S1814 Leveraging Artificial Intelligence for Endoscopic Healing in Ulcerative Colitis: A Systematic Review Anchored on Mayo Endoscopic Subscoring

Introduction: Mucosal healing is a key treatment goal in ulcerative colitis (UC), offering prognostic value beyond conventional endoscopic remission. The Mayo Endoscopic Subscore (MES), which classifies inflammation into 4 stages: 0 (normal/inactive), 1 (mild), 2 (moderate), and 3 (severe), is widely used to assess healing but is limited by subjectivity and variability. Histological assessment is more precise but invasive and impractical for routine use. Artificial intelligence (AI) offers a promising solution by automating and standardizing endoscopic evaluation. This systematic review assesses the diagnostic accuracy and clinical utility of AI models in detecting mucosal healing in UC, as defined by MES. Methods: A systematic search of PubMed and Google Scholar was performed through 2025 in accordance with PRISMA guidelines to identify studies evaluating machine learning (ML) and deep learning (DL) models for detecting mucosal healing in UC using the MES classifications. Eligible studies involved expert-labeled MES data to train and validate AI models. Data on study design, model development, diagnostic accuracy, and area under the curve (AUC) values were extracted. Results: Sixteen studies (2015–2025) involving over 11,000 patients and more than 200,000 endoscopic images/videos evaluated AI models for MES classification in UC. Most commonly used Convolutional neural networks (CNN) architectures (e.g., ResNet, DenseNet, InceptionV3) targeted MES 0 vs 1 or 0–1 vs 2–3, reflecting their focus on detecting mucosal healing vs active disease. Reported AUCs ranged from 0.70 to 0.99, with several models achieving ≥0.95, indicating high accuracy for detecting mucosal healing. Prospective studies demonstrated the clinical feasibility and reliability of AI models. Integration of vascular healing markers and multimodal imaging (WLI + NBI) enhanced relapse prediction. Across studies, AI consistently outperformed or supported expert assessments, especially in differentiating MES 0 from 1, an area prone to interobserver variability, highlighting its potential to standardize endoscopic scoring. Conclusion: This review highlights the potential of ML and DL models, specifically CNN, to accurately detect mucosal healing in ulcerative colitis, including alignment with histological remission. These AI tools offer a non-invasive, standardized approach that may improve disease monitoring. Further clinical validation and integration into treat-to-target strategies are needed to optimize patient management.