Establishment and validation of a computer-assisted colonic polyp localization system based on deep learning.

Area of Science:

• Medical Artificial Intelligence
• Gastrointestinal Endoscopy
• Computer-Aided Detection (CADe)

Background:

• Artificial intelligence (AI) in colonoscopy aims to enhance polyp and adenoma detection rates.
• Existing deep learning-based computer-assisted detection (CADe) systems often use limited datasets, potentially hindering real-world application.
• Clinical validation of CADe systems for real-time polyp identification is crucial.

Purpose of the Study:

• To develop and evaluate a deep learning-based CADe system using a multicenter, high-quality colonoscopy image dataset.
• To preliminarily validate the CADe system's performance in clinical colonoscopies.

Main Methods:

• A deep learning CADe model was trained and tested on over 71,000 images from 20 centers, curated by 55 colonoscopists.
• Real-time performance was assessed using 47 colonoscopy videos containing 86 confirmed polyps.
• A self-controlled observational study in a clinical setting evaluated the CADe's diagnostic performance.

Main Results:

• The CADe achieved 95.0% sensitivity and 99.1% specificity on the test dataset.
• Video analysis detected 92.2% of polyps, while prospective validation showed 98.4% sensitivity.
• CADe use increased polyp (0.90 vs. 0.82) and adenoma (0.32 vs. 0.30) detection rates, especially for small and flat polyps.

Conclusions:

• Computer-assisted detection (CADe) is feasible for clinical colonoscopy.
• The AI system can assist endoscopists in detecting more polyps and adenomas.
• Further validation is warranted to confirm efficacy across diverse clinical scenarios.