Japanese Medical VQA 12M: 12 Million Multimodal Medical Image-Text Pairs

Japanese Medical VQA 12M provides 12 million multimodal records for medical visual question answering, developed by MIL-UT and released in 2026. It consists of medical images paired with English and Japanese text across five distinct data-construction stages including captions and Q&A pairs.

Use Cases

• Training medical visual question answering (VQA) models using the question-answering pairs
• Fine-tuning vision-language models for Japanese medical image captioning using Japanese translations of source captions
• Evaluating cross-lingual medical text alignment using the Japanese translations of enriched captions

Strengths

• 12 million multimodal records
• Bilingual support in Japanese and English
• Derived from Open-PMC-18M medical source

Limitations

• Potential translation artifacts from automated Japanese translation
• Large dataset size requires significant storage and compute resources

Provenance

Source

MIL-UT, derived from Open-PMC-18M

Collection Method

Automated data construction including translation and enrichment of PMC medical literature

Freshness

Last updated March 2026.

Geography

Global medical literature with Japanese localization