Reproduce “REAL-Colon: A dataset for developing real-world AI applications in colonoscopy” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Carlo Biffi, Giulio Antonelli, Sebastian Bernhofer, Cesare Hassan, Daizen Hirata, Mineo Iwatate, Andreas Maieron, Pietro Salvagnini and 1 more
Cosmo Intelligent Medical Devices and an international clinical consortium introduce REAL-Colon, a dataset of 60 full-length, high-resolution colonoscopy videos with 351,264 bounding box annotations for 132 polyps. This resource is designed to facilitate the development and benchmarking of AI models for real-world colonoscopy, addressing limitations of prior datasets by including extensive 'negative frames' and diverse clinical contexts.
Key insights
- A vast majority (87.6%) of real-world colonoscopy frames do not contain polyps, underscoring the necessity of including 'negative frames' in datasets for training realistic AI models.
- Polyp visibility in real colonoscopy is transient, with an average appearance duration of 1 minute 15 seconds, and less than 0.3% of frames containing multiple polyps, posing distinct challenges for real-time tracking.
- Training object detection models like SSD on the full REAL-Colon dataset, including all negative frames, significantly improves performance (e.g., AP 0.216) and reduces false positive rates compared to training only on positive samples.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.