Implementing Boston Bowel Preparation Scale Using Artificial Intelligence for Colonic Segmentation and Preparation Scoring

Background & Purpose: The Boston Bowel Preparation Scale (BBPS) relies on subjective assessment with inter-observer variability. This study aimed to develop an AI-based system for real-time automated evaluation of bowel preparation quality to improve accuracy and consistency. Methods: A three-module AI system was developed: 1) EfficientNet-B0 for ileocecal valve detection (trained on 10,000 images); 2) ResNet-50 for colon segmentation (right/left colonic flexures, 12,350 images); 3) ResNet-50 for BBPS scoring (50,000 images). Models were trained on 1,015 annotated colonoscopy videos and validated via independent test datasets and a prospective real-time study. Results: The ileocecal valve detection achieved 99.5% accuracy (AUC 0.996). Colon segmentation showed 89.0% overall accuracy (AUC 0.908–0.913 across regions). The BBPS scoring model achieved 88.3% accuracy, with strong agreement with experts (p<0.001). Prospective real-time evaluation reached 91.5% accuracy, endorsed by endoscopists for clinical utility. Strengths & Limitations: The system integrates anatomical localization and scoring with real-time efficiency, reducing subjectivity. However, multicenter validation is needed for generalizability. Future work should optimize architectures (e.g., Transformers) and assess long-term clinical impacts. Conclusion: This AI system serves as an auxiliary tool for real-time BBPS scoring, aiding standardized assessment. Future multicenter studies and technical integration are needed to validate its clinical utility.