EndoAtlas

Dataset Description

EndoAtlas is a multi-center and multi-modal gastrointestinal endoscopy database developed to support artificial intelligence research in digestive endoscopy. The database contains endoscopic images, videos, clinical information, pathological findings, and expert annotations collected from approximately 40 medical centers in China. The current version includes data from approximately 70,000 patients, comprising around 50 million high-definition endoscopic images and 200,000 endoscopic videos. The database covers a wide range of gastrointestinal endoscopy procedures, including gastroscopy, colonoscopy, capsule endoscopy, endoscopic ultrasound, and endoscopic retrograde cholangiopancreatography.

EndoAtlas is structured using a technique-site-task framework. This organization allows data to be indexed and accessed according to examination technique, anatomical location, and artificial intelligence task type. The database is designed to support multiple endoscopy-related AI tasks, including image classification, lesion detection, semantic segmentation, regression analysis, quality control, and model validation.

Because the database contains clinically linked information and potentially privacy-sensitive data, the complete EndoAtlas database is released through a controlled application process. De-identified sample data, data dictionaries, and training code are provided to support database demonstration, method development, and reproducibility.