Endoscopy 2020; 52(S 01): S109
DOI: 10.1055/s-0040-1704336
ESGE Days 2020 oral presentations
Friday, April 24, 2020 17:00 – 18:30 ERCP: Ductal access The Liffey B
© Georg Thieme Verlag KG Stuttgart · New York

IDENTIFYING AMPULLA AND DIFFICULTY OF SELECTIVE CANNULATION WITH ARTIFICIAL-INTELLIGENCE ASSISTED ANALYSIS OF ERCP IMAGE

KW Lee
1   Korea University Anam Hospital, Division of Gastroenterology and Hepatology, Department of Internal Medicine, Seoul, Korea, Republic of
,
JM Lee
1   Korea University Anam Hospital, Division of Gastroenterology and Hepatology, Department of Internal Medicine, Seoul, Korea, Republic of
,
HJ Jeon
1   Korea University Anam Hospital, Division of Gastroenterology and Hepatology, Department of Internal Medicine, Seoul, Korea, Republic of
,
SH Kim
1   Korea University Anam Hospital, Division of Gastroenterology and Hepatology, Department of Internal Medicine, Seoul, Korea, Republic of
,
S Jang
1   Korea University Anam Hospital, Division of Gastroenterology and Hepatology, Department of Internal Medicine, Seoul, Korea, Republic of
,
SJ Choi
1   Korea University Anam Hospital, Division of Gastroenterology and Hepatology, Department of Internal Medicine, Seoul, Korea, Republic of
,
HS Choi
1   Korea University Anam Hospital, Division of Gastroenterology and Hepatology, Department of Internal Medicine, Seoul, Korea, Republic of
,
ES Kim
1   Korea University Anam Hospital, Division of Gastroenterology and Hepatology, Department of Internal Medicine, Seoul, Korea, Republic of
,
B Keum
1   Korea University Anam Hospital, Division of Gastroenterology and Hepatology, Department of Internal Medicine, Seoul, Korea, Republic of
,
YT Jeen
1   Korea University Anam Hospital, Division of Gastroenterology and Hepatology, Department of Internal Medicine, Seoul, Korea, Republic of
,
HS Lee
1   Korea University Anam Hospital, Division of Gastroenterology and Hepatology, Department of Internal Medicine, Seoul, Korea, Republic of
,
HJ Chun
1   Korea University Anam Hospital, Division of Gastroenterology and Hepatology, Department of Internal Medicine, Seoul, Korea, Republic of
,
CD Kim
1   Korea University Anam Hospital, Division of Gastroenterology and Hepatology, Department of Internal Medicine, Seoul, Korea, Republic of
,
J Kim
2   Korea University College of Informatics, Department of Computer Science and Engineering, Seoul, Korea, Republic of
,
T Kim
2   Korea University College of Informatics, Department of Computer Science and Engineering, Seoul, Korea, Republic of
,
J Choo
2   Korea University College of Informatics, Department of Computer Science and Engineering, Seoul, Korea, Republic of
,
SY Han
3   Pusan National University College of Medicine, Department of Internal Medicine, Pusan, Korea, Republic of
,
DU Kim
3   Pusan National University College of Medicine, Department of Internal Medicine, Pusan, Korea, Republic of
,
S Kwon
4   Catholic University of Daegu, Department of Anatomy, Daegu, Korea, Republic of
› Author Affiliations
Further Information

Publication History

Publication Date:
23 April 2020 (online)

 
 

    Aims The advancement of artificial intelligence (AI) made it possible to apply AI into medical fields. The aim of this study was to investigate an AI-assisted classification for ERCP through convolutional neural network (CNN) using documented endoscopic images.

    Methods We used deep convolutional neural networks pre-trained on the ImageNet dataset such as ResNet18, ResNet50, and VGG19 and fine-tuned them on dataset. We performed 5-fold cross-validation on our dataset and formed subtasks, four-class classification and binary classification, according to the cannulation difficulty.

    Results ERCP Data of 456 patients were included in the analysis. The averaged training results of 5-fold cross-validation for the detection task were as follows: mean intersection over union (mIoU) (0.544 ± 0.021), mean absolute error (MAE) (6.360 ± 0.522), and root mean squared error (10.009 ± 1.390). Fig. 1 shows the success plot of detection task. Each point on the plot shows the mean success rates of entire folds for each threshold. Fig. 2 shows the comparison of the estimated bounding box and the ground truth bounding box on a sample image. In the difficulty prediction task, we achieved the accuracy of 69.1 % on the four-class classification and that of 67.1 % on the binary classification.

    Conclusions AI-assisted system mostly differentiated the ampulla and has a potential to improve the quality of ERCP.


    #