Poster Session I - A57 ARTIFICIAL INTELLIGENCE–ASSISTED ENDOSCOPIC ASSESSMENT OF CROHN’S DISEASE SEVERITY: A SYSTEMATIC REVIEW

TL;DR

This paper reviews how artificial intelligence can help assess the severity of Crohn’s Disease through endoscopic images, showing promising results but needing more standardization.

Contribution

The study systematically reviews AI-based tools for Crohn’s Disease endoscopic assessment, highlighting their potential and current limitations.

Findings

Five studies showed AI model accuracy for CD assessment ranged from 62.4% to 98.4%.

AI models demonstrated high specificity (71.2-99.8%) and area-under-the-curve values (0.565-0.989).

Variability in classification systems and model performance suggests a need for standardization.

Abstract

Endoscopic assessment is integral in guiding the management and monitoring of Crohn’s Disease (CD). While manual scoring systems are widely used, they are subject to significant inter-rater variability. Artificial intelligence (AI) has shown utility in assessing disease activity in ulcerative colitis, yet its role in CD remains underexplored. AI-based tools could improve standardization and efficiency in CD assessment with potential benefits for both clinical practice and research.

Review the accuracy and potential clinical utility of artificial intelligence-based systems in the endoscopic assessment of CD activity and severity.

A systematic search of Ovid MEDLINE, Embase, Pubmed, Scopus, and Cochrane database was performed using PRISMA guidelines to identify publications exploring the use of AI-based tools to assess endoscopic disease activity in CD. Studies meeting inclusion criteria were reviewed and statistical measures and other performance metrics were extracted.

Five studies, published between 2021 and 2024, were included. All used convolutional neural networks to analyze still images from capsule or double-balloon endoscopy, comparing model output against expert readings. Each study used a different classification system for mucosal abnormalities, including ulcer severity, inflammatory stricturing, and villous changes. The accuracy, sensitivity, specificity, area-under-the-curve for the models ranged from 62.4-98.4%, 32.4-98.9%, 71.2-99.8%, and 0.565-0.989 respectively.

AI-based tools show considerable promise in the endoscopic assessment of CD severity. However, heterogeneity in their design and performance underscores the need for further validation and standardization prior to adoption in clinical practice.

A57 Table 1: Publications assessing the utility of artificial intelligence in the endoscopic assessment of Crohn’s disease

None

AUC: Area under curve