Endoscopy 2022; 54(S 01): S140-S141
DOI: 10.1055/s-0042-1744936
Abstracts | ESGE Days 2022
ESGE Days 2022 Digital poster exhibition

EVALUATION OF ULCERATIVE COLITIS ENDOSCOPIC MAYO SCORE WITH ARTIFICIAL INTELLIGENCE

H.T. Kani
1   Marmara University, School of Medicine, Department of Gastroenterology, Istanbul, Turkey
,
I. Ergenc
1   Marmara University, School of Medicine, Department of Gastroenterology, Istanbul, Turkey
,
G. Polat
2   Middle East Technical University, Institute of Informatics, Ankara, Turkey
,
Y. Ozen Alahdab
1   Marmara University, School of Medicine, Department of Gastroenterology, Istanbul, Turkey
,
A. Temizel
2   Middle East Technical University, Institute of Informatics, Ankara, Turkey
,
O. Atug
1   Marmara University, School of Medicine, Department of Gastroenterology, Istanbul, Turkey
› Author Affiliations
 
 

    Aims Multilayer artificial neural networks are artificial intelligence (AI) algorithms with high predictive power that allow processing large volumes of data sets. Ulcerative colitis (UC) Endoscopic Mayo Score (EMS) is a subjective assessment that varies between the endoscopists (1, 2). Our aim was to develop an AI algorithm to evaluate endoscopist-independent EMS with high accuracy and minimize the subjectivity.

    Methods We enrolled the images of UC patients who were evaluated with colonoscopy between December 2011 and July 2019. EMS evaluation was performed individually and blindly for each image by three different experienced gastroenterologists. Artificial intelligence algorithm developed in Python programming language and by using the PyTorch library. Seventy percent of the data set was defined as training set, 15% as the validation set, and 15% as test set. Artificial intelligence was trained by ResNet152 model.

    Results A total of 19537 images were evaluated. Images with artifact, terminal ileum and ileo-anal pouch images were excluded. A total of 11276 images were included to the data set [EMS 0: 6105 (54.1%), EMS 1: 3052 (27.1%), EMS 2: 1254 (9.9%) and EMS 3: 865 (7.7%)]. Success rate was 79.2% for differentiation of each EMS classes (Mayo 0,1,2,3) (specificity: 0.92; sensitivity: 0.73; AUC: 0.94). Success rate was 87.25% for the evaluation of remission (Mayo 0 vs Mayo 1,2,3) and success rate was 95.9% for the differentiation of severe disease (Mayo-0.1 vs Mayo-2,3) ([Fig.1]).

    Zoom Image
    Fig. 1

    Conclusions Our artificial intelligence algorithm developed with the ResNet152 model evaluated the EMS with a very high accuracy and 100% consistency from endoscopic images.


    #

    Publication History

    Article published online:
    14 April 2022

    © 2022. European Society of Gastrointestinal Endoscopy. All rights reserved.

    Georg Thieme Verlag KG
    Rüdigerstraße 14, 70469 Stuttgart, Germany

     
    Zoom Image
    Fig. 1