RadImageNet-VQA: 7.5 Million VQA Samples for CT and MRI Exams

RadImageNet-VQA contains 750,000 CT and MRI images paired with 7.5 million generated visual question answering samples and 750,000 medical captions. Developed by Raidium and updated in late 2025, the dataset is built upon expert-curated anatomical and pathological annotations from the RadImageNet corpus.

Use Cases

• Training multimodal models for radiologic visual question answering using the 7.5M generated samples
• Fine-tuning medical image captioning systems using the 750K provided captions
• Benchmarking zero-shot performance of vision-language models on anatomical and pathological features

Strengths

• 7.5 million generated VQA samples
• 750,000 expert-curated medical captions
• Derived from expert-curated anatomical and pathological annotations
• Includes both CT and MRI modalities

Limitations

• VQA samples are generated rather than human-authored, which may affect linguistic naturalness
• Specific column names and data schema are not provided in the primary metadata

Provenance

Source

Raidium, derived from the RadImageNet dataset

Collection Method

Generated from expert-curated anatomical and pathological annotations

Freshness

Last updated December 2025.

Geography

United States