MedicalNarratives: 4.7M Multimodal Medical Image-Text Pairs with Spatial Grounding

4.7 million image-text pairs form a large-scale multimodal corpus for clinical imaging. The dataset, created by kzhang20, includes approximately 1 million samples with spatial grounding annotations like cursor traces or bounding boxes. It spans 12 imaging domains such as X-ray, CT, and MRI, and was last updated on December 4, 2025.

Use Cases

• Train medical image classification models based on the 12 imaging domains mentioned.
• Develop image-text retrieval systems for clinical imaging based on the multimodal pairs.
• Build object detection or segmentation models for medical images based on the bounding-box or cursor-trace spatial grounding.
• Create vision-language models for medical education or video understanding based on the pedagogical video source.

Strengths

• Contains 4.7 million image-text pairs, indicating a large-scale resource.
• Includes approximately 1 million samples with spatial grounding annotations.
• Covers 12 diverse medical imaging domains, including X-ray, CT, and MRI.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Data may reflect source bias inherent to its construction from pedagogical videos and scientific articles.

Provenance

Source

huggingface, author kzhang20

Collection Method

Built from pedagogical videos and scientific articles.

Time Range

null

Freshness

Last updated 2025-12-04 01:30:18; freshness should be verified.

Geography

null