Med-Banana-50K is a large-scale medical image editing dataset released by the National University of Singapore in 2025. The related research paper is as follows: Med-Banana-50K: A Cross-modality Large-Scale Dataset for Text-guided Medical Image EditingIt aims to cover bidirectional editing tasks of lesions under various medical imaging modalities and disease conditions.

This dataset contains over 50,000 medical image editing samples, along with approximately 37,000 failed editing attempts and their complete evaluation logs. Based on real clinical medical images, the dataset covers three medical imaging modalities: chest X-ray, brain MRI, and fundus imaging, encompassing 23 disease types. Each sample supports both lesion addition and removal editing, with the editing process generated using the Gemini-2.5-Flash-Image model. All samples underwent an automated evaluation process based on LLM-as-Judge, assessing aspects such as instruction compliance, structural rationality, image realism, and fidelity of original information. Up to five rounds of iterative optimization were performed as needed, and all evaluation results and processes were fully recorded.