Deep learning-based assessment of ulcerative colitis activity from full-length endoscopic videos with spatial characterisation and histological correlation

Background and study aims: Endoscopic assessment of ulcerative colitis (UC) is central to clinical decision-making however remains subjective and limited in characterisation of disease distribution. Artificial intelligence (AI) enables analysis of entire endoscopic examination, rather than relying on selected views. We aimed to develop and validate a deep learning model for automated assessment of UC severity from full-length endoscopic videos, introducing spatial representation of inflammation (Continuous Disease Score, CDS), and assess histological correlation. Patients and methods: Full-length endoscopy videos from adult patients with UC undergoing colonoscopy or flexible sigmoidoscopy were analysed; isolated proctitis was excluded. Videos were segmented and annotated using Mayo Endoscopic Score (MES) and Ulcerative Colitis Endoscopic Index of Activity (UCEIS). A deep learning model was trained for frame-level quality control and severity prediction, enabling analysis of full-length videos. Performance was evaluated using quadratic weighted kappa (QWK) and Cohen's kappa, with patient-level separation between datasets. CDS was derived from UCEIS predictions to quantify cumulative inflammatory burden, spatial extent of disease and histological prediction. Results: A total of 67 videos from 59 patients were included. The model demonstrated agreement for remission classification (MES=0 {kappa} 0.76; UCEIS[≤]1 {kappa} 0.84). CDS enabled quantification of inflammatory burden and revealed spatial heterogeneity not reflected in categorical scores. Agreement with histology was strong (AUROC 0.84-0.87). Conclusions: AI-based analysis enables automated assessment of UC activity from full-length endoscopic videos, including remission detection and estimation of histological healing. CDS provides continuous characterisation of inflammatory burden and disease extent beyond conventional categorical scores, with potential to support more standardised assessment in clinical trials and practice.