Artificial Intelligence-Assisted Flexure and Landmark Identification Improves Reporting Completeness in Colonoscopy

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Gastroenterology

Background:

• Colorectal cancer poses a significant global health challenge.
• Colonoscopy is the primary method for colorectal cancer detection and prevention.
• Accurate localization of findings during colonoscopy is crucial but often challenging, impacting reporting objectivity.

Purpose of the Study:

• To develop and evaluate an artificial intelligence (AI) system for automated colon segment localization.
• To identify key anatomical landmarks including the appendiceal orifice, ileocecal valve, and flexures during colonoscopy.
• To assess the AI system's potential to improve the accuracy and completeness of colonoscopy reporting.

Main Methods:

• The AI system was trained on a large dataset of 7264 manually annotated colonoscopy images.
• Internal validation involved 1238 images from 215 patients, assessing performance metrics like accuracy, sensitivity, specificity, and F1-score.
• External validation was performed on public video datasets and videos from an independent clinic to assess real-world applicability.

Main Results:

• The AI achieved high performance on internal testing, with 94.5% accuracy and 97.0% specificity for landmark identification.
• External validation demonstrated the AI's ability to identify cecum segments (91.7%) and flexures (66%).
• AI integration significantly improved flexure identification in external clinical data, increasing detection rates and the proportion of videos with both flexures identified.

Conclusions:

• The developed AI reliably detects critical anatomical landmarks, enabling automated colon segment localization.
• Integration of this AI into reporting workflows can enhance lesion localization accuracy and report completeness.
• The AI system shows potential to streamline documentation and reduce endoscopist workload in colonoscopy reporting.