UEGJournal Abstracts 2021 - 270

Oesophageal, gastric and duodenal disorders
P0017
PERFORMANCE OF A DEEP LEARNING ALGORITHM FOR AUTOMATIC
DETECTION OF BLOOD IN THE GASTRIC LUMEN USING CAPSULE
ENDOSCOPY IMAGES
T. Ribeiro1,2, M.J. Mascarenhas Saraiva1,2,3, J. Afonso1,2
J. Ferreira4,5, R. Gonçalves1,2
, P. Cardoso1,2
A.P. Andrade2,3,1
1
, M. Parente4,5, R. Natal Jorge4,5
,
, H. Cardoso1,2,3
,
, G. Macedo1,2,3
Centro Hospitalar Universitário de São João, Department of
Gastroenterology, Porto, Portugal, 2
WGO Gastroenterology and
Hepatology Training Center, Porto, Portugal, 3Faculty of Medicine
of the University of Porto, Porto, Portugal, 4
INEGI - Institute of Science and Innovation in
Mechanical and Industrial Engineering, Porto, Portugal
Contact E-Mail Address: tiagofcribeiro@outlook.com
Introduction: Upper gastrointestinal bleeding (UGB) is a frequent cause
of admission to emergency department. Esophagogastroduodenoscopy
(EGD) is the current gold standard for diagnosis, risk stratification and
treatment in patients presenting with UGB. Recent evidence has shown
that capsule endoscopy (CE) may facilitate risk stratification in these situations,
thus selecting patients who may benefit the most with prompt endoscopic
intervention (1).
The development of artificial intelligence (AI) systems for automatic analysis
of medical images has been the target of intense research over the last
decade. Convolutional Neural Networks (CNN) are a type of deep neural
architecture with high performance levels for image analysis. The application
of these tools for automatic identification of luminal blood in CE images
has provided auspicious results (2,3). Nevertheless, the application
of these algorithms for identification of blood in the gastric lumen has not
been evaluated.
Aims & Methods: We aimed to develop and test a deep learning algorithm
for automatic detection of blood or hematic residues in the gastric lumen.
A total of 3070 CE images were included for construction of the CNN, 2180
containing evidence of free blood or hematic residues within the gastric
lumen and 890 showing normal gastric mucosa. These images were inserted
into a CNN model with transfer learning. The full image dataset was
split for construction of training and validation datasets (80% and 20% of
all images, respectively). The performance of the network was assessed
using the validation dataset. The output provided by the CNN was compared
to a consensus classification provided by two endoscopists with experience
in CE. The performance of the network was evaluated by calculating
its sensitivity, specificity, accuracy, positive predictive and negative
predictive values (PPV and NPV, respectively), and area under the receiving
operating characteristic curve (AUC).
Results: The CNN was able to detect blood or hematic residues within the
gastric lumen with an overall accuracy of 97.4%. The deep learning system
had a sensitivity of 97.2%, a specificity of 97.8%, a PPV of 99.1% and a NPV
of 93.5%. The CNN analyzed the validation dataset in 12 seconds (average
image processing speed of 45 images per second).
Conclusion: The development of AI algorithms for application to CE may
allow to improve the diagnostic yield of this modality and expand its indications
in gastrointestinal bleeding beyond the evaluation for suspected
small bowel haemorrhage. Our group developed, for the first time, an automated
system for detection of blood within the gastric lumen.
References: 1. Gralnek IM, Ching JY, Maza I, et al. Capsule endoscopy in
acute upper gastrointestinal hemorrhage: a prospective cohort study. Endoscopy
2013;45:12-9.
2. Aoki T, Yamada A, Kato Y, et al. Automatic detection of blood content
in capsule endoscopy images based on a deep convolutional neural network.
J Gastroenterol Hepatol 2020;35:1196-1200.
270
Faculty of Engineering
of the University of Porto, Department of Mechanical Engineering,
Porto, Portugal, 5
P0018
THE UTILITY OF CT ANGIOGRAPHY IN IDENTIFICATION OF ACUTE
GASTROINTESTINAL BLEEDING SOURCE AND PREDICTORS OF A
POSITIVE EXAMINATION
W. Sbeit1
, A. Kadah1
, A. Shahin1
, T. Khoury1
1Galilee Medical Center, Gastroenterology, Naharia, Israel
Contact E-Mail Address: tawfikkhoury1@hotmail.com
Introduction: Acute Gastrointestinal Bleeding (GIB) constitutes a major
cause for hospital admissions around the world. Actually most bleeding
episodes stop spontaneously, but in cases of ongoing bleeding the mortality
rate approaches 14% for acute upper GIB and up to 3.9% for acute
lower GIB cases. Endoscopy is the first investigatory and therapeutic exam
used to localize and treat the bleeding source, however in cases where the
bleeding is obscure, computed tomography angiography (CTA) is used to
localize the bleeding source. According to American College of Radiology
(ACR) Criteria, the appropriateness of CT Angiography (CTA) and angiography
is comparable in cases of upper GIB with failure to localize the bleeding
source by endoscopy
Aims & Methods: To evaluate the utility of CTA in identification of acute
upper and lower GIB source and to try to identify predictors of a positive
CTA in these cases. We performed a cross sectional, retrospective study,
using the databases of Galilee Medical Center. All hospitalized patients
who were referred for CTA as part of acute GIB evaluation during a 10 years
period were included in the study. Univariate and multivariate logistic regression
was used to estimate odds ratio (OR) of baseline factors, backward
selection was applied to select the final most significant variables for
predicting positive CTA examinations.
Results: Overall, we included 154 patients with gastrointestinal bleeding
who underwent CTA examinations. Among them, 25 patients (16.2%) had
positive CTA as defined by extravasation of blood. On univariate analysis,
several parameters were found to be associated with a lower rate of positive
CTA, INR (OR 0.75, 95% CI 0.59-0.96, P=0.02), congestive heart failure
(OR 0.4, 95% CI 0.17-0.95, P=0.03), chronic coumadin use (OR 0.21, 95% CI
0.06-0.67, P=0.008), while albumin level (OR 2.68, 95% CI 1.25-5.73, P=0.01)
and American Society of Anesthesiologists (ASA) category 2 and 3 vs. 1 (OR
4.38, 95% CI 1.27-15.1, P=0.01) were risk factors for positive CTA. On the
other hand, there was no effect of blatchford and AIMS scores on the rate
of CTA positivity (OR 0.91, 95% CI 0.8-1.04, P=0.18 and OR 0.78, 95% CI
0.51-1.17, P=0.23), respectively. On multivariate logistic regression analysis,
only ASA score of 2 and 3 vs. 1 (OR 12.76, 95% CI 2.86-56.8, P=0.008)
and lower albumin (OR 3.36, 95% CI 1.42-7.95, P=0.005) remained a significant
risk factor for positive CTA.
Conclusion: Only one sixth of all CTA examination showed blood extravasation.
As only higher ASA categories and lower albumin level were the
only positive significant predictors for positive CTA. Further studies are
needed to validate our findings and to precisely define clinical and laboratory
predictors regarding the performance of CTA in this setting.
Disclosure: Nothing to disclose.
3. Ding Z, Shi H, Zhang H, et al. Gastroenterologist-Level Identification of
Small-Bowel Diseases and Normal Variants by Capsule Endoscopy Using a
Deep-Learning Model. Gastroenterology 2019;157:1044-1054.e5.
Disclosure: Nothing to disclose.
UEG Journal | Abstract Book
UEGJournal Abstracts 2021

Table of Contents for the Digital Edition of UEGJournal Abstracts 2021
