MedTrinity-25M: 25 Million Multimodal Medical Images with ROI Annotations

MedTrinity-25M provides 25 million medical images across 10 modalities and 65 diseases, released by UCSC-VLAA in 2024. The collection features multigranular annotations including global textual descriptions and local region-of-interest bounding boxes.

Use Cases

• Medical visual question answering using region-specific descriptions and disease types
• Localizing medical anomalies using the provided region-of-interest bounding boxes
• Cross-modal retrieval tasks using the 10-modality image-text pairs

Strengths

• 25 million medical images
• 10 distinct imaging modalities
• Annotations for 65+ diseases
• Multigranular labels including local bounding boxes

Limitations

• Geographic bias toward United States medical data
• High computational overhead for processing 25 million high-resolution images

Provenance

Source

UCSC-VLAA

Freshness

Last updated 2024-10-11

Geography

United States