ReXVQA: 653,834 Visual Question Answering Pairs for Chest Radiology

ReXVQA provides 653,834 visual question-answering pairs linked to 160,000 radiological studies and 273,004 chest X-ray images. Created by the Rajpurkar Lab, the data originates from 109,487 patients across 79 medical sites within three U.S. health systems.

Use Cases

• Training multimodal models to generate text answers from chest X-ray image inputs
• Benchmarking model accuracy on clinical questions across 79 different medical sites
• Fine-tuning vision-language models for radiological study interpretation

Strengths

• 653,834 question-answer pairs
• 273,004 unique chest X-ray images
• Multi-site diversity from 79 medical locations

Limitations

• Geographic bias to U.S. healthcare systems
• Restricted 'other' license type

Provenance

Source

Rajpurkar Lab

Collection Method

Constructed from ReXGradient-160K

Freshness

Last updated November 2024

Geography

United States