Artificial Intelligence for Quantifying Endoscopic Mucosal Ulceration in Crohn's Disease

Endoscopic scoring of Crohn's disease (CD) is challenging, as mucosal disease is patchy with highly variable morphology, size, and severity. Computer vision may help quantify disease activity with similar performance as standard instruments like the Simple Endoscopic Score for Crohn's Disease (SES-CD).

Colonoscopy videos from the STARDUST and SEAVUE phase 3 clinical trials underwent post-hoc computer vision endoscopic (CVE) assessment to quantify CD mucosal ulceration and injury. A segmentation model was trained on hand annotations of images performed by 2 gastroenterologists, predicting ulcer area, severity, and relative size. Using complete endoscopic video, predicted ulceration and general mucosal injury were then spatially mapped to the ileum and colon to quantify CD burden. CVE ulceration and general injury values were compared with the SES-CD in terms of disease quantification, localization, and agreement with end-of-study clinical remission (Crohn's Disease Activity Index [CDAI] <150).

Ulcer semantic segmentation models matched the performance of gastroenterologist annotators (Dice similarity coefficient, 0.591 vs 0.462), with neither performing better on qualitative review of disagreements. CVE measures were highly correlated with SES-CD scores (r = 0.73-0.85; P < .0001), although there was expected poor correlation with the degree of stenosis (r = 0.12-0.21). CVE ulcer measurements (28.4 vs 52.3; P = .0012) and SES-CD (6.3 vs 9.0; P = .0193) separated end-of-study clinical remission status, although CVE measures had greater effect size than SES-CD (g = 0.416 vs g = 0.290). Performance of CVE for CD was similar in the SEAVUE validation cohort.

CVE provides a means for automated ulceration and mucosal injury quantitation that shows conceptual agreement with SES-CD. CVE offers new capabilities to improve the granularity and personalization of endoscopic disease assessment in CD.