Deep Learning Localizes and Identifies Polyps in Real Time With 96% Accuracy in Screening Colonoscopy

Area of Science:

• Gastroenterology
• Artificial Intelligence
• Medical Imaging

Background:

• Colonoscopy's effectiveness in preventing colorectal cancer relies on the adenoma detection rate (ADR).
• Current ADRs vary significantly among endoscopists, and improved detection is crucial for cancer prevention.
• New strategies are needed to enhance ADR during colonoscopy procedures.

Purpose of the Study:

• To evaluate the efficacy of computer-assisted image analysis using convolutional neural networks (CNNs) for improving polyp detection during colonoscopies.
• To assess the potential of deep learning models to serve as a surrogate for ADR improvement.

Main Methods:

• Deep convolutional neural networks (CNNs) were designed and trained on a dataset of 8,641 hand-labeled colonoscopy images from over 2,000 patients.
• The trained CNN models were tested on 20 colonoscopy videos (5 hours total duration).
• Expert colonoscopists reviewed 9 colonoscopy videos with and without CNN assistance, comparing their findings to the CNN's performance.

Main Results:

• The CNN achieved high accuracy (96.4%) and an area under the receiver operating characteristic curve of 0.991 in identifying polyps on manually labeled images.
• With CNN assistance, expert reviewers identified 17 additional polyps compared to unassisted review, detecting a total of 45 polyps.
• The CNN system demonstrated real-time polyp detection and localization capabilities suitable for clinical use.

Conclusions:

• The developed CNN system shows significant promise for increasing ADR and consequently reducing interval colorectal cancers.
• The computer-assisted system achieved high accuracy and efficiency in polyp detection, comparable to expert performance.
• Further validation in large, multicenter clinical trials is recommended to confirm the system's effectiveness and clinical utility.