arXiv:2511.00801v2 Announce Type: replace 
Abstract: Recent advances in multimodal large language models have enabled remarkable medical image editing capabilities. However, the research community's progress remains constrained by the absence of large-scale, high-quality, and openly accessible datasets built specifically for medical image editing with strict anatomical and clinical constraints. We introduce Med-Banana-50K, a comprehensive 50K-image dataset for instruction-based medical image editing spanning three modalities (chest X-ray, brain MRI, fundus photography) and 23 disease types. Our dataset is constructed by leveraging Gemini-2.5-Flash-Image to generate bidirectional edits (lesion addition and removal) from real medical images. What distinguishes Med-Banana-50K from general-domain editing datasets is our systematic approach to medical quality control: we employ LLM-as-Judge with a medically grounded rubric (instruction compliance, structural plausibility, realism, and fidelity preservation) and history-aware iterative refinement up to five rounds. Beyond single-turn editing, Med-Banana-50K includes 37K failed attempts with full conversation logs for preference learning and alignment research. By providing this large-scale, medically validated, and fully documented resource, Med-Banana-50K establishes a foundation for training and evaluating the next generation of medical image editing models.Our dataset and code are publicly available at [https://github.com/richardChenzhihui/med-banana-50k].

تقديم مجموعة بيانات Med-Banana-50K يمثل تقدمًا كبيرًا في مجال تحرير الصور الطبية. تتكون هذه المجموعة الشاملة من 50,000 صورة، وهي مصممة لدعم التحرير القائم على التعليمات مع الالتزام بمعايير تشريحية وسريرية صارمة. تعتبر توفرها أمرًا حيويًا لأنها تعالج القيود الحالية التي يواجهها الباحثون في الوصول إلى مجموعات بيانات عالية الجودة، مما يمهد الطريق لتطبيقات أكثر ابتكارًا في التصوير الطبي.

La introducción del conjunto de datos Med-Banana-50K marca un avance significativo en el campo de la edición de imágenes médicas. Este conjunto de datos integral, que consta de 50,000 imágenes, está diseñado para respaldar la edición basada en instrucciones mientras se adhiere a estrictos estándares anatómicos y clínicos. Su disponibilidad es crucial, ya que aborda las limitaciones actuales que enfrentan los investigadores para acceder a conjuntos de datos de alta calidad, allanando el camino para aplicaciones más innovadoras en la imagen médica.

L'introduction du jeu de données Med-Banana-50K marque une avancée significative dans le domaine de l'édition d'images médicales. Ce jeu de données complet, composé de 50 000 images, est conçu pour soutenir l'édition basée sur des instructions tout en respectant des normes anatomiques et cliniques strictes. Sa disponibilité est cruciale car elle répond aux limitations actuelles auxquelles les chercheurs sont confrontés pour accéder à des jeux de données de haute qualité, ouvrant ainsi la voie à des applications plus innovantes en imagerie médicale.

The introduction of the Med-Banana-50K dataset marks a significant advancement in the field of medical image editing. This comprehensive dataset, consisting of 50,000 images, is designed to support instruction-based editing while adhering to strict anatomical and clinical standards. Its availability is crucial as it addresses the current limitations faced by researchers in accessing high-quality datasets, ultimately paving the way for more innovative applications in medical imaging.

Med-Banana-50K: A Cross-modality Large-Scale Dataset for Text-guided Medical Image Editing

One More Thing in AI – Your Shortcut to AI Mastery

Med-Banana-50K: A Cross-modality Large-Scale Dataset for Text-guided Medical Image Editing

Was this article worth reading? Share it

One More Thing in AI

Https

Twofold Health

Republiclabs.ai

Media Workbench AI

4o Image Gen

Ready to build your own newsroom?