Training and deploying a deep learning model for endoscopic severity grading in ulcerative colitis using multicenter...

Area of Science:

• Gastroenterology
• Artificial Intelligence
• Medical Imaging

Background:

• The Mayo Clinic Endoscopic Subscore is crucial for assessing ulcerative colitis severity.
• Current grading methods suffer from significant inter- and intrarater variability.
• Machine learning offers potential for standardizing endoscopic assessment.

Purpose of the Study:

• To develop and validate a deep learning system for automated Mayo Clinic Endoscopic Subscore grading from raw colonoscopy videos.
• To mimic the clinical practice of assessing the entire colonoscopy for severity scoring.
• To improve the accuracy and reproducibility of ulcerative colitis endoscopic grading.

Main Methods:

• An end-to-end deep learning system was designed to predict a binary Mayo Clinic Endoscopic Subscore directly from colonoscopy videos.
• The system processed entire videos, identifying informative regions to compute an overall score.
• Training and deployment utilized raw colonoscopy data with colon-section-level ground truth, avoiding manual frame selection.

Main Results:

• The system achieved high accuracy in grading endoscopic videos, with Area Under the ROC Curve values of 0.84 (≥1), 0.85 (≥2), and 0.85 (≥3).
• Evaluation on 1672 videos from a multisite dataset demonstrated robustness.
• The methodology significantly reduced the need for manual annotation.

Conclusions:

• Artificial intelligence can accurately and automatically grade ulcerative colitis severity from full endoscopic videos.
• Training AI models on diverse, multisite datasets is essential for developing robust systems for real-world deployment.
• This AI approach enhances standardization and reproducibility in endoscopic assessment.