International Session(Panel Discussion)1(JGES・JSGE・JSGS・JSGCS)
Sat. November 4th   14:00 - 17:00   Room 9: Portopia Hotel Main Building Kairaku 3
Deep neural network for video colonoscopy of ulcerative colitis
Kento Takenaka
Department of Gastroenterology and Hepatology, Tokyo Medical and Dental University
Endoscopic assessment of ulcerative colitis (UC) is important; however, individual clinician experiences can differ. Histological evaluation is also important, but specimen biopsies are necessary. We aimed to develop a deep neural network system based on endoscopy (DNUC) to achieve consistent, objective, and real-time evaluations.
We constructed the algorithm using 40758 images of colonoscopies and 6885 biopsy results from an endoscopic database. We then prospectively evaluate the accuracy of the DNUC and patient prognosis. Finally, we adopted our algorithm to full video-colonoscopy and evaluated its validity in a prospective multicenter study.
Regarding the prediction of histological remission from endoscopic images, the DNUC showed high diagnostic accuracy (92.9%). According to the Kaplan-Meier curve analysis, mucosal healing assessed using the DNUC was associated with a significantly lower risk for hospitalization and colectomy (p<0.001 for both). There were no significant differences in the accuracy of experts and the DNUC in predicting hospitalization and colectomy. Regarding video validation, DNUC was able to evaluate the presence or absence of histological inflammation in 81% of biopsy specimens, and had a sensitivity of 97.9% and a specificity of 94.6% for predicting histological remission. The intraclass correlation coefficient between DNUC and experts for endoscopic scoring was 0.927.
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.
Index Term 1: Ulcerative colitis
Index Term 2: Artificial inteligence
Page Top