REAL-Colon: A dataset for developing real-world AI applications in colonoscopy

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Colon polyp detection is crucial for colorectal cancer prevention.
• AI-based computer-aided detection (CADe) and diagnosis (CADx) systems show promise in enhancing colonoscopy.
• Existing datasets lack the quality and real-world representation needed for robust AI development.

Purpose of the Study:

• Introduce the REAL-Colon dataset, a large-scale, high-quality resource for AI research in colonoscopy.
• Provide a comprehensive dataset that accurately reflects real-world colonoscopy procedures.
• Facilitate the development and benchmarking of advanced AI algorithms for polyp detection and diagnosis.

Main Methods:

• Compiled 2.7 million native video frames from sixty full-resolution, multi-center colonoscopy recordings.
• Included 350,000 bounding-box annotations supervised by expert gastroenterologists.
• Integrated comprehensive patient clinical, acquisition, and polyp histopathological data.

Main Results:

• The REAL-Colon dataset is the largest and highest quality publicly available resource for colonoscopy AI research.
• It features full-resolution videos and detailed annotations, overcoming limitations of previous datasets.
• The dataset's heterogeneity and included clinical data support diverse AI model development.

Conclusions:

• REAL-Colon is a unique resource for advancing AI in colonoscopy due to its size, quality, and heterogeneity.
• Open access to this dataset promotes rigorous, reproducible research in AI-driven polyp detection.
• This resource will foster the creation of more accurate and reliable AI tools for colonoscopy.