MedPMC scales high-fidelity medical image-text data for multimodal foundation models

Researchers introduce MedPMC, an automated framework that transforms permissively licensed literature from PubMed Central into high-fidelity infrastructure for medical multimodal models. Applied to 6.1 million articles, the system curates 11 million medical image-text pairs with strong component evaluation scores and high clinical relevance.

• Curated 11 million image-text pairs from 6.1 million PMC articles using automated screening and separation.
• Achieved 95.3% medical relevance in manual review, compared to 19.7% in prior datasets.
• A MedPMC-trained CLIP-style model improved average zero-shot AUC by 7.1 percentage points across 26 benchmarks.
• Enhanced medical visual question-answering by 1.9 and 16.9 percentage points when used as a vision encoder.
• Improved morphology-to-image retrieval Recall@5 by 11.7 percentage points on dermatology photographs.

The findings demonstrate that high-fidelity literature curation strengthens medical multimodal foundation models across both benchmark and clinical settings, with the framework, corpus, and models publicly released.