Med-Banana-50K: 50,000 Text-Guided Medical Image Edits Across Three Modalities

Med-Banana-50K is a dataset of 50,000 images for instruction-based medical image editing. It spans three imaging modalities (Chest X-ray, Brain MRI, Fundus Photography) and 23 disease types, with edits generated from real images using the Gemini-2.5-Flash-Image model. The dataset was created by RichardChenZH and was last updated on March 12, 2026.

Use Cases

• Training models for text-guided medical image editing based on the described bidirectional lesion edits.
• Developing cross-modality medical image synthesis based on the inclusion of Chest X-ray, MRI, and Fundus Photography.
• Benchmarking deepfake detection algorithms for medical images based on the dataset's focus on generated edits.
• Researching disease representation in AI-generated medical imagery based on the coverage of 23 disease types.

Strengths

• Contains 50,000 images, providing a substantial scale for model training.
• Covers three distinct medical imaging modalities: Chest X-ray, Brain MRI, and Fundus Photography.
• Includes edits for 23 different disease types, offering diversity in medical conditions.
• Features bidirectional edits (lesion addition and removal) for each image.

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 bias inherent to the source images and the generative model used for editing.

Provenance

Source

RichardChenZH via Hugging Face.

Collection Method

Edits were generated from real medical images using the Gemini-2.5-Flash-Image model.

Freshness

Last updated 2026-03-12 10:41:51; freshness should be verified.