modeling被引用 95

一句话总结

BiomedCLIP 在 PubMed Central 的 PMC-15M 上预训练一个领域特定的视觉–语言基础模型，PMC-15M 是一个来自 PubMed Central 的 15 million 图像-字幕 数据集，在生物医学领域的检索、分类和 VQA 任务中达到 state-of-the-art 结果。

ABSTRACT

Biomedical data is inherently multimodal, comprising physical measurements and natural language narratives. A generalist biomedical AI model needs to simultaneously process different modalities of data, including text and images. Therefore, training an effective generalist biomedical model requires high-quality multimodal data, such as parallel image-text pairs. Here, we present PMC-15M, a novel dataset that is two orders of magnitude larger than existing biomedical multimodal datasets such as MIMIC-CXR, and spans a diverse range of biomedical image types. PMC-15M contains 15 million biomedical image-text pairs collected from 4.4 million scientific articles. Based on PMC-15M, we have pretrained BiomedCLIP, a multimodal foundation model, with domain-specific adaptations tailored to biomedical vision-language processing. We conducted extensive experiments and ablation studies on standard biomedical imaging tasks from retrieval to classification to visual question-answering (VQA). BiomedCLIP achieved new state-of-the-art results in a wide range of standard datasets, substantially outperforming prior approaches. Intriguingly, by large-scale pretraining on diverse biomedical image types, BiomedCLIP even outperforms state-of-the-art radiology-specific models such as BioViL in radiology-specific tasks such as RSNA pneumonia detection. In summary, BiomedCLIP is a fully open-access foundation model that achieves state-of-the-art performance on various biomedical tasks, paving the way for transformative multimodal biomedical discovery and applications. We release our models at https://aka.ms/biomedclip to facilitate future research in multimodal biomedical AI.

研究动机与目标

• Address the need for a large, diverse, open biomedical vision-language dataset.
• Pretrain a domain-specific multimodal model that leverages long biomedical captions and high-resolution images.
• Evaluate BiomedCLIP across retrieval, zero-shot classification, and medical VQA to establish state-of-the-art performance.
• Demonstrate that large-scale diverse biomedical pretraining can outperform radiology-centric models on certain tasks.

提出的方法

• Create PMC-15M, a public dataset of 15 million figure-caption pairs from PubMed Central articles.
• Adapt CLIP with a biomedical-tailored text encoder (PubMedBERT) and larger vision encoders at higher resolutions.
• Implement patch dropout and tailored batch sizing to optimize biomedical pretraining efficiency and performance.
• Enhance data diversity by constructing PMC-Fine-Grained-46M, including split panels and in-line citances for finer-grained image-text pairs.
• Evaluate using eight standard biomedical vision-language tasks spanning retrieval, classification, and VQA, and compare to general-domain CLIP, PubMedCLIP, MedCLIP, and BioViL.

实验结果

研究问题

• RQ1 Does training on a large, diverse, open biomedical image-text dataset improve vision-language generalization in biomedicine?
• RQ2 How does BiomedCLIP compare to general-domain and radiology-focused models on retrieval, classification, and VQA tasks?
• RQ3 What domain-specific adaptations (text and image encoders, tokenization, image resolution) maximize biomedical multimodal performance?
• RQ4 Can diverse pretraining data enable cross-domain transfer within biomedical disciplines (e.g., radiology vs pathology)?

主要发现

• BiomedCLIP achieves high cross-modal retrieval accuracy, with top-1 and top-5 recalls substantially outperforming general-domain CLIP and PubMedCLIP.
• BiomedCLIP delivers superior zero-shot image classification across five datasets and outperforms radiology-focused BioViL on RSNA benchmarks with less labeled data.
• BiomedCLIP attains state-of-the-art results on medical VQA benchmarks (VQ-RAD, SLAKE) and shows strong open-ended question performance.
• Pretraining on PMC-15M’s diverse image types yields robust representations that surpass radiology-only pretraining in several tasks.
• The study demonstrates that large-scale, domain-specific multimodal pretraining can enable open-access, high-performance biomedical AI models.