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

• DOI: 10.48550/arxiv.2403.02163
• Publication Date: 2024-03-04
• Authors: Carlo Biffi; Giulio Antonelli; Sebastian Bernhofer; Cesare Hassan; Daizen Hirata; Mineo Iwatate; A Maieron; Pietro Salvagnini; Andrea Cherubini

Detection and diagnosis of colon polyps are key to preventing colorectal cancer. Recent evidence suggests that AI-based computer-aided detection (CADe) (CADx) systems can enhance endoscopists' performance boost colonoscopy effectiveness. However, most available public datasets primarily consist still images or video clips, often at a down-sampled resolution, do not accurately represent real-world procedures. We introduce the REAL-Colon (Real-world multi-center Endoscopy Annotated Library) dataset: compilation 2.7M native frames from sixty full-resolution, recordings across multiple centers. The dataset contains 350k bounding-box annotations, each created under supervision expert gastroenterologists. Comprehensive patient clinical data, acquisition information, polyp histopathological information also included in video. With its unprecedented size, quality, heterogeneity, is unique resource for researchers developers aiming advance AI research colonoscopy. Its openness transparency facilitate rigorous reproducible research, fostering development benchmarking more accurate reliable colonoscopy-related algorithms models.