Development and validation of a three-dimensional deep learning-based system for assessing bowel preparation on...

Area of Science:

• Artificial Intelligence in Medicine
• Medical Imaging Analysis
• Gastroenterology Technology

Background:

• Existing image-based models for evaluating colonoscopy bowel preparation have shown limited performance and generalization capabilities.
• There is a need for more precise and stable AI systems to assess bowel preparation quality from colonoscopy videos.

Purpose of the Study:

• To develop and validate a novel AI system, ViENDO, for accurate assessment of bowel preparation quality in colonoscopy videos.
• To evaluate the performance of ViENDO on full-length colonoscopy videos from multiple centers and compare it with endoscopist performance.

Main Methods:

• Proposed ViENDO system comprising two Convolutional Neural Networks (CNNs): Information-Net for frame filtering and BBPS-Net for bowel preparation assessment.
• BBPS-Net utilized 3D CNN technology trained on 5,566 video clips to detect insufficient bowel preparation based on the Boston Bowel Preparation Scale (BBPS).
• ViENDO was applied to full-length colonoscopy videos for predicting BBPS segment scores and underwent a human-machine contest.

Main Results:

• BBPS-Net achieved an area under the curve of 0.98 and 95.2% accuracy in identifying inadequate bowel preparation from video clips.
• ViENDO demonstrated high accuracy in assessing bowel cleanliness on full-length videos: 93.8% (internal) and 91.7% (external dataset).
• The human-machine contest indicated ViENDO's accuracy was slightly superior to most endoscopists, without statistical significance.

Conclusions:

• The 3D-CNN-based AI model (ViENDO) exhibits strong performance in evaluating full-length bowel preparation from colonoscopy videos.
• ViENDO holds significant potential as a tool to assist or substitute endoscopists in providing BBPS-based assessments during routine clinical practice.