The optimal labelling method for artificial intelligence-assisted polyp detection in colonoscopy.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Gastroenterology

Background:

• Colon polyp labeling for machine learning lacks standardized methodology.
• Developing accurate AI models for polyp detection requires optimized annotation techniques.

Purpose of the Study:

• To determine the optimal colon polyp annotation method for enhancing AI model accuracy.
• To compare different bounding box extension percentages against exact segmentation for AI training.

Main Methods:

• Utilized 3542 colonoscopy images for manual polyp annotation.
• Compared exact outline segmentation with rectangle boxes extended by 10% to 50%.
• Developed a U-Net convolutional neural network model for automatic segmentation.

Main Results:

• Extending the bounding box by 20% yielded the best performance: 95.42% accuracy, 94.84% sensitivity, and 95.41% F1-score.
• Exact outline segmentation achieved 99.6% sensitivity but only 77.47% precision.
• The 20% extended bounding box model demonstrated superior performance with an AUC of 0.971.

Conclusions:

• Annotation methodology directly impacts AI model predictability in colon polyp detection.
• A 20% bounding box extension provides the most accurate AI predictive model.
• Standardized colon polyp labeling protocols are essential for comparing AI model precision.