Automated assessment of the ulcerative colitis endoscopic index of severity using a multi-task deep learning model

Assessment of the Ulcerative Colitis Endoscopic Index of Severity (UCEIS) is limited by subjectivity and interobserver variability. We developed UC-MTLNet, a multi-task deep learning model to predict UCEIS descriptors, total score, endoscopic remission, and severity strata. This multicenter diagnostic study included 405 patients and 10,269 white-light endoscopic images from a development cohort, a prospectively collected internal test cohort, and an independent external test cohort. UC-MTLNet was evaluated at image and patient levels against a consensus reference standard. Descriptor-level QWK ranged from 0.9077 to 0.9337 in the internal test set and from 0.9142 to 0.9313 in the external test set. For total UCEIS, QWK was 0.9690 at image level and 0.9729 at patient level in the internal test set, with corresponding external values of 0.9646 and 0.9783. Image-level accuracies were 94.25%–95.66% for remission identification and 86.74%–89.21% for four-category severity stratification. Under standardized image-based scoring conditions, UC-MTLNet showed significantly higher exact-score accuracy and QWK than five experienced endoscopists across all settings. Compared with three single-task models, UC-MTLNet provided comparable performance with approximately one-third as many parameters and FLOPs and 2.78-fold faster inference. UC-MTLNet may assist standardized UCEIS assessment in research and central reading, pending prospective video-based and outcome-linked validation.