Multicenter Validation of a ResNet50-Based Deep Convolutional Network for Endoscopic Grading of Ulcerative Colitis Severity

Abstract Objective To tackle inconsistent endoscopic severity scores in ulcerative colitis (UC) caused by subjective clinician interpretation, this study developed a deep convolutional neural network (DCNN)-based system. Unlike many studies relying on single-center or public data, this model was externally validated using multi-institutional data to enhance generalizability. Methods Colonoscopy images from patients with UC were retrospectively collected from one hospital (Dec 2019–May 2024) for training and testing. An AI system based on ResNet50 was developed to assess endoscopic severity. Additional images from two other tertiary hospitals (May–Dec 2024) were used for external validation. Performance was measured by sensitivity, specificity, accuracy, PPV, NPV, AUC, and weighted Kappa. Results The optimized model achieved sensitivities of 91.8% (normal), 81.8% (mild), 78.8% (moderate), and 64.8% (severe), with a mean AUC of 0.899. External validation on 1,732 images showed improved sensitivities: 96.0% (normal), 84.2% (mild), 79.1% (moderate), and 92.1% (severe), with specificities > 94% across categories and a mean AUC of 0.885. AI-endoscopist agreement (ICC = 0.840, QWK = 0.840) was comparable to inter-expert consistency (ICC = 0.880, QWK = 0.880). Conclusions The ResNet50-based AI system demonstrated high diagnostic accuracy for UC severity evaluation, performing comparably to experienced endoscopists. It offers a reliable and generalizable tool for supporting clinical assessment and treatment planning.