Deep neural network for video colonoscopy of ulcerative colitis: a cross-sectional study

Lancet Gastroenterol Hepatol. 2022 Mar;7(3):230-237. doi: 10.1016/S2468-1253(21)00372-1. Epub 2021 Nov 29.

Abstract

Background: A combination of endoscopic and histological evaluation is important in the management of patients with ulcerative colitis. We aimed to adapt our previous deep neural network system (deep neural ulcerative colitis [DNUC]) to full video colonoscopy and evaluate its validity in the real-time detection of histological mucosal inflammation.

Methods: In this multicentre, cross-sectional study, we prospectively enrolled consecutive patients (≥15 years) with ulcerative colitis who had an indication for colonoscopy at five hospitals in Japan. Patients in clinical remission were randomly assigned (1:2) to study 1 and study 2. Those with clinically active disease were assigned to study 2 only. Study 1 assessed the validity of real-time histological assessment using DNUC and study 2 validated the consistency of endoscopic scoring between DNUC and experts. The primary endpoint for study 1 was comparison of the results judged by DNUC (healing or active) with biopsy specimens evaluated by pathologists. In study 2, the primary endpoint was the ability of DNUC to determine the Ulcerative Colitis Endoscopic Index of Severity score compared with centrally evaluated scoring by inflammatory bowel disease endoscopy experts.

Findings: From April 1, 2020, to March 31, 2021, 770 patients (180 in study 1 and 590 in study 2) were enrolled. Using real-time histological evaluation, DNUC was able to evaluate the presence or absence of histological inflammation in 729 (81%) of 900 biopsy specimens. For predicting histological remission, the DNUC had a sensitivity of 97·9% (95% CI 97·0-98·5) and a specificity of 94·6% (91·1-96·9). Moreover, its positive predictive value was 98·6% (97·7-99·2) and negative predictive value was 92·1% (88·7-94·3). The intraclass correlation coefficient between DNUC and experts for endoscopic scoring was 0·927 (95% CI 0·915-0·938).

Interpretation: DNUC provided consistently accurate endoscopic scoring and showed potential for reducing the number of biopsies required. This system is an objective and consistent application for video colonoscopy that has potential for use in various medical situations.

Funding: Tokyo Medical and Dental University and Sony.

Publication types

  • Multicenter Study
  • Research Support, Non-U.S. Gov't

MeSH terms

  • Adult
  • Biopsy
  • Colitis, Ulcerative / pathology*
  • Colonoscopy*
  • Cross-Sectional Studies
  • Female
  • Humans
  • Intestinal Mucosa / pathology
  • Male
  • Middle Aged
  • Neural Networks, Computer*
  • Prospective Studies
  • Random Allocation
  • Remission Induction
  • Sensitivity and Specificity
  • Video Recording*