Endoscopy 2020; 52(S 01): S26
DOI: 10.1055/s-0040-1704084
ESGE Days 2020 oral presentations
Friday, April 24, 2020 11:00 – 13:00 Artificial Intelligence inGI-endoscopy:Is the future here? Wicklow Meeting Room 3
© Georg Thieme Verlag KG Stuttgart · New York

ARTIFICIAL INTELLIGENCE, TRAINED WITH A ROUGH BINARY CLASSIFICATION, CAN SELECT SIGNIFICANT IMAGES OF CAPSULE ENDOSCOPY

J Park
1   Soonchunhyang University College of Medicine, Digestive Disease Center, Institute for Digestive Research, Department of Internal Medicine, Seoul, Korea, Republic of
,
Y Hwang
2   Chungbuk National University, Department of Electronics Engineering, Chungju, Korea, Republic of
,
YJ Lim
3   Dongguk University Ilsan Hospital, Dongguk University College of Medicine, Division of Gastroenterology, Department of Internal Medicine, Goyang, Korea, Republic of
,
JH Nam
3   Dongguk University Ilsan Hospital, Dongguk University College of Medicine, Division of Gastroenterology, Department of Internal Medicine, Goyang, Korea, Republic of
,
DJ Oh
3   Dongguk University Ilsan Hospital, Dongguk University College of Medicine, Division of Gastroenterology, Department of Internal Medicine, Goyang, Korea, Republic of
,
KB Kim
4   Chungbuk National University College of Medicine, Department of Internal Medicine, Cheongju, Korea, Republic of
,
HJ Song
5   Jeju National University School of Medicine, Department of Internal Medicine, Jeju, Korea, Republic of
,
SH Kim
6   Seoul Metropolitan Government Seoul National University Boramae Medical Center, Department of Internal Medicine, Seoul, Korea, Republic of
,
MK Jung
7   Kyungpook National University Hospital, Division of Gastroenterology and Hepatology, Department of Internal Medicine, Daegu, Korea, Republic of
› Author Affiliations
Further Information

Publication History

Publication Date:
23 April 2020 (online)

Also available at
 

Aims Since the introduction of computer vision technology using Deep-learning, various acceptable results have been reported for the recognition of small bowel pathologies in capsule endoscopy. However, the results are limitedly dealt with lesions such as erosions, ulcers, and angioectasia, which are easy to apply machine learning technologies. We classified capsule endoscopy images into those with and without significant lesions, and studied whether artificial intelligence, which learned the images of binary classification, can correctly suggest images containing significant lesions.

Methods Seventy capsule endoscopy cases using PillCam SB3 (Medtronic, Minneapolis, MN, USA) were collected at 3 university hospitals. Under an agreement, two experienced endoscopist classified the extracted small bowel images into 2 categories according to the presence of pathologic features. Forty-eight thousand images containing inflamed mucosa, atypical vascularity and blood were categorized as significant images. Forty-eight thousand images representing normal mucosa were classified as insignificant. Normal contaminants like bile, bubble, and debris were allowed to be included in the insignificant images.

Sixty percent of total images (57,600) are used to train the recent Inception ResNet V2 model that has 467 layers. After pre-training on the ImageNet dataset, we retrained all the layers of model and obtained 97.98% accuracy on the validation set of images (20% of total, 19,200 frames).

Results Finally, we apply our trained model to a test set, 20 percent (19,200) of total images, that is not used for training and validation. The accuracy of testing is 98.13%.

Conclusions Artificial intelligence for accurate recognition of small intestine pathology requires highly classified and annotated learning materials. However, a well-organized large database can contribute to artificial intelligence for capsule endoscopy even with a rough classification.