A Clinician-Guided Framework for Endoscopic AI: Developing PanEndoAtlas and Benchmarking Foundation Models Across the...

Endoscopic procedures play a central role in the diagnosis and management of gastrointestinal (GI) diseases, yet the field lacks large-scale, clinically diverse benchmarks and unified datasets to evaluate vision foundation models. We introduce PanEndoSuite, the first unified ecosystem for endoscopic AI, developed through systematic collaboration between AI researchers and practicing gastroenterologists. PanEndoSuite consists of three complementary components: PanEndoAtlas, PanEndoX, and PanEndoFM. PanEndoAtlas is a harmonized dataset of over 420,000 labeled images from 30 public endoscopy datasets across 13 countries and 26 hospitals, creating a clinically-grounded hierarchical taxonomy that mirrors diagnostic reasoning patterns across 111 GI diseases. PanEndoX is a benchmark of 10 clinically grounded tasks, including hierarchical GI-tree classification, Barrett's esophagus grading, ulcerative colitis scoring, polyp subtyping, Boston Bowel Preparation Scale assessment, multi-organ disease classification, and anatomical landmark identification-designed to probe generalization across anatomical regions, disease presentations, and annotation granularities. PanEndoFM is a foundation model pretrained on a 10 million-image corpus curated from public data sources, spanning the entire GI tract. We benchmark PanEndoFM against two endoscopy-specific foundation models (EndoFM-LV, EndoSSL) and two general-purpose vision models (ViT-B/16, ResNet-50). PanEndoFM achieves the highest macro-AUC on 6 of 10 tasks, demonstrating broad clinical generalization; EndoFM-LV performs best on colon-focused tasks, EndoSSL excels in polyp subtyping, and ViT-B/16 shows strengths on small-intestine conditions. Together, PanEndoSuite establishes a foundation for building robust, generalist AI systems in gastrointestinal endoscopy that bridge current AI capabilities and clinical practice.