Deep learning layer-specific collagen quantification correlates with activity and is associated with outcomes in...

Intestinal fibrosis remains a major unmet need in inflammatory bowel disease (IBD). Although collagen deposition is the defining feature of this process, its quantification across intestinal layers has relied on semiquantitative, low-throughput methods that provide limited clinical insight. We developed and validated the first machine learning-based pipeline for layer-specific collagen quantification in intestinal biopsies and assessed its association with disease activity and clinical outcomes.

In this retrospective, single-center study, 1317 intestinal biopsies from 191 IBD patients (92 Crohn's disease [CD], 99 ulcerative colitis [UC]) and 74 controls were analyzed. Sirius Red-stained whole-slide images were segmented using a deep learning model trained on manual annotations to delineate mucosa, muscularis mucosa (MM), and submucosa. Collagen proportional area (CPA) was calculated within each layer using K-means clustering. Associations with clinical, endoscopic, and histological parameters were examined, and the prognostic value of collagen burden for IBD-related outcomes was explored.

Machine learning-based collagen quantification showed good reproducibility across layers (intraclass correlation coefficient [ICC] 0.72-0.75). MM and submucosal CPA were significantly higher in active CD than in remission or controls, correlating with disease activity indices and inflammation markers. Elevated baseline submucosal CPA was independently associated with IBD-related surgery (adjusted hazard ratio [aHR] = 1.024; P = .043), while MM CPA was independently associated with hospitalization during follow-up (adjusted odds ratio [aOR] = 1.044; P = .008). In UC, submucosal CPA was higher in remission versus controls. In paired biopsies, bio-experienced patients showed greater reductions in MM and submucosal CPA than bio-naive patients.

Layer-specific collagen burden reflects disease activity and is associated with adverse outcomes in CD, supporting its potential as a biomarker for risk stratification and therapeutic monitoring.