Ulcerative colitis is a chronic inflammatory disease in which treatment decisions depend on knowing how much inflammation remains in the colon. Endoscopic appearance is important, but histology provides additional information about microscopic disease activity and remission. The Nancy Histological Index grades this inflammation from N0 to N4. Obtaining it requires tissue biopsy during colonoscopy followed by microscopic assessment by trained pathologists.
Biopsy-based histology is clinically valuable, but it is invasive, samples only limited areas of the bowel, and is not available immediately during the procedure. Colonoscopy video, by contrast, is already acquired as part of routine care and captures the appearance of a much larger mucosal surface. The open question is whether those visual patterns contain enough information to estimate pathology-derived inflammation. This is difficult because histological and endoscopic severity are related but not identical measurements, particularly in intermediate disease.
Gastroenterologists need a reproducible estimate of inflammatory activity that uses the video already collected during colonoscopy and reflects the full inspected segment. Any useful system must preserve the ordered nature of Nancy scoring, distinguish severe disease without hiding failures in less common intermediate grades, and make clear that a video-derived estimate is not identical to microscopic histology. The clinically meaningful unit is the colon segment linked to the pathology sample, so predictions should combine information across multiple frames and be evaluated with patient-level separation to avoid artificially optimistic results from repeated views of the same patient.
We are studying whether pathology-derived Nancy inflammation scores can be estimated from colonoscopy video already collected during routine care. Predictions are made for the colon segment linked to the pathology sample, combining information across many frames of the segment, using a retrospective single-centre cohort split at the patient level.
The model achieved quadratic weighted kappa of 0.601±0.023 and macro-F1 of 0.344±0.025. Performance was strongest for Nancy 4 (F1=0.719) and weakest for Nancy 2 (F1=0.019), indicating useful ordinal signal but insufficient middle-grade discrimination for standalone clinical use.