Radiology Images With Captions And Medical Concepts

ROCOv2 is a multimodal dataset containing 79,789 radiological images paired with captions and extracted medical concepts. It was created by eltorio and updated in November 2024. The dataset expands upon its predecessor by adding 35,705 new images and improving its concept extraction process.

Use Cases

• Train image-to-text models to generate captions from radiological images.
• Develop concept detection models to identify medical concepts from radiology images.
• Fine-tune vision-language models for medical question answering using paired images and captions.
• Benchmark multimodal retrieval systems for finding relevant medical images based on textual concepts or captions.

Strengths

• Contains 79,789 radiological images, providing a substantial collection for training.
• Adds 35,705 new images compared to the previous version, indicating growth.
• Includes multimodal data with images, captions, and medical concepts per sample.

Limitations

• Specific row counts for captions and concepts are not provided, making scope unclear.
• The geographic and temporal coverage of the source radiology images is unknown.
• Potential class imbalance or bias exists as the data is sourced from a specific subset of medical literature.

Provenance

Source

PMC Open Access Subset of medical literature.

Collection Method

Images and associated text extracted and processed from medical publications.

Freshness

Last updated on 2024-11-13.