Digestive and Liver Disease 54 (2022) 1513–1519

Contents lists available at ScienceDirect

Digestive and Liver Disease

journal homepage: www.elsevier.com/locate/dld

Alimentary Tract

A deep learning method to assist with chronic atrophic gastritis

diagnosis using white light images
Ju Luo a, Suo Cao a, Ning Ding b, Xin Liao c, Lin Peng c, Canxia Xu a,∗
a
Gastroenterology department, Third Xiangya Hospital, Central South University, China
b
Emergency department, The Affiliated Changsha Central Hospital, Hengyang medical School, University of South China, China
c
College of Computer Science and Electronic Engineering, Hunan University, China

a r t i c l e i n f o a b s t r a c t

Article history: Background: Chronic atrophic gastritis is a common preneoplastic condition of the stomach with a low
Received 20 January 2022 detection rate during endoscopy.
Accepted 20 April 2022
Aims: This study aimed to develop two deep learning models to improve the diagnostic rate.
Available online 21 May 2022
Methods: We collected 10,593 images from 4005 patients including 2280 patients with chronic atrophic
Keywords: gastritis and 1725 patients with chronic non-atrophic gastritis from two tertiary hospitals. Two deep
Artificial intelligence learning models were developed to detect chronic atrophic gastritis using ResNet50. The detection ability
Chronic atrophic gastritis of the deep learning model was compared with that of three expert endoscopists.
Precancerous lesions Results: In the external test set, the diagnostic accuracy of model 1 for detecting gastric antrum atrophy
Auxiliary diagnostic model was 0.890. The identification accuracies for the severity of gastric antrum atrophy were 0.773 and 0.590
in the internal and external test sets, respectively. In the other two external sets, the detection accuracies
of model 2 for chronic atrophic gastritis were 0.854 and 0.916, respectively. Deep learning model 1 s
ability to identify gastric antrum atrophy was comparable to that of human experts.
Conclusion: Deep-learning-based models can detect chronic atrophic gastritis with good performance,
which may greatly reduce the burden on endoscopists, relieve patient suffering, and improve the disease’s
detection rate in primary hospitals.
© 2022 Editrice Gastroenterologica Italiana S.r.l. Published by Elsevier Ltd. All rights reserved.

1. Introduction procedure has several limitations. First, making a diagnosis using

standard endoscopy alone (white light endoscopy) is reliant on
Chronic atrophic gastritis (CAG) is a common gastric disease highly specialized doctors; second, the false-negative rate of biopsy
that has been recognized as a precancerous lesion [1]. The popula- is high; and third, some patients have contraindications for biopsy.
tion prevalence of CAG generally ranged from 2.1% to 8.2% world- A multicenter survey in China showed that the biopsy sensitivity
wide. In China, the prevalence of CAG was 55.7 per 10 0 0 patients of pathologic atrophic gastritis was only 42% [5]. Although multi-
with gastric cancer [2]. Patients with CAG have an increased risk of point biopsy could increase the positivity rate, it would also likely
gastric adenocarcinoma [3]. Increased severity of atrophy and ex- increase the risk of hemorrhage and other complications.
tent of intestinal metaplasia are associated with an increased risk Artificial intelligence (AI) has profoundly transformed the way
of cancer [3,4]. Therefore, early detection of CAG and the evalua- we live our lives, and this has been especially evident in the field
tion of the severity of atrophy would be helpful in identifying pop- of medical science. In recent years, with the accumulation of abun-
ulations at high risk for gastric cancer. dant medical data and improvement in computer algorithms, AI
Gastric atrophy and intestinal metaplasia are usually diagnosed development has made great strides in assisting diagnoses and
by endoscopic examination and subsequent biopsy. However, the predicting prognosis – for example, AI-based cardiovascular disease
risk identification systems and diabetic retinopathy detection sys-
tems have been created [6,7].
Abbreviations: CAG, Chronic Atrophic Gastritis; CNAG, Chronic Non-Atrophic
Deep learning (DL), a branch of AI, allows computational mod-
Gastritis; AI, Artificial Intelligence; DL, Deep Learning; CNN, Convolutional Neural
Network; ADR, Adenoma Detection Rate; OLGA, Operative Link on Gastritis Assess- els composed of multiple processing layers to learn representations
ment. of data with multiple levels of abstraction. Convolutional Neural
∗
Corresponding author. Network (CNN) is the most important basic unit in the field of
E-mail addresses: [email protected] (J. Luo), xucanxia20 0 [email protected] DL. The typical architecture of a CNN usually includes a convolu-
(C. Xu).

https://doi.org/10.1016/j.dld.2022.04.025
1590-8658/© 2022 Editrice Gastroenterologica Italiana S.r.l. Published by Elsevier Ltd. All rights reserved.
J. Luo, S. Cao, N. Ding et al. Digestive and Liver Disease 54 (2022) 1513–1519

tional layer, pooling layer, rectified linear unit (ReLU) layer, and a matched images. Images were excluded if they were of poor qual-
fully connected layer. When images and video files are input into a ity or depicted CAG or CNAG with other lesions, such as ulcers and
CNN, the CNN can extract useful image features automatically and cancerous lesions.
exploit the softmax function (also known as multinomial logistic Images were collected and classified according to the hospital
regression) for classification [9]. in which the procedure took place and the location, atrophy, and
These methods have dramatically improved the state-of-the- severity of the atrophy. A total of 3885 images were collected from
art speech recognition, visual object recognition, object detection, 2280 CAG patients, including 2931 images of the antrum, 590 im-
and many other domains, such as drug discovery and genomics. ages of the angle, 309 images of the corpus, 45 images of the fun-
DL discovers intricate structures in large datasets using the back- dus, 1825 images of mild atrophic gastritis, 1902 images of moder-
propagation algorithm in order to dictate how a machine should ate atrophic gastritis, and 38 images of severe atrophic gastritis. A
change its internal parameters. These parameters compute the rep- total of 6708 images were collected from 1725 CNAG patients, in-
resentation in each layer from the representation in the previ- cluding 1650 images of the antrum, 1595 images of the angle, 1729
ous layer. DL have brought about breakthroughs in image, video, images of the corpus, and 1734 images of the fundus. Information
speech, and audio processing [8]. about the gastroscopic image dataset is shown in Supplementary
Image and video data, such as CT, MRI and endoscopic examina- Table 1.
tions, are crucial for diagnosing digestive system diseases. Hence, All gastroscopic images were obtained by doctors during
the application of DL has significant prospects for development. their daily clinical operations. Four different devices were used:
Progress has already been made in the study of digestive system Olympus Evis Lucera 260/290 (Tokyo, Japan) and FUJIFILM EG-
diseases. DL-based methods can detect not only early esophageal 530WR/601 WR (Tokyo, Japan). The gastric mucosa images were
cancer and early gastric cancer but also H. pylori infections obtained using the white light model; all images were in JPG for-
[10–12]. By utilizing colonoscopy images, a DL-based computer- mat with ordinary resolution, and the size of each image was 50–
assisted diagnosis system could help doctors significantly increase 300 kb.
the adenoma detection rate (ADR) [13,14]. The DL model can detect In order to identify gastric antrum atrophy and its severity, we
and grade ulcers in Crohn’s disease and ulcerative colitis based on established two datasets to develop model 1. Dataset 1 contained
capsule endoscopy images and videos [15–17]. However, there are all antrum images of CAG and CNAG from the Third Xiangya Hos-
few reports of DL being used for CAG detection. Thus, our study pital. Three hundred antrum images were randomly selected from
aimed to use the DL method to improve the diagnostic rate of CAG the Changsha Central Hospital as dataset 2. The details of these
using traditional white light images. two datasets are summarized in Supplementary Table 2.
In order to identify CAG, we established three datasets for de-
2. Materials and methods veloping model 2. Dataset 3 contained all CAG and CNAG images
from the Third Xiangya Hospital. We employed several data aug-
This study conformed to the ethical guidelines of the Declara- mentation methods to increase the size of dataset 3, including flip
tion of Helsinki and was approved by the ethics committee of the and 90° rotation. Two hundred CAG and CANG images were ran-
Third Xiangya Hospital of Central South University (No.21160). Ac- domly selected from Changsha Central Hospital as dataset 4. A to-
cording to national legislation and institutional requirements, in- tal of 132 images without a gastric antrum were randomly selected
formed consent was waived by the Ethics Committee of the Third from Changsha Central Hospital as Dataset 5. The details of these
Xiangya Hospital of Central South University and Changsha Central three datasets are summarized in Supplementary Table 3.
Hospital due to the retrospective nature of this study. The relevant codes and models can be freely accessed at https:
//github.com/philiplaw1984/chronic-atrophic-gastritis/.
2.1. Datasets and preprocessing
2.2. Deep learning networks
This retrospective study was conducted at the Gastroenterology
Department of the Third Xiangya Hospital of Central South Univer- Our study tested the current mainstream architectures, namely
sity and Changsha Central Hospital. Endoscopic images of patients Vgg-16, ResNet-50, DenseNet169, and Inception_V3 networks.
with CAG and chronic non-atrophic gastritis (CNAG) were collected Resnet-50 showed the best performance for CAG identification.
between January 2015 and December 2020. We collected 10,593 Therefore, we selected Resnet-50 as the basic network to develop
images of 4005 patients, including 2280 patients with CAG and models 1 and 2.
1725 patients with CNAG, from two hospitals. All CAG cases were Resnet-50 is a classic CNN. The “50’’ in the name ResNet-50
confirmed by pathological examination. First, the biopsies were refers to an architecture with 7 × 7 convolutional layers and 16
scored semi-quantitatively by two pathologists with >10 years of building blocks (each building block includes 3 convolutional lay-
experience each. The scoring was conducted according to the up- ers), forming a total of 48 convolutional layers and a fully con-
dated Sydney classification system and the Operative Link on Gas- nected layer. One of the most important features of ResNet-50 is
tritis Assessment (OLGA) method, which combines the degree and the shortcut connections, which skip one or more layers to solve
range of intestinal metaplasia and gastric mucosa. Second, the de- the vanishing gradient problem in deep neural networks by allow-
gree of atrophy was calculated on the basis of the degree of gland ing the gradient to flow through the layer. Fig. 1 shows the struc-
reduction: mild atrophy - the number of glands was reduced by ture of the ResNet-50.
less than 1/3, moderate atrophy - the number of glands was re-
duced between 1/3 and 2/3, severe atrophy - the number of glands 2.3. Experimental settings and evaluation values
was reduced by more than 2/3, with only a few remaining glands
or their complete disappearance. Cases were excluded when the All experiments using deep learning for model training were
two pathologists did not reach an agreement. performed on matpool, a GPU cloud platform, with an NVIDIA
All gastroscopic reports of CAG and CNAG were reviewed by GeForce RTX 3090 GPU. The experimental environment was Python
two experts with more than 10 years of experience in gastroscopy 3.8, CUDA 11.0, cuDNN 8.0, TensorFlow 2.4, Keras2.3.1, NVCC, and
each. All reports included a series of endoscopic images and text Ubuntu 18.04. The batch size was 64 and the number of epochs
descriptions. We first reviewed the text descriptions to deter- was 500. The main evaluation values were accuracy, sensitivity,
mine the location of the performed biopsy and then extracted the and specificity.

1514
J. Luo, S. Cao, N. Ding et al. Digestive and Liver Disease 54 (2022) 1513–1519

Fig. 1. The simple structure of the ResNet-50.

1a - the structure of ResNet-50;
1b - the structure of ID BLOCK;
1c - the structure of CONV BLOCK.

2.4. AI vs doctors were 0.859, 0.875, and 0.854, respectively. To test the robustness
and universality, we tested the performance of model 2 on dataset
To verify the performance of the DL model, we designed an 4, an external test set; the accuracy, sensitivity, and specificity for
AI-doctors comparison experiment. Fifty images were randomly CAG were 0.854, 0.870, and 0.850, respectively. We also tested the
selected from the test-set of datasets 1 and 2 as the test set performance of model 2 in dataset 5, an external test set contain-
of AI-doctors competition. We invited three gastroenterologists to ing all the images of all gastric segments apart from the antrum.
counter DL model 1. All three experts had more than five years of Accuracy, sensitivity, and specificity were 0.916, 0.912, and 0.920,
experience in gastroscopy. We then compared the performances of respectively. The detailed results of the two external test sets are
the three experts with the performance of the AI on this small test shown in Fig. 3.
set.
3.3. AI vs doctors
3. Result
The AI-doctors competition experiment showed that the abil-
3.1. Model 1 recognized gastric antrum atrophy and its severity ity of our DL model to identify gastric antrum atrophy was almost
equal to those of the doctors. The results of the three experts and
In the test set of dataset 1, the diagnostic accuracy, sensitiv- our DL model 1 of the AI-doctors competition test set are shown
ity, and specificity of model 1 for gastric antrum atrophy were in Figs. 4 and 5.
0.902, 0.891, and 0.915, respectively. The accuracy of identifying
the severity of gastric antrum atrophy was 0.773. To test the ro- 4. Discussion
bustness and universality, we tested the performance of model 1 in
dataset 2, an external test set. Accuracy, sensitivity, and specificity Research on deep learning and chronic atrophic gastritis is rela-
for gastric antrum atrophy were 0.890, 0.905, and 0.890, respec- tively scarce. With the help of the DL model based on DenseNet,
tively. The accuracy of identifying the severity of gastric antrum CAG can be diagnosed using high-resolution gastric atrophy im-
atrophy was 0.590 in the external test set. The details of the results ages [18]. A small study from Germany indicated that DL could
are shown in Fig. 2. The detection rates of CNAG and mild, moder- diagnose atrophic gastritis with high accuracy [19]. Using gastric
ate, and severe antrum atrophy in the internal test set were 91.5, X-ray images, a Japanese research team developed a DL model for
87.9, 90.4, and 91.7%, respectively. The detection rates of CNAG and automatic CAG detection [20]. However, the previous studies also
mild, moderate, and severe antrum atrophy in the external test set have several limitations. First, to achieve high accuracy, studies are
were 89.0, 93.0, 87.8, and 90.0%, respectively. prone to using high-resolution images or magnifying narrow-band
images. Only a few studies have focused on ordinary white light
3.2. Model 2 recognized chronic atrophic gastritis images, which are the images most widely used in clinical practice.
Second, current studies either focused on the gastric antrum or
Our results showed that model 1 could recognize gastric on the fundus and corpus. Few studies have focused on all stom-
antrum atrophy and its severity with a first-rate performance. ach regions where chronic atrophic gastritis occurs. Third, most of
However, chronic atrophic gastritis is not only limited to the the published studies were single-center ones. Due to limited data,
antrum but also occurs in other locations in the stomach, including many studies lacked external testing with which to evaluate the
the angle, corpus, and fundus. To overcome this limitation, we de- performance of the models.
veloped model 2 to recognize CAG. In the test set of dataset 3, the To overcome these limitations, we first collected abundant or-
diagnostic accuracy, sensitivity, and specificity of model 2 for CAG dinary white light images of the antrum, angle, corpus, and fun-

1515
J. Luo, S. Cao, N. Ding et al. Digestive and Liver Disease 54 (2022) 1513–1519

Fig. 2. The details of model 1 recognize gastric antrum atrophy and its severity.

Fig. 3. The details of model 2 recognize chronic atrophic gastritis.

1516
J. Luo, S. Cao, N. Ding et al. Digestive and Liver Disease 54 (2022) 1513–1519

Fig. 4. The details of three experts and deep learning model 1 on the artificial intellingence-doctors competition test-set.

Fig. 5. The artificial intelligence-doctors comparison results.

dus from two different medical centers. Subsequently, we estab- sistance of the DL model, we could detect moderate and severe
lished five independent datasets for developing and testing two DL atrophic gastritis as early as possible. Moreover, if the follow-up
models and one small dataset for AI-doctors competition. Then, we biopsy confirmed that the patient had intestinal metaplasia, we
developed two DL models, one for recognizing gastric antrum at- could identify the high-risk gastric cancer population earlier. If the
rophy and its severity and another for recognizing CAG. Both DL DL model predicts mild gastric atrophy, biopsy might be avoided.
models could detect CAG with good performance. Finally, we ver- However, CAG not only occurs in the gastric antrum but also in
ified that the ability of our DL model to identify gastric antrum other locations, including the angle, corpus, and fundus. Patients
atrophy was similar to that of trained endoscopists. with CAG should be followed up with high quality endoscopy,
Helicobacter pylori causes CAG with predominant localization meaning that doctors should carefully examine all gastric locations
in the gastric antrum [21]. Many lines of evidence indicate that ad- [27]. DL model 2 could detect CAG in any region of the stomach
vanced atrophy is a risk factor for gastric cancer [22,23]. Therefore, with excellent performance. This may greatly reduce the burden
grading atrophic gastritis is useful for the early detection and treat- on doctors and improve the CAG detection rate.
ment of gastric cancer. Our study developed a DL model with good Conventional white light endoscopy has moderate sensitiv-
performance for the detection and grading of gastric antrum atro- ity and specificity, as well as high inter-observer variability, and
phy. China has a high prevalence of Helicobacter pylori infection, is therefore not sufficient for reliably diagnosing gastric atrophy
resulting in a high incidence of gastric atrophy, gastric intestinal [27,28]. While advanced endoscopic techniques such as chromoen-
metaplasia, and gastric cancer [24,25]. For this reason, the Chinese doscopy, magnification endoscopy, and confocal laser endomi-
Society of Gastroenterology has suggested that doctors should fol- croscopy could increase the detection rate, primary hospitals are
low OLGA in clinical practice. The Chinese Society of Gastroenterol- often hindered by technical availability and high costs. However,
ogy also recommends that patients with moderate and severe gas- using only conventional white light images, our DL models could
tric atrophy with gastric intestinal metaplasia should be followed detect CAG with outstanding sensitivity and specificity. Moreover,
annually [26]. Our initial assumption was that by including the as- the computer hardware requirements of our DL models were very

1517
J. Luo, S. Cao, N. Ding et al. Digestive and Liver Disease 54 (2022) 1513–1519

cost-effective since the input images were small and had low res- Changsha Central Hospital due to the retrospective nature of this
olutions. With their good performance and relatively low cost, the study.
DL models we developed may be more suitable for primary hospi-
tals. Consent for publication
Our DL model performed equally well in detecting CAG in two
different hospitals. The results indicate that our model may be eas- Not applicable.
ily transferred from one hospital to another. While the accuracy
of grading atrophic gastritis was decreased, this may largely be Funding
because endoscopy images from different hospitals were acquired
by different devices and operators. It is unrealistic to expect ev- Not applicable.
ery hospital, especially primary hospitals, to invent their own DL
model. A possible solution to this issue is to use our pre-trained
Author contributions
DL model and add small-sized images to train their own models.
It is our future work to prospectively validate our methods and
Ju Luo and Suo Cao wrote the manuscript’s main text. Canxia
help other hospitals build their own models with our pre-trained
Xu and Xin Liao contributed to the conception and design of the
model.
study. Ju Luo and Lin Peng analyzed the data and developed the
In this study, ResNet-50 was selected as the framework since
deep learning model. Ning Ding created the tables and figures.
we required a deep network to extract the hidden features from
Canxia Xu and Lin Peng revised some of the chapters with con-
endoscopic images, which are more challenging than other im-
structive comments. All authors have approved the final draft of
ages. Macroscopically, as atrophy in CAG progresses, the gastric
the manuscript and have agreed to its submission for publication.
folds disappear. This loss of gastric rugae, combined with mucosal
pallor and increased visibility of mucosal vessels, constitutes the
Acknowledgment
main endoscopic features of atrophic gastritis [29,30]. On white
light endoscopy, intestinal metaplasia typically appears as small,
We thank Elsevier for their help in the English language editing
gray-white, elevated plaques surrounded by a mix of patchy pink
of this manuscript. We also thank Wei Chen, jiangfang Chui and
and pale areas of mucosa, causing an irregular surface appearance
Chaiwei He for their help in the artificial Intelligence-doctors com-
[31]. ResNet-50 is suitable for extracting endoscopic findings of this
petition test-set.
complexity, since it can detect multiple features including texture,
shape, and color [32].
Our study has several limitations. First, the accuracy of DL Supplementary materials
model 1 in grading atrophic gastritis, especially in the external test
set, was unsatisfactory. To improve the performance of DL model, Supplementary material associated with this article can be
more training data from different medical centers and a more ap- found, in the online version, at doi:10.1016/j.dld.2022.04.025.
propriate deep learning network architecture are required. Second,
References
due to the lack of sufficient training data, DL model 2 was unable
to grade the atrophic severity of the angle, corpus, and fundus. Fur- [1] Hatakeyama M. Helicobacter pylori CagA and gastric cancer: a paradigm for hit
ther studies are needed to collect more atrophic gastritis images of and-run carcinogenesis. Cell Host Microbe 2014;15:306–16.
these segments, especially images of severe atrophy. Finally, our re- [2] Wang R, Chen XZ. Prevalence of atrophic gastritis in southwest China and pre-
dictive strength of serum gastrin-17: a cross-sectional study (SIGES). Sci Rep
search was a retrospective study, and more prospective studies are
2020;10:4523.
needed to verify the diagnostic ability of the DL model. [3] Banks M, Graham D, Jansen M, et al. British Society of Gastroenterology guide-
lines on the diagnosis and management of patients at risk of gastric adenocar-
cinoma. Gut 2019;68:1545–75.
5. Conclusion
[4] Rugge M. Gastric Cancer Risk in Patients with Helicobacter pylori Infection and
Following Its Eradication. Gastroenterol Clin North Am 2015;44:609–24.
In conclusion, DL-based models could assist endoscopists in di- [5] Du Y, Bai Y, Xie P, et al. Chronic gastritis in China: a national multi-center
agnosing chronic atrophic gastritis with high efficacy and low cost. survey. BMC Gastroenterol 2014;14:21.
[6] Du T, Xie L, Zhang H, et al. Training and validation of a deep learning archi-
This may greatly reduce the burden on endoscopists, relieve the tecture for the automatic analysis of coronary angiography. EuroIntervention
suffering of patients, and allow for better detection of atrophic gas- 2021;17:32–40.
tritis in primary hospitals. [7] Gulshan V, Peng L, Coram M, et al. Development and Validation of a Deep
Learning Algorithm for Detection of Diabetic Retinopathy in Retinal Fundus
Photographs. JAMA 2016;316:2402–10.
Declaration of Competing Interest [8] Zhou L, Li Q, Huo G, et al. Image Classification Using Biomimetic Pattern
Recognition with Convolutional Neural Networks Features. Comput Intell Neu-
rosci 2017;2017:3792805. doi:10.1155/2017/3792805.
The authors declare no competing interests. [9] LeCun Y, Bengio Y, Hinton G. Deep learning. Nature 2015;521 436–444.
[10] Guo L, Xiao X, Wu C, et al. Real-time automated diagnosis of precancerous
Availability of data and materials lesions and early esophageal squamous cell carcinoma using a deep learning
model (with videos). Gastrointest Endosc 2020;91:41–51.
[11] Ikenoyama Y, Hirasawa T, Ishioka M, et al. Detecting early gastric cancer: com-
To protect patient privacy, the original data used to support the parison between the diagnostic ability of convolutional neural networks and
findings of this study could not be shared. endoscopists. Dig Endosc 2020;33:141–50.
[12] Nakashima H, Kawahira H, Kawachi H, et al. Helicobacter pylori Artificial intel-
ligence diagnosis of infection using blue laser imaging-bright and linked color
Ethics approval and consent to participate imaging: a single-center prospective study. Ann Gastroenterol 2018;31:462–8.
[13] Wang P, Berzin TM, Glissen Brown JR, et al. Real-time automatic detection sys-
tem increases colonoscopic polyp and adenoma detection rates: a prospective
This retrospective study conformed to the ethical guidelines of
randomised controlled study. Gut 2019;68:1813–19.
the Declaration of Helsinki and was approved by the ethics com- [14] Wang P, Liu X, Berzin TM, et al. Effect of a deep-learning computer-aided de-
mittee of the Third Xiangya Hospital of Central South University tection system on adenoma detection during colonoscopy (CADe-DB trial): a
(No.21160). According to national legislation and institutional re- double-blind randomised study. Lancet Gastroenterol Hepatol 2020;5:343–51.
[15] Stidham RW, Liu W, Bishu S, et al. Performance of a Deep Learning Model
quirements, informed consent was waived by the Ethics Commit- vs Human Reviewers in Grading Endoscopic Disease Severity of Patients with
tee of the Third Xiangya Hospital of Central South University and Ulcerative Colitis. JAMA Netw Open 2019;2:e193963.

1518
J. Luo, S. Cao, N. Ding et al. Digestive and Liver Disease 54 (2022) 1513–1519

[16] Klang E, Barash Y, Margalit RY, et al. Deep learning algorithms for automated Linqu County, China: baseline results and factors affecting the eradication. Gut
detection of Crohn’s disease ulcers by video capsule endoscopy. Gastrointest 2016;65:9–18.
Endosc 2020;91:606–13 e2. [26] The Chinese Society of Gastroenterology. China consensus report on chronic
[17] Barash Y, Azaria L, Soffer S, et al. Ulcer severity grading in video capsule im- gastritis (2017, Shanghai). Chin J Gastroenterol 2017;22:670–87.
ages of patients with Crohn’s disease: an ordinal neural network solution. Gas- [27] Pimentel-Nunes P, Libânio D, Marcos-Pinto R, et al. Management of epithe-
trointest Endosc 2021;93:187–92. lial precancerous conditions and lesions in the stomach (MAPS II): european
[18] Zhang Y, Li F, Yuan F, et al. Diagnosing chronic atrophic gastritis by gas- Society of Gastrointestinal Endoscopy (ESGE), European Helicobacter and Mi-
troscopy using artificial intelligence. Dig Liver Dis 2020;52:566–72. crobiota Study Group (EHMSG), European Society of Pathology (ESP), and So-
[19] Guimarães P, Keller A, Fehlmann T, et al. Deep-learning based detection of gas- ciedade Portuguesa de Endoscopia Digestiva (SPED) guideline update 2019. En-
tric precancerous conditions. Gut 2020;69:4–6. doscopy 2019;51:365–88.
[20] Kanai M, Togo R, Ogawa T, et al. Chronic atrophic gastritis detection with a [28] Redéen S, Petersson F, Jönsson K-A, et al. Relationship of gastroscopic features
convolutional neural network considering stomach regions. World J Gastroen- to histological findings in gastritis and Helicobacter pylori infection in a gen-
terol 2020;26:3650–9. eral population sample. Endoscopy 2003;35:946–50.
[21] Kuipers EJ, Uyterlinde AM, Peña AS, et al. Increase of Helicobacter pylori-as- [29] Uedo N, Yao K. Endoluminal Diagnosis of Early Gastric Cancer and Its Precur-
sociated corpus gastritis during acid suppressive therapy: implications for sors: bridging the Gap Between Endoscopy and Pathology. Adv Exp Med Biol
long-term safety. Am J Gastroenterol 1995;90:1401–6. 2016;908:293–316.
[22] Kaji K, Hashiba A, Uotani C, et al. Grading of Atrophic Gastritis is Useful for [30] Kanemitsu T, Yao K, Nagahama T, et al. Extending magnifying NBI diagnosis
Risk Stratification in Endoscopic Screening for Gastric Cancer. Am J Gastroen- of intestinal metaplasia in the stomach: the white opaque substance marker.
terol 2019;114:71–9. Endoscopy 2017;49:529–35.
[23] Choi IJ, Kook MC, Kim YI, et al. Helicobacter pylori Therapy for the Prevention [31] Uedo N. Advanced Imaging in the Diagnosis of Gastric Intestinal Metapla-
of Metachronous Gastric Cancer. N Engl J Med 2018;378:1085–95. sia: the Expert’s Approach. Video Journal and Encyclopedia of GI Endoscopy
[24] Hooi JK, Lai WY, Ng WK, et al. Global Prevalence of Helicobacter pylori 2013;1:112–14.
Infection: systematic Review and Meta-Analysis. Gastroenterology 2017;153: [32] Wei W, Can T, Xin W, et al. Image Object Recognition via Deep Fea-
420–429. ture-Based Adaptive Joint Sparse Representation. Comput Intell Neurosci
[25] Pan K, Zhang L, Gerhard M, et al. A large randomised controlled interven- 2019;2019:8258275.
tion trial to prevent gastric cancer by eradication of Helicobacter pylori in

1519