Abstract / Summary
Abstract Ulcerative colitis (UC) is a chronic inflammatory bowel disease characterized by continuous mucosal inflammation, where accurate assessment of endoscopic severity is essential for treatment selection, disease monitoring, and prognosis. Although colonoscopy provides direct visualization of mucosal abnormalities, subjective interpretation, inter-observer variability, subtle lesion characteristics, and limited availability of experienced gastroenterologists can compromise consistent severity assessment. To address these challenges, this study proposes L-GTACNet (Local–Global Transformer Attention-based Graph Convolutional Network), a hybrid deep learning framework for automated multi-level severity classification of UC from colonoscopy images. The framework integrates image enhancement using Retinex-based illumination correction and Contrast-Limited Adaptive Histogram Equalization (CLAHE), a Local–Global Transformer for hierarchical lesion representation, and an attention-guided graph convolutional module for modeling spatial and relational dependencies among discriminative mucosal regions. The proposed architecture is designed to jointly capture fine-grained local manifestations, including erythema, vascular alteration, mucosal friability, and ulceration, together with global contextual patterns across the colonoscopic field. Disease severity is categorized according to the Mayo Endoscopic Score (MES), enabling clinically interpretable severity stratification. The experimental study utilizes a publicly accessible UC colonoscopy dataset comprising 564 patient records and 11,276 images, with annotations established through expert gastroenterologist assessment. Performance is evaluated using accuracy, precision, recall, F1-score, specificity, and area under the receiver operating characteristic curve (AUROC). The proposed L-GTACNet achieves 94.8% classification accuracy, demonstrating its potential as a robust computer-aided system for objective and reproducible assessment of UC severity.