Semantic Segmentation Dataset for AI-Based Quantification of Clean Mucosa in Capsule Endoscopy

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Gastroenterology

Background:

• Capsule endoscopy (CE) bowel cleanliness assessment is subjective.
• Objective evaluation using AI is needed.
• Developing a robust AI model requires a large, high-quality dataset.

Purpose of the Study:

• To create a large-scale semantic segmentation dataset from CE images.
• To train and validate a convolutional neural network (CNN) for objective bowel cleanliness assessment.
• To develop and verify a formula for quantifying clean mucosal regions.

Main Methods:

• Extracted and classified 10,033 CE frames into 2 or 3 classes (clean, dark, floats/bubbles).
• Trained a CNN model using 169 videos and a semantic segmentation dataset.
• Developed a visualization scale (VS) formula and evaluated its performance using mIoU and Dice index.

Main Results:

• Achieved high performance metrics: 0.7716-0.8927 mIoU and 0.8627-0.9457 Dice index for 2- and 3-class segmentation.
• Demonstrated high clean mucosal prediction accuracy: 94.4% (3-class) and 95.7% (2-class).
• The developed VS formula showed performance nearly identical to ground truth.

Conclusions:

• Established a comprehensive 10-stage semantic segmentation dataset from 179 patients.
• Validated an AI-based approach for accurate ( >94%) and quantitative bowel cleanliness assessment.
• The developed VS equation provides a reliable method for measuring clean mucosal regions in CE.