Deep Neural Network, Endoscopy Images Predict Ulcerative Colitis Remission

The researchers sought to develop a deep neural network system for consistent, objective, and real-time analysis of endoscopic images from patients with ulcerative colitis.

According to a report published in Gastroenterology, researchers in Japan have developed a deep neural network algorithm that predicts remission in patients with ulcerative colitis (UC) with greater than 90% accuracy using real-time analysis of endoscopy images.

Kento Takenaka, PhD, of the department of gastroenterology and hepatology, Tokyo Medical and Dental University, Japan, and colleagues created the deep neural network for evaluation of UC (DNUC) algorithm based on a training set of 40,758 endoscopy images and biopsy results from patients with UC (n=2012) who underwent colonoscopy at a single center from January 2014 to March 2018.

The DNUC algorithm was then validated using data from a prospective cohort of patients with UC (n=875) who underwent colonoscopy from April 2018 to April 2019 (a total of 4187 endoscopy images and 4104 biopsy results). Remission was defined endoscopically as a UC endoscopic index of severity (UCEIS) score of zero and histologically as a Geboes score ≤3 points.

In the prospective portion of the study, the algorithm identified patients with endoscopic remission with 90.1% accuracy (95% CI, 89.2%-90.9%); the findings reported by endoscopists were used as a reference standard and Cohen’s kappa coefficient was 0.798 (95% CI, 0.78-0.814) between the DNUC and endoscopists’ findings. The DNUC findings and endoscopists’ UCEIS scores were strongly correlated (intraclass correlation coefficient 0.917; 95% CI, 0.911-0.921).

Related Articles

The algorithm identified patients in histologic remission with 92.9% accuracy (95% CI, 92.1%-93.7%). Cohen’s kappa coefficient was 0.859 (95% CI, 0.841-0.875) between the DNUC and biopsy findings.

Study limitations included potential selection bias because the endoscopy images and results for the training set were not controlled. The investigators also noted that using the most severe endoscopic findings according to the endoscopists’ judgment could create a selection bias during the algorithm’s training. Finally, variation in the deep learning methods and the details of the algorithm are not definitively understood, therefore, interpretation of results still require review by a physician.

“In conclusion, the accuracy of the DNUC was comparable to that of endoscopists in evaluating mucosal inflammation in patients with UC, and the DNUC could predict histologic remission without the need for obtaining a mucosal specimen,” wrote the authors. “The DNUC is a new, objective, consistent evaluation method that could be used in several medical situations.”

Follow @Gastro_Advisor


Takenaka K, Ohtsuka K, Fujii T, et al. Development and validation of a deep neural network for accurate evaluation of endoscopic images from patients with ulcerative colitis [published online February 11]. Gastroenterology. doi:10.1053/j.gastro.2020.02.012