Endoscopy 2021; 53(S 01): S58
DOI: 10.1055/s-0041-1724399
Abstracts | ESGE Days
ESGE Days 2021 Oral presentations
Friday, 26 March 2021 16:00 – 16:45 AI in the colon: Innovations and new developments Room 5

Preliminary Results of Artificial Intelligent Application in Colonoscopic Polyps Detection in Vietnam

H Dao
1   Hanoi Medical University, Internal Medicine Faculty, Hanoi, Vietnam
2   Hanoi Medical University Hospital, Endoscopic Centre, Hanoi, Vietnam
3   Institute of Gastroenterology and Hepatology, Hanoi, Vietnam
,
H Le
2   Hanoi Medical University Hospital, Endoscopic Centre, Hanoi, Vietnam
,
B Nguyen
3   Institute of Gastroenterology and Hepatology, Hanoi, Vietnam
,
H Nguyen
3   Institute of Gastroenterology and Hepatology, Hanoi, Vietnam
,
H Lam
3   Institute of Gastroenterology and Hepatology, Hanoi, Vietnam
,
L Dao
4   Institute of Gastroenterology and Hepatology, HANOI, Vietnam
,
T Nguyen
5   Vietnam National University of Agriculture, Informatics Technology Department, Hanoi, Vietnam
,
S Dinh
6   Hanoi University of Science and Technology, SOICT, Hanoi, Vietnam
,
H Vu
7   VDSENSE Joint Stock Company, Hanoi, Vietnam
› Author Affiliations
› Further Information
Also available at
 
 

    Aims The data from AI applications in colorectal polyps detection in developing countries like Vietnam is still lacking. This study aims to build a deep learning model for colorectal polyps detection in colonoscopy images and validate the accuracy in testing dataset.

    Methods The study was conducted from June 2019 to June 2020 in Institute of Gastroenterology and Hepatology and Hanoi Medical University hospital. Three datasets were collected, including a training dataset of 8186 colonoscopy images with at least one polyp and 4000 colonoscopy images without polyp; a validation dataset consisted of 1498 colonoscopy images with polyp(s) and 1000 colonoscopy images without polyps; and a testing dataset of 1321 colonoscopy images with 1549 polyps. The proposed model was built based on the U-Net architecture with an EfficientNet encoder and was trained in 150 epochs before testing on the testing dataset. The model’s accuracy was then evaluated using the F1 score, the positive predictive value (PPV) index, the sensitivity (Se) and specificity (Sp) index.

    Results The F1 scores of both the training set and the testing set were reported over 95 %. The PPV, Se and Sp scores of AI model on the validation in pixels were 94.6 %, 96.4 % and 99.8 %, respectively. The PPV, Se and Sp scores of AI model on the testing dataset were 94.4 %, 96.2 % and 95.4 %, respectively. In term of characteristics of polyps detected in the testing dataset, 63.58 % were less than 5mm in diameter, and 81.14 % were categorized as Is according to Paris classification. 59 polyps missed by the AI model were under 5mm in diameter and classified as Is. 88 cases classified as false positive were areas with mucosal folds (35.6 %), mucus (26.7 %), foam (13.3 %) or optical flares (6.7 %) in the images.

    Conclusions Application of AI in colonoscopy polyps detection is then feasible and requires further investigation in Vietnam.

    Citation: Dao H, Le H, Nguyen B et al. OP141 PRELIMINARY RESULTS OF ARTIFICIAL INTELLIGENT APPLICATION IN COLONOSCOPIC POLYPS DETECTION IN VIETNAM. Endoscopy 2021; 53: S58.


    #

    Publication History

    Article published online:
    19 March 2021

    © 2021. European Society of Gastrointestinal Endoscopy. All rights reserved.

    Georg Thieme Verlag KG
    Rüdigerstraße 14, 70469 Stuttgart, Germany