Automated endoscopic detection and classification of colorectal polyps using convolutional neural networks

Area of Science:

• Gastroenterology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate in vivo differentiation of colorectal polyps (CP) is crucial for the 'resect and discard' strategy.
• Artificial intelligence (AI) and deep learning show promise for object recognition in medical imaging.
• There is a need for AI systems to assist in the detection and classification of CP during colonoscopy.

Purpose of the Study:

• To construct an AI system capable of accurately detecting and classifying colorectal polyps (CP) using stored colonoscopy images.
• To evaluate the diagnostic performance of the AI system in terms of speed and accuracy.
• To assess the potential of AI as a support tool for colonoscopy.

Main Methods:

• A deep convolutional neural network (CNN) architecture, Single Shot MultiBox Detector, was employed.
• The CNN was trained on a dataset of 16,418 CP images and 4013 normal colorectum images.
• Performance was validated on 7077 colonoscopy images, including 1172 CP images from 309 types of CP.

Main Results:

• The CNN processed images at 20ms per frame, detecting 1246 CP with 92% sensitivity and 86% positive predictive value (PPV).
• Sensitivity and PPV varied between white light (90%/83%) and narrow band imaging (97%/98%).
• 83% of detected CPs were accurately classified, with 97% of adenomas identified under white light imaging.

Conclusions:

• The developed CNN demonstrates significant potential for detecting and classifying colorectal polyps (CP) from endoscopic images.
• This AI system could serve as a valuable diagnostic support tool in colonoscopy.
• Further application of AI in CP diagnosis is promising for improving patient care.