Computer-aided automated diminutive colonic polyp detection in colonoscopy by using deep machine learning system;...

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Colonic polyps are precursors to colorectal cancer and can be missed during colonoscopy due to size, location, or human error.
• Missed polyps are estimated to occur in about 25% of colonoscopies.
• Artificial intelligence (AI) offers a potential solution to improve polyp detection rates and reduce colorectal cancer incidence.

Purpose of the Study:

• To develop an indigenous AI system for detecting diminutive colonic polyps in real-life colonoscopy scenarios.
• To ensure compatibility with standard high-definition colonoscopy equipment and video capture software.

Main Methods:

• A masked region-based convolutional neural network was trained to detect and localize colonic polyps.
• The model was trained on 1,039 image frames from three colonoscopy video datasets, including real-life and publicly available sources.
• Testing involved augmented image frames to simulate real-life distortions, with the AI localizing polyps using bounding boxes.

Main Results:

• The AI system achieved a mean average precision (specificity) of 88.63% for automatic polyp detection.
• The system demonstrated 100% sensitivity, correctly identifying all polyps in the testing dataset (no false negatives).
• The average polyp size was 5 mm (±4 mm), with a processing time of 96.4 minutes per frame.

Conclusions:

• The developed AI system accurately detects colonic polyps, even with variations in bowel preparation and small polyp sizes.
• This AI technology has the potential to significantly improve diagnostic accuracy in colonoscopy.
• By enhancing polyp detection, the AI system can contribute to reducing the incidence of colorectal cancer.