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

Area of Science:

• Artificial Intelligence in Medicine
• Gastroenterology
• Medical Imaging Analysis

Background:

• Endoscopic procedures are crucial for diagnosing and managing gastrointestinal (GI) diseases.
• Current AI development in endoscopy lacks large-scale, clinically diverse benchmarks and unified datasets.
• Evaluating vision foundation models in this field is challenging due to data fragmentation.

Purpose of the Study:

• To introduce PanEndoSuite, the first unified ecosystem for AI in gastrointestinal endoscopy.
• To establish a comprehensive benchmark for evaluating AI models in clinical endoscopy.
• To develop and assess a novel foundation model for endoscopic AI applications.

Main Methods:

• Developed PanEndoSuite, comprising PanEndoAtlas (harmonized dataset), PanEndoX (benchmark tasks), and PanEndoFM (foundation model).
• PanEndoAtlas includes over 420,000 labeled images with a hierarchical taxonomy for 111 GI diseases.
• PanEndoX features 10 clinically grounded tasks, and PanEndoFM is pretrained on a 10 million-image corpus.

Main Results:

• PanEndoFM achieved the highest macro-AUC on 6 out of 10 benchmark tasks, demonstrating broad clinical generalization.
• Specialized models showed strengths in specific areas: EndoFM-LV for colon tasks, EndoSSL for polyp subtyping, and ViT-B/16 for small intestine conditions.
• The study established performance baselines for AI models across various endoscopic diagnostic challenges.

Conclusions:

• PanEndoSuite provides a foundation for developing robust, generalist AI systems in gastrointestinal endoscopy.
• The developed ecosystem bridges the gap between current AI capabilities and practical clinical needs.
• This work facilitates the advancement of AI-driven diagnostic and management tools in clinical endoscopy.