A Machine Learning-Based System for Real-Time Polyp Detection (DeFrame): A Retrospective Study.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Existing AI-based computer-aided detection systems for endoscopy often rely on limited, static image datasets, hindering clinical application.
• There is a need for AI systems that can accurately detect intestinal polyps in real-time during actual clinical endoscopy procedures.

Purpose of the Study:

• To develop and validate the DeFrame system, integrating multiple deep learning algorithms for accurate, real-time intestinal polyp detection.
• To assess the system's performance using diverse datasets, including retrospective video data and public image collections.

Main Methods:

• Collected 681 colonoscopy videos for analysis, extracting images for training and validation.
• Utilized a combination of retrospective clinical data and public datasets to train and validate the DeFrame system.
• Evaluated system performance using metrics such as sensitivity, specificity, precision, recall, and F1 score on various datasets and full colonoscopy videos.

Main Results:

• The DeFrame system achieved a sensitivity of 79.54% and specificity of 95.83% for intestinal polyp detection.
• Recall and precision for polyp detection were 95.43% and 92.12%, respectively.
• In full colonoscopy video testing, the system demonstrated 100% recall and 80.80% precision.

Conclusions:

• The developed DeFrame system is fast, accurate, and reliable for detecting intestinal polyps.
• The DeFrame system shows feasibility for integration into routine clinical endoscopy practices.