Deep Learning-Based Semantic Segmentation for Objective Colonoscopy Quality Assessment.

Area of Science:

• Medical imaging analysis
• Artificial intelligence in healthcare
• Gastroenterology

Background:

• Colonoscopy quality assessment is crucial for effective diagnosis.
• Traditional methods like the Boston Bowel Preparation Scale have limitations in objective evaluation.
• Deep learning offers a potential solution for automated and detailed analysis of colonoscopy videos.

Purpose of the Study:

• To develop and validate a deep learning-based semantic segmentation network for objective colonoscopy quality evaluation.
• To compare the AI-driven assessment with expert evaluations using the Boston scale.
• To introduce a method that quantifies colonic mucosa, residues, and artifacts for a comprehensive quality assessment.

Main Methods:

• Thousands of colonoscopy frames were processed using color-based image analysis to extract features.
• A semantic segmentation neural network was trained on annotated frames to classify intestinal mucosa, residues, artifacts, and lumen.
• Pixel statistics from the network's analysis were correlated with expert Boston Bowel Preparation Scale (BBPS) scores.

Main Results:

• The deep learning model accurately classified key regions in colonoscopy frames.
• Spearman correlation showed moderate to strong agreements between AI pixel scores and BBPS (e.g., 0.69 for overall pixel scores, 0.63 for mucosa).
• AI-based evaluation demonstrated fair compatibility with expert assessments (Cohen's Kappa = 0.28).

Conclusions:

• The proposed deep learning semantic segmentation approach is a promising tool for objective colonoscopy quality evaluation.
• This AI method provides a more comprehensive assessment than the Boston scale by quantifying multiple image components.
• The AI model's ability to analyze mucosa, residues, and artifacts enhances the objectivity and detail of colonoscopy quality assessment.