The application of artificial intelligence in gastrointestinal endoscopy: a state-of-the-art review

Chenxia  Zhang; Lianlian Wu

doi:10.55976/jdh.120221423-18

Home /

Archives /

Journal of Digital Health, Volume 1 Issue 1, 2022 /

Review

The application of artificial intelligence in gastrointestinal endoscopy: a state-of-the-art review

Chenxia Zhang ¹ , Lianlian Wu ² ^*

1. Department of Gastroenterology；Key Laboratory of Hubei Province for Digestive System Disease；Hubei Provincial Clinical Research Center for Digestive Disease Minimally Invasive Incision；Renmin Hospital of Wuhan University, Wuhan, China
2. Department of Gastroenterology；Key Laboratory of Hubei Province for Digestive System Disease；Hubei Provincial Clinical Research Center for Digestive Disease Minimally Invasive Incision；Renmin Hospital of Wuhan University, Wuhan, China

Correspondence: wu_leanne@whu.edu.cn

DOI: https://doi.org/10.55976/jdh.120221423-18

Received

22 November 2021
Revised

07 February 2022
Accepted

11 February 2022
Published

18 February 2022

Artificial intelligence Deep learning Gastrointestinal endoscopy Digestive diseases Quality control

Abstract

Endoscopy is an important tool for detecting and diagnosing digestive diseases. However, the performance of endoscopists varies, which significantly impacts the health outcomes of patients. In recent years, with the continuous development of science and technology, artificial intelligence (AI) has also set off a new wave. Nowadays, AI has been widely studied in the medical field and has shown great potential, especially in gastrointestinal endoscopy. The application of AI in endoscopy mainly includes the detection of lesions, classification of diseases, selection of best therapy, prognosis judgment, quality control, etc. What's more, the effectiveness and safety of the application of AI in gastrointestinal endoscopy have been confirmed in clinical trials. In this paper, we review the current research status and future development of AI in gastrointestinal endoscopy. Core Tips: Endoscopy is an important tool for detecting and diagnosing digestive diseases. However, the level of endoscopists varies, which significantly impacts the health outcomes of patients. In recent years, with the continuous development of science and technology, artificial intelligence (AI) has also set off a new wave. AI is increasingly widely used in the medical field nowadays and has shown great potential, especially in gastrointestinal endoscopy. The application of AI in endoscopy mainly includes the detection of lesions, classification of diseases, selection of best therapy, prognosis judgment, quality control, etc. What's more, the effectiveness and safety of the application of AI in gastrointestinal endoscopy have been confirmed in clinical trials. In this paper, we review the current research status and future development of AI in gastrointestinal endoscopy.

Introduction

Gastrointestinal cancer is one of the most common and deadly malignant tumors worldwide. According to the latest statistics (2020) <a href="#reference_list_0">[1]</a>, colorectal cancer and stomach cancer rank in the top five cancers amongst women; amongst men, they are the third and fourth most common cancers, respectively. In addition, nearly 2.5 million people die of gastrointestinal cancer every year. Among the leading causes of cancer-related deaths, colorectal cancer and gastric cancer rank second and fourth respectively, causing a huge social and economic burden. The basic reason that malignant tumors are life-threatening is that they cannot be detected in the early stage <a href="#reference_list_1">[2]</a>. Gastrointestinal tumors are generally treatable in their early stages <a href="#reference_list_2">[3, 4]</a>. More importantly, the 5-year survival rate for early cancer of the digestive tract is higher than 90%, however, if cancer progresses to the middle and advanced stage, the 5-year survival rate will be less than 25% <a href="#reference_list_4">[5]</a>. Therefore, early diagnosis and early treatment are the keys to improve the survival rate of patients with gastrointestinal cancer. Endoscopy is the most commonly used method for screening and detecting gastrointestinal diseases. In recent decades, rapid progress in endoscopy has been seen worldwide. However, behind the vigorous development, some quality and safety concerns also exist. Mucosal changes in early gastrointestinal cancer are slight and difficult to identify, which requires endoscopists to be armed with rich experience and strong knowledge <a href="#reference_list_5">[6, 7]</a>. However, there are great variations among the level and performance of endoscopists, resulting in a high rate of missed detection of lesions and poor quality of gastrointestinal endoscopy, which seriously threatens the health and prognosis of patients <a href="#reference_list_7">[8, 9]</a>. With the rapid development of computer technology and big data, artificial intelligence (AI) has gradually entered the public eye and the medical industry in recent years <a href="#reference_list_9">[10-12]</a>. Nowadays, AI has made tremendous advances in medical fields such as radiology, pathology, and dermatology <a href="#reference_list_12">[13-15]</a>. At present, numerous studies have shown the great potential of AI in endoscopy, which is expected to bring revolutionary changes to the diagnosis and treatment of digestive endoscopy (16). This article summarizes the AI technology and its research in the field of digestive endoscopy, including (1) introduction of AI and deep learning; (2) the application of AI technology in the field of endoscopy: diagnosis of cancer or pre-cancer diseases, diagnosis of other diseases, and quality control; (3) challenges and prospects of artificial intelligence.

<h5 class="journalThemeColor">Overview of artificial intelligence</h5>
Artificial intelligence and deep learning
Broadly speaking, artificial intelligence (AI) is a new technical science that enables computers to simulate certain thinking processes and intelligent behaviors of human beings, such as learning, reasoning, thinking, and planning <a href="#reference_list_16">[17, 18]</a>. As an important branch of computer science, AI is one of the three cutting-edge technologies (space technology, energy technology, and artificial intelligence) in the world. The development of AI has experienced three waves (12). The first two waves occurred in the 1970s and 1990s <a href="#reference_list_18">[19, 20]</a>, respectively. Due to the constraints of algorithms and computing power, AI did not achieve good results in the practical application of various industries at that time. With the deep learning technology proposed in 2006, AI ushered in the third wave (21). In 2012, the ImageNet competition brought a breakthrough to the application of artificial intelligence in the field of image recognition and also made deep learning a widely used method in the field of image recognition. Different from the previous two waves, AI is not limited to academic theory but has entered into wide fields, showing extraordinary practical results and gradually changing our lives.
Deep Learning, whose principle is to establish a neural network that analyzes and interprets data by imitating the operation mechanism of the human brain, is an important branch of technology in the field of AI and the latest development trend of artificial neural networks (17). By extracting more abstract features of the input data layer by layer from the lower level to the higher level, the deep learning model forms the most appropriate network weight structure of the required features, to achieve an accurate classification. Deep learning mainly includes Deep Neural Network (DNN), Convolutional Neural Network (CNN), Recurrent Neural Network (RNN), and so on (22). Although various deep learning architectures have been explored to solve different tasks, convolutional neural networks (CNNs) remain the most popular type of deep learning architecture in medical imaging today. A typical CNN is mainly composed of the input layer, convolution layer, pooling layer, and output layer (23). The input layer receives input values and passes them to the next layer, but does not perform operations on the input value. The first hidden layer is the nonlinear mapping of the input data, and the second hidden layer is also the nonlinear mapping of the previous layer. They get the maps that are easy to classify through weight updating and pass the values to the output layer, to get the ideal results within a reasonable range. The artificial neurons of CNN respond to part of the surrounding units within the coverage, which helps with processing large images (24).
Deep learning in computer vision
Computer Vision (CV), a discipline that teaches machines to see, has a rich history of decades. Advances in neural networks and deep learning have greatly promoted the development of computer vision recognition systems (25). Four types of tasks are commonly applied to computer vision using deep learning, namely image classification, object detection, semantic segmentation, and instance segmentation. Image classification is the most basic application in computer vision, the purpose of it is to determine the classification of a given picture. The training data set contains a variety of different objects, and the objects included in a given picture are required to output. Object detection refers to further clarifying the specific positions of various objects in the picture with a rectangular frame based on image classification. The essence of semantic segmentation is to classify every pixel in the image to distinguish various objects in the image. For the positions of objects of different categories, each pixel should be distinguished, but not for different objects belonging to the same category. Instance segmentation is the combination of object detection and semantic segmentation. It not only uses object detection to box out different objects in the picture but also needs to use semantic segmentation to understand the given box at the pixel level. Compared with the boundary box given by object detection, instance segmentation can be further refined to the edge of the object. What's more, it can further annotate different objects of the same categories in the picture compared with semantic segmentation.
<h5 class="journalThemeColor">Application of AI in upper gastrointestinal endoscopy</h5>
<h5 class="journalThemeColor">Esophageal cancer</h5>
Esophageal cancer is one of the most fatal tumors. In 2020, more than 604,000 people were newly diagnosed with esophageal cancer worldwide, accounting for 3.1% of new cancer cases globally each year <a href="#reference_list_0">[1]</a>. Furthermore, esophageal squamous cell carcinoma (ESCC) is the major subtype of Esophageal cancer, accounting for more than 90% of esophageal cancer in China <a href="#reference_list_25">[26]</a>. However, in recent years, the incidence of esophageal adenocarcinoma has increased significantly. And it has surpassed the incidence of esophageal squamous cell carcinoma gradually in many western countries <a href="#reference_list_26">[27]</a>. With early diagnosis and treatment, the 5-year survival rate of patients with esophageal cancer is up to 90% <a href="#reference_list_27">[28]</a>. However, most esophageal cancers are in the middle and late stages at the time of diagnosis, and the overall 5-year survival rate is less than 20% <a href="#reference_list_0">[1]</a>. There is no doubt that early diagnosis of esophageal cancer is crucial to improve the prognosis.
LinJie Guo et al. used 6473 NBI images, including dysplasia, early ESCC, and non-cancerous lesions from 549 patients to train the semantic segmentation model SegNet for real-time diagnosis of esophageal precancerous lesions and early ESCC <a href="#reference_list_28">[29]</a>. The sensitivity, specificity, and the area under the curve (AUC) in predicting precancerous lesions and early ESCC in a test set consisting of 6,671 NBI images from 59 cancerous patients and 2004 non-cancerous patients were 98.04%, 95.03%, and 0.989, respectively. In the test set of 47 patients with cancer and 33 patients without cancer, the model achieved a per-case sensitivity of 100% and a per-case specificity of 90.9%. The authors conclude that the deep learning model achieved high sensitivity and specificity in images and videos, and has the potential to assist endoscopists in diagnosing precancerous lesions and ESCC in real-time. As we all know, Intrapapillary capillary loops (IPCLs) are an effective endoscopic marker for the detection of squamous dysplasia and esophageal squamous carcinoma. Everson MA et al. used 67,742 magnification endoscopy narrow-band(ME-NBI) images of 115 patients to construct a convolutional neural network that could classify IPCL patterns <a href="#reference_list_29">[30]</a>. The model predicted dysplasia with an accuracy of 91.7% and sensitivity of 93.7%, which was preferable to previous studies <a href="#reference_list_30">[31]</a>. In short, this kind of clinically interpretable CNN based on the IPCL model puts forward a new idea for our future research.
The key point of early esophageal cancer is not the diagnosis, but the judgment of invasion depth, which is very crucial for the selection of treatment <a href="#reference_list_31">[32]</a>. Nevertheless, the current diagnostic theory of invasion depth is complex, and the judgment is often subjective. Kentaro et al. developed a deep learning system to the evaluate invasion depth of esophageal superficial squamous cell carcinoma using 8660 non-ME images and 5678 ME images from 804 patients whose invasion depth was confirmed by pathology <a href="#reference_list_32">[33]</a>. The system performed well in diagnosing the depth of invasion of superficial esophageal squamous cell carcinoma with a sensitivity of 90.1%, a specificity of 95.8%, and an accuracy of 91.0%, comparable to experienced endoscopists.
Barrett's esophagus
Barrett's esophagus (BE) refers to a pathological phenomenon in which the lamellar squamous epithelium of the lower mucosa of the esophagus is replaced by a single columnar epithelium. BE is the only known precancerous lesion of esophageal adenocarcinoma (EAC), and early detection and follow-up monitoring of patients with BE are conducive to the improvement of prognosis <a href="#reference_list_33">[34, 35]</a>. However, the lesion of BE is relatively mild, which could be easily ignored or missed. Guidelines recommend an endoscopic biopsy of the four quadrants of BE at random intervals of 1-2cm to detect dysplasia in time <a href="#reference_list_31">[32]</a>, but this method is invasive, time-consuming, and difficult to adhere to in some cases. The use of deep learning to detect BE and distinguish BE from neoplasia is expected to improve detection efficiency.
Hashimoto R et al. collected 916 images from 65 cases with early esophageal neoplasia in BE containing high-grade dysplasia or T1 cancer and 919 images from 119 cases of BE without high-grade dysplasia or T1 cancer, to train image classification and target detection nested model (Inception-ResNet-v2 as the main model to detect whether there is a target lesion in the image, and adding Yolo to locate the lesion), to identify and locate early tumors from BE images (36). In the 458 test images (225 early tumors and 233 controls), the sensitivity, specificity, and accuracy of the model to detect early tumors were 96.4%, 94.2%, and 95.4%, respectively. The intersection over union (IOU) between the model and the endoscopist on the localization of the lesion was 0.3. The authors conclude that the model can detect early esophageal tumors in BE images with high accuracy and map a positioning box around the areas of early esophageal neoplasia with high accuracy. What's more, Ebigbo A et al developed a real-time deep learning artificial intelligence system based on their previous work <a href="#reference_list_36">[37]</a>. The system differentiated and predicted normal BE and early oesophageal adenocarcinoma (EAC) accurately. This is the first real-time application of a deep learning AI system in the evaluation and diagnosis of early EAC in BE in a real-life setting.
The differentiation between Barrett's cancer mucosa (T1a) and submucosal invasion (T1b) is the key to the choice of treatment and prognosis <a href="#reference_list_31">[32]</a>. Unfortunately, it is quite challenging (38). Ebigbo A et al. trained a deep learning system to distinguish T1a from T1b Barrett's-related cancer using 230 white light endoscopy images (108 T1a and 122 T1b) from 116 patients with Barrett's cancer <a href="#reference_list_38">[39]</a>. The sensitivity, specificity, and accuracy of the system were 0.77, 0.64, 0.71, respectively. This suggested its diagnostic performance was comparable to that of experts. However, this study was based on static images, so it still needed further improvement to be put into the practical clinical application as soon as possible.
Esophageal protruded lesions
Nearly 30% of the protruded lesions of the upper digestive tract originate in the esophagus (40). In clinical practice, it is important to distinguish esophageal protruded lesions, such as esophageal leiomyoma (EL), esophageal cyst (EC), and esophageal papilloma (EP). But it is difficult to judge by white light image alone, and endoscopic ultrasonography (EUS) is often required <a href="#reference_list_40">[41]</a>. Unfortunately, EUS has a steep learning curve and remains a challenging task for endoscopists of all levels of experience <a href="#reference_list_41">[42]</a>. Min Zhang et al. developed a CNN-based model using WLE and EUS images of 1217 patients with benign esophageal protruded lesions to identify and differentiate the three subtypes (EL, EC, EP) (43). For the identification of esophageal benign lesions from healthy controls, the AUC of the CNN model was 0.751. The model achieved an AUC of 0.907, 0.897, 0.868 for the identification of EP, EL, and EC under WL, while 0.739 and 0.724 for EL and EC under EUS. Compared with all endoscopists, the model achieved higher sensitivity and specificity than all endoscopists in correctly classifying EL and EC using EUS images.
Esophagogastric varices
Esophagogastric varices (EGV) are one of the main complications of cirrhosis, and the death rate of acute variceal rupture and bleeding can reach 15%-50% <a href="#reference_list_43">[44]</a>. Early identification of patients with a high risk of variceal hemorrhage in cirrhosis is very important for the primary prevention of variceal hemorrhage. Esophagogastroduodenoscopy (EGD) is recommended in the Guidelines as a tool for assessing the risk of bleeding from esophageal and gastric varices, but this is often subjective in practical clinical practice <a href="#reference_list_44">[45]</a>. Chen M et al. trained a DCNN system (ENDOANGEL) based on 8,566 endoscopic images of gastroesophageal varices in 3021 patients and 6,152 normal esophagus/stomach images in 3168 patients for the diagnosis of gastroesophageal varices and the prediction of rupture risk <a href="#reference_list_45">[46]</a>. As a result, ENDOANGEL detected esophageal varices (EVs) and gastric varices (GVs) at the accuracy of 97.00% and 92.00% respectively, comparable to that of experts. Moreover, it has achieved excellent results in the detection of endoscopic risk factors for esophagogastric variceal bleeding. The authors point out that although this system still has some limitations in the judgment of treatment decisions, prospective studies are needed to further verify the performance of the model. As for patients with compensated advanced chronic liver disease, Agarwal S et al developed a machine learning model to predict bleed in esophageal varices <a href="#reference_list_46">[47]</a>. The model improved the performance of endoscopic stratification to predict VB with an accuracy of 98.7%.
Gastric cancer
More than 1 million new cases of gastric cancer occur globally each year, making it the fourth leading cause of cancer-related death <a href="#reference_list_0">[1]</a>. The 5-year survival rate of advanced gastric cancer is 5%-25%, while early endoscopic diagnosis and treatment can improve the survival rate to more than 90% <a href="#reference_list_4">[5]</a>. Early detection and treatment is a critical issue to reduce mortality and improve survival rates. The recognition and diagnosis of EGC based on deep learning are expected to reduce the missed diagnosis of gastric cancer and enhance survival.
Our previous work used a total of 3,170 gastric cancer and 5,981 benign images to train the vgg-16 and resnet-50 for the detection of EGC <a href="#reference_list_47">[48]</a>. Based on the image classification model, a class-like activation map is further developed to automatically cover suspicious cancerous areas. One hundred EGC and one hundred non-cancerous images were used as a test set to evaluate and compare the diagnostic capabilities of the model and endoscopists. The results showed that the accuracy, sensitivity, specificity, positive predictive value, and negative predictive value of the model were 92.5%, 94.0%, 91.0%, 91.3%, and 93.8% respectively, which were better than that of endoscopists. At the same time, the study of real-time monitoring of EGC in early cancer and non-cancer videos has demonstrated a good effect. Li L et al. collected 386 images of non-cancerous lesions and 1702 images of early gastric cancer to train the classification model Inception-v3 to distinguish early and non-cancer by magnifying endoscopy with narrow-band imaging (ME-NBI) <a href="#reference_list_48">[49]</a>. The model showed comparable accuracy (90.9%) and superior sensitivity (91.2%) to that of experts. In addition, Tang D et al. used 3407 endoscopic images from 666 patients with gastric cancer to construct a deep convolutional neural network model to differentiate intramucosal GC from advanced GC <a href="#reference_list_49">[50]</a>. With the help of this model, novice endoscopists(84.6%) achieved the same level of accuracy as specialists(85.5%).
With the development of endoscopic treatment for EGC, endoscopic submucosal dissection (ESD) is becoming more and more popular (51). However, whether ESD can be performed depends to a large extent on the depth of invasion, degree of differentiation, and boundary of gastric cancer <a href="#reference_list_51">[52, 53]</a>. Zhu Y et al. developed a CNN-CAD system for judging the depth of gastric cancer invasion, which can effectively distinguish early gastric cancer from deep submucosal invasion and minimize the overestimation of the depth of invasion to avoid overtreatment <a href="#reference_list_53">[54]</a>. Ling T et al. retrospectively collected 2217 images of differentiated gastric cancer and 11,870 images of undifferentiated gastric cancer to establish a model to evaluate the degree of differentiation of gastric cancer <a href="#reference_list_54">[55]</a>. In addition, 694 images of differentiated cancer and 234 images of undifferentiated cancer labeled by experts were used to train and test the gastric cancer boundary delineation model to accurately identify the differentiation state and delineate the edge of EGC under ME-NBI endoscopy. The accuracy of the model to identify the differentiation degree of EGC reached 83.3%. At the overlap rate of 0.80, the accuracy rate for differentiated and undifferentiated EGC was 82.7% and 88.1%, respectively, which was better than that of endoscopy experts and had potential clinical practical value. The system performed excellently in a multicenter, prospective, real-time, competitive comparative, diagnostic study <a href="#reference_list_55">[56]</a>.
Precancerous lesions
Chronic inflammation of gastric mucosa induces a series of precancerous conditions (gastric atrophy, intestinal metaplasia) and lesions (dysplasia), leading to the occurrence and development of gastric cancer <a href="#reference_list_56">[57]</a>. In patients with extensive atrophy of gastric mucosa, the 5-year incidence of gastric cancer can reach 1.9%-10%, while in patients with intestinal metaplasia, the incidence is as high as 5.3%-9.8% <a href="#reference_list_57">[58]</a>. Therefore, early identifying and monitoring the precancerous conditions is beneficial to the healthy outcome of patients. Zhang Y et al. collected 3042 images of atrophic gastritis and 2428 images of non-atrophic gastritis from the gastric antrum of 1470 patients, aiming to construct a classification model to diagnose chronic atrophic gastritis <a href="#reference_list_58">[59]</a>. The authors conclude that the deep learning model has high sensitivity (94.5%) and accuracy (94.2%) in the diagnosis of atrophic gastritis, which can improve the diagnostic ability of endoscopists on atrophic gastritis. Based on 6250 endoscopic images from 760 patients with precancerous gastric cancer and 98 videos from 77 patients who underwent image-enhanced endoscopy (IEE), Xu M et al. constructed ENDOANGEL, a deep convolutional neural network system, for the detection of precancerous gastric cancer <a href="#reference_list_59">[60]</a>. The system showed high accuracy, sensitivity, and specificity in detecting precancerous lesions, and the diagnostic accuracy (GA 0.901; IE 0.908) is close to the expert level, which provides the possibility of being widely used in the diagnosis of precancerous lesions of gastric cancer.
Helicobacter pylori (Hp) infection
Helicobacter pylori (HP) is a ubiquitous microbe, which can be found in 50% of the world's population (61). Chronic infection with Hp can lead to gastric atrophy and even metaplasia. The International Agency for Research on Cancer (IARC) of the World Health Organization (WHO) believes that 78% of gastric cancers can be attributed to chronic infection caused by Hp and has classified it as a class I carcinogen <a href="#reference_list_61">[62]</a>. Timely diagnosis and eradication of HP are considered to be an important strategy to prevent gastric cancer. With the continuous progress of endoscopic technology, more and more studies have found that morphological changes of esophageal and gastric mucosa under endoscopy (such as atrophy, punctured redness, mucosal swelling, erosion, etc.) are related to HP infection <a href="#reference_list_62">[63, 64]</a>. However, there are no objective indicators for these endoscopic features, which greatly limits their clinical application. Shichijo S et al. constructed a convolutional neural network that could diagnose Hp infection <a href="#reference_list_64">[65]</a>. The sensitivity and accuracy of the model in determining Hp infection through gastroscopic images reached 81.9% and 83.1%, showing excellent performance. Zheng W et al. collected 11,729 gastric images from 1,507 patients (847 of whom were infected with Hp) to train classification model ResNet-50 and tested the model in 3,755 images from 452 patients (310 of whom were infected with Hp) <a href="#reference_list_65">[66]</a>. The results showed that the AUC per image and per patient was 0.93, 0.97, respectively. with sensitivity, specificity, and accuracy of 91.6%, 98.6%, and 93.8%, which indicated a high diagnostic accuracy. To evaluate the infection status of Helicobacter pylori (not infected, past infection, current infection), Yoshii S et al. recruited 498 subjects and developed a predictive model by evaluating their gastritis endoscopic results based on the Kyoto classification <a href="#reference_list_62">[63]</a>. The overall diagnostic accuracy of the model for gastritis was 82.9%. With further improvement, this model might be very helpful to novice endoscopists.
Quality control
Gastroscopy is a key step in the diagnosis of upper gastrointestinal diseases. However, endoscopists vary in their performance levels, which affected the detection of gastric cancer and precancerous lesions. At present, most of the published articles focus on the diagnosis or detection of lesions, while a few studies focus on the quality of endoscopic examination.
Our previous work used more than 60,000 EGD images and divided them into training set and test set at the ratio of 9:1. Endoscopists studied the guidelines of the European Society of Gastrointestinal Endoscopy (ESGE) and the Japanese systematic screening protocol and independently labeled EGD images into 26 different sites <a href="#reference_list_66">[67]</a>. Then VGG-16 was trained to realize the function of classifying gastric sites <a href="#reference_list_67">[68]</a>. The model monitored blind spots with an accuracy of 90.40% in real EGD videos. What's more, the blind spot rate of the patients in AI-assisted group (5.9%) was significantly lower than that of the control group (22.5%) in our clinical trial, which suggested that this system had the potential to improve the quality of daily endoscopic examination. In our most recent prospective randomized controlled study, the number of blind areas during gastroscopy was reduced from 9.82% to 5.39% with the aid of AI <a href="#reference_list_68">[69]</a>.
Table 1. Examples of the application of artificial intelligence in digestive endoscopy
<table style="margin: 0; padding: 0; border-collapse: collapse; border-spacing: 0; border: 1px solid #777777;">
<tbody>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Lesions</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">Diagnostic or predictive modality</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">AI model</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">Number of images/casesin training dataset</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">Number of images/casesin test dataset</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">result</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="8">Esophageal cancer</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Guo L et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">narrow-band imaging (NBI)</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">SegNet</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">6473 NBI images</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">6670 NBI images80 videos</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">SensitivityImage dataset: 93.6%Nonmagnifying video dataset: 46.9%Magnifying video dataset: 85.8%</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Everson MA et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">magnification endoscopy</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">CNN</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">54,193 magnification endoscopy narrow-band images</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">13,548 magnification endoscopy narrow-band images</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">F1 score: 94%Sensitivity: 93.7%Accuracy: 91.7%</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Nakagawa K et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">upper GI endoscopes</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">CNN</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">8660 non-magnified endoscopic (non-ME) and 5678 ME images from 804 superficial esophageal SCCs</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">405 non-ME images and 509 ME images from 155 patients</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">SM1, SM2/3：Sensitivity: 90.1%Specificity: 95.8%Accuracy : 91.0%</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="8">Barrett's esophagus</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Hashimoto R et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">chromoendoscopy</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">CNN</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">960 images from 65 patients</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">458 images (225 dysplasia and 233 non-dysplasia)</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">Sensitivity: 96.4%Specificity: 94.2% Accuracy: 95.4%</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Ebigbo A et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">upper GI endoscopes</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">CNN</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">230 white-light endoscopic images (108 T1a and 122 T1b)</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">Sensitivity: 0.77 Specificity: 0.64Accuracy: 0.71</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="8">Esophageal protruded lesions</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Zhang M et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">white-light endoscopy EUS</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">CNN</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">identification of esophageal benign lesions from healthy controls: 17279 WLI images from 598 patientsdifferentiation esophageal leiomyoma, esophageal cyst, esophageal papilloma: 3226 WLI images from 619 patients discrimination between EL and EC using EUS images: 3411 EUS images from 248 patients</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">AUC: 0.751AUC: EP: 0.907 EL: 0.897 EC: 0.868AUC: EL:0.739 EC: 0.724</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="8">Esophagogastric varices</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Chen M et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">upper GI endoscopes</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">DCNN</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">8566 images (gastroesophageal varices) 6152 images (normal esophagus/stomach)</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">-</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">Accuracy: esophageal varices (EVs): 97.00% gastric varices (GVs): 92.00%</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="8">Gastric cancer</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Wu L et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">esophagogastroduodenoscopy</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">DCNN</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">24,549 images of different parts of the stomach were taken to train the model to monitor the blind spot3170 images of gastric cancer and 5981 images of benign lesions were used to train the model to recognize early cancer</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">Accuracy：10 parts classification : 90% 26 parts classification: 65.9% early cancer in WL images: 92.4%</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Li L et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">magnifying endoscopy</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">CNN</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">386 images of non-cancerous lesions and 1702 images of early gastric cancer</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">341 endoscopic images (171 non-cancerous lesions and 170 early gastric cancer)</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">Sensitivity: 91.18% Specificity: 90.64% Accuracy: 90.91%</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Tang D et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">esophagogastroduodenoscopy</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">DCNN</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">3407 endoscopic images from 666 gastric cancer patients</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">228 images from 62 patients</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">AUC: 0.942 Sensitivity: 90.5% Specificity: 85.3%</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Zhu Y et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">esophagogastroduodenoscopy</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">CNNResNet50</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">790 images</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">203 images</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">Ensitivity: 76.47% Specificity: 95.56%Accuracy: 89.16%</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Ling T et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">magnifying narrow-band imaging (ME-NBI) endoscopy</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">CNN</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">1131 images of differentiated and 1086 images of undifferentiated gastric cancer</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">1526 images from 105 differentiated gastric cancers and 344 images from 34 undifferentiated gastric cancers</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">Accuracy: 83.3 %</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Wu L et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">magnifying narrow-band imaging endoscopy</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">-</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">-</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">37 EGCs and 63 noncancerous lesions</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">Sensitivity: Detect neoplasms 87.81% Diagnose EGCs 100%Accuracy: Predict EGC invasion depth 78.57% Differentiate status 71.43%</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="8">Precancerous lesions</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Zhang Y et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">esophagogastroduodenoscopy</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">CNN</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">5470 images of the gastric antrums of 1699 patients</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">Diagnose atrophic gastritis Accuracy: 0.942 Sensitivity: 0.945Specificity: 0.940</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Xu M et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">image-enhanced endoscopy (IEE)</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">DCNN</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">3,049 images of precancerous states and 2,149 images of chronic inflammation</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">1052 images98 videos</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">Accuracy:gastric atrophy: 0.901intestinal metaplasia: 0.908</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="8">Helicobacter pylori (Hp) infection</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Shichijo S et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">esophagogastroduodenoscopy</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">CNN</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">32,208 images either positive or negative for H. pylori</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">11,481 images from 397 patients</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">CNN1: sensitivity 81.9% specificity: 83.4%accuracy: 83.1%CNN2: sensitivity 88.9% specificity: 87.4%accuracy : 87.7%</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Zheng W et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">upper endoscopy</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">ResNet-50</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">11,729 gastric images</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">3,755 images</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">sensitivity：81.4%specificity: 90.1%accuracy: 84.5%</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Yoshii S et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">upper endoscopy</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">CNN</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">498 patients</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">Accuracy: Kyoto classification of gastritis: 82.9%without H. pylori eradication history: 88.6% with eradication history: 93.4%</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="8">Quality control</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Wu L et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">esophagogastroduodenoscopy</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">VGG-16DenseNet</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">DCNN1:12 220 in vitro, 25 222 in vivo and 16 760 unqualified EGD imagesDCNN2:34 513 qualified EGD images</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">DCNN1:3000 images (1000 per category)DCNN2:2160 images (80 per site)</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">monitor blind spots with an average accuracy of 90.02%</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Wu L et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">esophagogastroduodenoscopy</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">-</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">ENDOANGEL groups: 498 patients Control groups:504 patients</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">The number of blind spots: ENDOANGEL group: 5.39%Control group: 9.82%</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="8">Application of AI in capsule colonoscopy</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Aoki T et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">wireless capsule endoscopy</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">CNN</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">5360 WCE images of erosions and ulcerations</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">440 images of erosions and ulcerations</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">Ensitivity: 88.2% Specificity: 90.9% Accuracy: 90.8%</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Ding Z et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">small bowel capsule endoscopy (SB-CE)</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">CNN</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">158,235 SB-CE images from 1970 patients</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">5000 patients</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">Sensitivity of identifying abnormalitiesper-patient analysis: 99.88% per-lesion analysis: 99.90%</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Yamada A et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">colon capsule endoscopy</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">Single Shot MultiBox Detector</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">15 933 CCE images of colorectal neoplasms</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">4784 images(1850 images of colorectal neoplasms and 2934 normal colon images)</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">Sensitivity:79.0 % Specificity: 87.0 % Accuracy: 83.9 %</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Leenhardt R et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">small-bowel capsule endoscopy</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">CNN</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">-</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">-</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">sensitivity : 90.3 %Specificity: 83.3 %Accuracy: 89.7 %</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="8">colorectal cancer</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Luo X et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">white-light colonoscopy</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">CNN</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">7734 nonmagnified white-light colonoscopy (WLC) imagesfrom 657 lesions</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">1634 WLC images from 156 lesions</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">Accuracy: 91.1% Sensitivity: 91.2%Specificity:91.0%</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Ichimasa K et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">colonoscopy</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">SVM</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">590 patients</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">100 patients</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">Sensitivity: 100 %Specificity: 66%Accuracy: 69%</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="8">Polyps and adenomas</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Kudo SE et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">colonoscopy</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">CNN</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">69,142 endocytoscopic images</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">100 lesions from 89 patients</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">Sensitivity: 96.9% Specificity: 100% Accuracy: 98%</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Wang P et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">colonoscopy</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">-</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">1,290 patients</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">612 polyp-containing images</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">per-image-sensitivity, 88.24%</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Wang P et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">colonoscopy</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">-</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">standard colonoscopy group:536 patientscomputer-aided group:522 patients</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">ADR standard colonoscopy group: 29.1%computer-aided group: 20.3%</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="8">Inflammatory bowel disease</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Takenaka K et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">colonoscopy</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">DNUC</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">40,758 images of colonoscopies</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">4187 endoscopic images and 4104 biopsy specimens</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">The DNUC identified patients in histologic remission with 92.9% accuracy</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Gottlieb K et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">colonoscopy</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">-</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">795 prospectively recorded full-length endoscopy</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">quadratic weighted kappa (QWK):eMS : 0.844UCEIS: 0.855</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="8">Quality control</td>
</tr>
<tr style="padding: 0;">
<td style="padding: 2px 5px; border: 1px solid #777777;">Zhou J et al.</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">colonoscopy</td>
<td style="padding: 2px 5px; border: 1px solid #777777;" colspan="2">CNN</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">5583 clear colonoscopic images</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">120 images (30 in each classification)</td>
<td style="padding: 2px 5px; border: 1px solid #777777;">BBPS four classification scoring accuracy: 93.33%</td>
</tr>
</tbody>
</table>
<h5 class="journalThemeColor"> </h5>
<h5 class="journalThemeColor">Application of AI in capsule colonoscopy</h5>
Capsule endoscopy (CE) entered the domain of small intestine disease diagnosis in 2001 (70). As a non-invasive deglutition diagnostic device, CE can not only reduce patients' discomfort during the examination but also clearly observe various types of intestinal mucosal abnormalities (71). However, the major limitation is that CE is time-consuming and has a high risk of missed detection <a href="#reference_list_71">[72]</a>. Aoki T et al. collected 5,360 CE images of erosion and ulcer to train the detection model SSD and tested it in 10,440 CE images (including 440 erosion and ulcer) <a href="#reference_list_72">[73]</a>. The deep learning model can automatically detect erosion and ulcers with an accuracy of 90.8% in the CE image, thus reducing the burden on endoscopists. Furthermore, Ding Z et al. collected 113,426,569 CE colonoscopy images from 6970 patients who had undergone small bowel capsule endoscopy in 77 medical centers to conduct a classification model ResNet which could distinguish normal between abnormal images <a href="#reference_list_73">[74]</a>. With the assistance of AI model, endoscopists identified abnormalities with a higher sensitivity(99.88%) and shorter time( 5.9±2.23 min per case) with the assistant of AI. Recently, Yamada A et al. used 15 933 colon capsule endoscopy (CCE) images of colon tumors, such as polyps and cancers, to develop a deep convolutional neural network based on a Single Shot MultiBox Detector for automatic detection of colorectal tumors (75). Leenhardt R et al. also developed a neural network-based algorithm that can automatically evaluate the cleanliness of the small intestine during capsule endoscopy with a high sensitivity of 90.3% <a href="#reference_list_75">[76]</a>. What's more, Saraiva MM et al developed a Convolutional Neural Network model using a database of CCE images to detect protruding lesions <a href="#reference_list_76">[77]</a>. This model performed well in detecting protruding lesions with sensitivity and specificity of 90.7%, 92.6%, respectively.
<h5 class="journalThemeColor">Application of AI in colonoscopy</h5>
colorectal cancer
Colorectal cancer is the second most deadly malignancy in the world1, and new cases of colorectal cancer are expected to rise to 2.5 million worldwide by 2035 <a href="#reference_list_77">[78]</a>, which will place huge health and economic burden on society. Screening for colorectal cancer, early detection and treatment of precancerous lesions and asymptomatic early cancer are essential to delay progression and reduce mortality. Colonoscopy is an important means for early diagnosis and treatment of lower gastrointestinal diseases and is the gold standard for colorectal cancer screening <a href="#reference_list_78">[79]</a>. Unfortunately, the current application of AI in colonoscopy mainly focuses on the detection and classification of polyps <a href="#reference_list_79">[80, 81]</a>. However, once colorectal cancer is diagnosed, evaluation of the depth of invasion is our first task, which is crucial for the selection of treatment strategy <a href="#reference_list_51">[52]</a>. Luo X et al. constructed a deep convolutional neural network with tumor localization branches based on GoogLeNet architecture to evaluate the invasion depth using 7734 white light colonoscopy images of 657 lesions <a href="#reference_list_81">[82]</a>. Pathological results were the gold standard. Consequently, the model achieved an accuracy of 91.1% in the prediction of non-invasive and superficial invasive tumors. In addition, the model could distinguish between superficial submucosal invasive tumors and deep invasive CRC with an accuracy similar to that of experienced endoscopists. Some studies also focus on the treatment and prognosis of patients with colorectal cancer, trying to use AI to select the optimal treatment plan for patients. Ichimasa K et al. developed a machine learning model to predict the risk of lymph node metastasis. Consequently, it showed that AI could significantly reduce unnecessary additional surgery after endoscopic resection of T1 colorectal cancer <a href="#reference_list_82">[83]</a>.
Polyps and adenomas
Removal of adenomas found by colonoscopy can significantly reduce the incidence and mortality of colorectal cancer <a href="#reference_list_83">[84, 85]</a>. Each 1.0% increase in the adenoma detection rate was associated with a 3.0% decrease in the risk of cancer <a href="#reference_list_85">[86]</a>. However, the level of endoscopists varies, and polyps may be missed during colonoscopy. In one study, the missed rate of adenomas was as high as 26% <a href="#reference_list_86">[87]</a>. At the same time, it is important to distinguish between non-neoplastic and neoplastic lesions for the treatment and prognosis of patients. The traditional method is to take biopsy to identify neoplastic lesions <a href="#reference_list_87">[88]</a>, but this method is invasive and time-consuming. Accurate optical diagnosis of colorectal polyps can decrease the cost of colonoscopy and reduce the complications related to polyp biopsy.
Kudo SE et al. collected 69,142 images of endoscopy from patients with colorectal polyps in 5 academic centers in Japan to train the system EndoBRAIN <a href="#reference_list_88">[89]</a>. Compared with 30 endoscopists (20 trainees and 10 specialists), EndoBRAIN showed a better performance in the diagnosis of neoplasia in stained endocytoscopic images with a sensitivity of 96.9%. Furthermore, the sensitivity and accuracy of EndoBRAIN in narrow-band images were 96.9% and 96.0%, which were significantly better than that of non-experts and comparable with that of experts.
To improve the detection rate of colorectal polyps, early detection and removal of colorectal polyps is of great significance for the prevention of colorectal cancer. Wang P et al. conducted a segmentation model SegNet to detect colorectal polyps, which performed well with the sensitivity of 94.38% and specificity of 95.92% <a href="#reference_list_89">[90]</a>. Subsequently, a randomized, controlled and single-blind clinical trial, of which 536 received conventional colonoscopy and 522 received AI-assisted colonoscopy was conducted <a href="#reference_list_90">[91]</a>. The results showed that the AI system could significantly improve the adenoma detection rate (ADR) from 20.3% to 29.1% and the mean number of adenomas detected per patient (MNA) from 0.31 to 0.53. Recently, To study the influence of artificial intelligence system on ADR after eliminating the operational bias, Wang P et al. further designed and conducted a randomized double-blind clinical trial involving 1,010 patients <a href="#reference_list_91">[92]</a>. The AI-assisted group simulates the alarm box on polyps; while the sham control group simulates the alarm box on the polypoid non-polyp mucosas. Only the observer can see the output of the systems on a second monitor and report the alert to endoscopists. Only when the area predicted by the AI system is not thought to be a polyp by the operating endoscopist and is about to leave the field of vision, will the observer point at the area to alert the endoscopist. The results showed that the ADR of the AI-assisted group(34%) was significantly higher than that of the sham control group(28%), which indicated that the AI system could help endoscopists detect more polyps and adenomas in real time. Furthermore, Sinonquel P et al also used a second observer in a non-RCT setting to assess the miss rate of AI system and endoscopists <a href="#reference_list_92">[93]</a>. Both the system and the endoscopists performed well with a real-time accuracy of 96.5% and 98.2%, which suggested the system was non-inferior in terms of sensitivity and might aid in guaranteeing quality in colonoscopy.
Inflammatory bowel disease
At present, the prevalence of inflammatory bowel disease (IBD) is on the rise and it has become a global disease (94). As we all know, delayed diagnosis of IBD significantly increases the risk of surgical resection and complications <a href="#reference_list_94">[95, 96]</a>. Objective endoscopic evaluation of inflammatory bowel disease is of great significance in promoting the treatment of patients and improving the prognosis <a href="#reference_list_96">[97]</a>. However, endoscopic assessment requires training for endoscopists, and there are usually differences in the assessment among different endoscopists (98). It is difficult to diagnose IBD, misdiagnosis and missed diagnosis are widespread in the world. Takenaka K et al. collected 40,758 colonoscopy images from 2,012 patients with ulcerative colitis (UC) to train the classification model to assess the severity of UC and validated it prospectively with 4187 endoscopic images from 875 patients with UC <a href="#reference_list_98">[99]</a>. Endoscopic remission was defined as the ulcerative colitis endoscopic index of severity (UCEIS) score is 0. Histological remission was defined as Geboes score ≤ 3. The deep learning model is highly accurate in assessing UC with an accuracy of 90.1% and 92.9% in predicting endoscopic remission and histological remission, and is expected to help endoscopists identify remission in UC patients without biopsy. What's more, Bossuyt P et al also develop an operator-independent computer-based tool to determine UC activity based on endoscopic images, which provided an objective computer-based score that accurately assesses disease activity in UC <a href="#reference_list_99">[100]</a>. In their validation study, they tested the correlation between the algorithm RD score and clinical, endoscopic and histological features. Consequently, RD correlated with endoscopic and histological disease activity. To identify the endoscopic disease activity scoring in UC in the video, Gottlieb K et al. collected 795 endoscopy videos prospectively and constructed a system based on convolutional neural network and recurrent neural network to evaluate UC endoscopy Mayo score(eMS) and Ulcerative Colitis Endoscopic Index of Severity (UCEIS) score <a href="#reference_list_100">[101]</a>. Quadratic weighted kappa (QWK), which was used to evaluate the model, was 0.844 and 0.855 for eMS and UCEIS respectively. Soffer S et al. was encouraged by this innovative research and believed that it would be a stepping stone for the future development of AI in Esophagogastroduodenoscopy (102).
Quality control
In recent years, the incidence and mortality of colorectal cancer have been increasing and it has become one of the major malignant tumors threatening human health. Colonoscopy is the gold standard for CRC screening, and high-quality colonoscopy is the basis for the early detection of lesions. Studies have shown that adenoma detection rate (ADR), as an important indicator of colonoscopy quality control <a href="#reference_list_85">[86]</a>, can decrease the risk of interphase colon cancer by 3% and the mortality of colorectal cancer by 5% with an increase of 1%. The key indicators to evaluate the quality of colonoscopy mainly include intestinal preparation, cecal intubation rate, ADR, etc <a href="#reference_list_87">[88]</a>. However, the existing number of endoscopists cannot meet the increasing demand for colonoscopy, and the level of endoscopists varies, which will lead to the uneven quality of colonoscopy. Therefore, we should further strengthen the quality control of colonoscopy, and reduce the rate of missed diagnosis of the lower gastrointestinal tract during colonoscopy.
Inadequate intestinal preparation is associated with a reduced adenoma detection rate (ADR). Our previous work used 4,764 colonoscopy images, which were graded from 0 to 3 according to the Boston bowel score (BBPS) by endoscopies to train the classification model DenseNet <a href="#reference_list_102">[103]</a>. In predicting the bowel preparation in real-time, the model achieved an accuracy of 93.33% in a test set of 480 colonoscopy images (120 for each category) and 89.04% in 20 colonoscopy videos. We have recently developed an intestinal preparation assessment system for calculating automatic BBPS (E-BBPS) scores (range 0-20) <a href="#reference_list_103">[104]</a>. The system classifies colonoscopy video images into qualified and unqualified images, then calculates the proportion of unqualified frames in real-time to reflect intestinal preparation. In addition, the correlation between the proportion of intestinal unqualified preparation and the ADR was explored based on 616 prospective colonoscopy screening videos. We were pleasantly surprised to find that there was a significant negative correlation between the proportion of intestinal preparation unqualified frames and the ADR (ρ = -0.976, P<0.01), ADR was significantly lower in patients with scores greater than 3 than in patients with scores less than 3 (15.93% vs 28.03%, OR 0.43). This suggests that our system has the potential to provide more objective and refined thresholds for the quantification of adequate intestinal preparation.
It is well known that the withdrawal time is significantly associated with the incidence of interval colorectal cancer <a href="#reference_list_104">[105]</a>. As is recommended in the guidelines, it shouldn't be less than 6 minutes <a href="#reference_list_87">[88]</a>. Our previous work used more than 20,000 images of colonoscopy to construct a real-time quality improvement system through VGG-16 and perceptual hashing algorithm to monitor real-time withdrawal speed, withdrawal time and remind endoscopists of blind spots caused by scope slipping <a href="#reference_list_105">[106]</a>. Subsequently, 791 patients were recruited for a randomized controlled clinical trial. The results showed that the ADR of the AI-assisted group (16.34%) was significantly higher than that of the control group (7.74%). We conclude that the system has greatly improved the detection of adenomas and polyps by improving the quality of colonoscopy, and it is expected to be a powerful tool to narrow the skill gap among endoscopists.
<h5 class="journalThemeColor">Summary and prospect</h5>
With the rapid development of computer technology, the application of AI in the field of digestive endoscopy has mushroomed. Needless to say, the future of artificial intelligence is promising. However, it is also full of challenges. Behind the rapid development of technology, there are still many problems and difficulties to be solved. Firstly, the sample data set is important to the deep learning model, and the quantity and quality of the data set directly affect the accuracy and generalization ability of the model. How to evaluate the quality of the data set, how to improve the quality of sample labeling, and how to ensure that there is no cross contamination between training and testing sets, etc., are problems worth further discussion. In addition, there are "black boxes" in the logic of deep-learning algorithms' decision-making processes that are difficult for humans to understand, preventing doctors from discovering potential confounding factors. More importantly, AI is still in the development stage, and there are no relevant rules or formation of industry norms yet. In future practical applications, AI cannot know the comprehensive situation of patients, nor does it understand the laws and responsibilities except the picture information read. If misdiagnosis occurs, how to solve the ethical problem and who should bear the corresponding responsibility?
The free and open-source sharing of algorithms ensures us to use the most advanced deep learning model to study and explore the application of deep learning in medical imaging. We need to start from the actual clinical needs, take the basic theory as the starting point, and form a set of independent digestive endoscopy AI technology development mechanism with the accumulation of time experience. Hope we could see that AI bring revolutionary changes to digestive endoscopy and even the whole medical field, contributing to improve patients' outcome worldwide.
<h5 class="journalThemeColor">Acknowledgments</h5>
My deepest gratitude goes first and foremost to my senior, Lianlian Wu, for her constant encouragement and guidance. She helped me work out the outline of the manuscript and walked me through all the stages of the writing of this manuscript. I believe that without her consistent and illuminating instruction, this review could not have reached its present form. Then, my thanks would go to the reviewers and editors, who spent much time and effort on reading and revising my paper.
<h5 class="journalThemeColor">Authors' contributions</h5>
CXZ analyzed the research data and drafted the manuscript; LLW revised the manuscript. All the authors discussed, reviewed and edited its content before submission.
<h5 class="journalThemeColor">Disclosure</h5>
There are no financial conflicts of interest to disclose.

References

[1]Sung H, Ferlay J, Siegel RL, Laversanne M, Soerjomataram I, Jemal A, et al. Global Cancer Statistics 2020: GLOBOCAN Estimates of Incidence and Mortality Worldwide for 36 Cancers in 185 Countries. CA: A Cancer Journal for Clinicians. 2021; 71(3):209-249. doi: 10.3322/caac.21660.

[2]Cai Q, Zhu C, Yuan Y, Feng Q, Feng Y, Hao Y, et al. Development and validation of a prediction rule for estimating gastric cancer risk in the Chinese high-risk population: a nationwide multicentre study. Gut. 2019; 68(9):1576-1587. doi: 10.1136/gutjnl-2018-317556.

[3]Smyth EC, Nilsson M, Grabsch HI, van Grieken NC, Lordick F. Gastric cancer. Lancet. 2020; 396(10251):635-648. doi: 10.1016/S0140-6736(20)31288-5.

[4]Ladabaum U, Dominitz JA, Kahi C, Schoen RE. Strategies for Colorectal Cancer Screening. Gastroenterology. 2020; 158(2):418-432. doi: 10.1053/j.gastro.2019.06.043.

[5]Van Cutsem E, Sagaert X, Topal B, Haustermans K, Prenen H. Gastric cancer. Lancet. 2016; 388(10060):2654-2664. doi: 10.1016/S0140-6736(16)30354-3.

[6]Yoshida N, Doyama H, Yano T, Horimatsu T, Uedo N, Yamamoto Y, et al. Early gastric cancer detection in high-risk patients: a multicentre randomised controlled trial on the effect of second-generation narrow band imaging. Gut. 2021; 70(1):67-75. doi: 10.1136/gutjnl-2019-319631.

[7]Emura F, Rodriguez-Reyes C, Giraldo-Cadavid L. Early Gastric Cancer: Current Limitations and What Can Be Done to Address Them. The American Journal of Gastroenterology. 2019; 114(6):841-5. doi: 10.14309/ajg.0000000000000220.

[8]Zhang X, Li M, Chen S, Hu J, Guo Q, Liu R, et al. Endoscopic Screening in Asian Countries Is Associated With Reduced Gastric Cancer Mortality: A Meta-analysis and Systematic Review. Gastroenterology. 2018; 155(2):347-54.e9. doi: 10.1053/j.gastro.2018.04.026.

[9]Shah SC, Canakis A, Peek RM, Jr., Saumoy M. Endoscopy for Gastric Cancer Screening Is Cost Effective for Asian Americans in the United States. Clinical Gastroenterology Hepatology. 2020; 18(13):3026-39. doi: 10.1016/j.cgh.2020.07.031.

[10]Yu KH, Beam AL, Kohane IS. Artificial intelligence in healthcare. Nature Biomedical Engineering. 2018; 2(10):719-731. doi: 10.1038/s41551-018-0305-z.

[11]Silver D, Huang A, Maddison CJ, Guez A, Sifre L, van den Driessche G, et al. Mastering the game of Go with deep neural networks and tree search. Nature. 2016; 529(7587):484-9. doi: 10.1038/nature16961.

[12]Krittanawong C, Kaplin S. Artificial Intelligence in Global Health. European Heart Journal. 2021; 42(24):2321-2322. doi: 10.1093/eurheartj/ehab036.

[13]Hosny A, Parmar C, Quackenbush J, Schwartz LH, Aerts H. Artificial intelligence in radiology. Nature reviews. Cancer. 2018; 18(8):500-510. doi: 10.1038/s41568-018-0016-5.

[14]Niazi MKK, Parwani AV, Gurcan MN. Digital pathology and artificial intelligence. The Lancet. Oncology. 2019; 20(5):e253-e61. doi: 10.1016/S1470-2045(19)30154-8.

[15]Lee EY, Maloney NJ, Cheng K, Bach DQ. Machine learning for precision dermatology: Advances, opportunities, and outlook. Journal of the American Academy Dermatology. 2021; 84(5):1458-1459. doi: 10.1016/j.jaad.2020.06.1019.

[16]Chahal D, Byrne MF. A primer on artificial intelligence and its application to endoscopy. Gastrointestinal Endoscopy. 2020; 92(4):813-20.e4. doi: 10.1016/j.gie.2020.04.074.

[17]LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015; 521(7553):436-44. doi: 10.1038/nature14539.

[18]Mahmud M, Kaiser MS, Hussain A, Vassanelli S. Applications of Deep Learning and Reinforcement Learning to Biological Data. IEEE Transactions on Neural Networks and Learning System. 2018; 29(6):2063-2079. doi: 10.1109/TNNLS.2018.2790388.

[19]Rumelhart DE, Hinton GE, Williams RJ. Learning representations by back-propagating errors. Nature. 1986; 323(6088):533-536. doi: 10.1038/323533a0.

[20]LeCun Y, Boser B, Denker JS, Henderson D, Howard RE, Hubbard W, et al. Backpropagation Applied to Handwritten Zip Code Recognition. Neural Computation. 1989; 1(4):541-551. doi: 10.1162/neco.1989.1.4.541.

[21]Hinton GE, Osindero S, Teh YW. A fast learning algorithm for deep belief nets. Neural Computation. 2006; 18(7):1527-54. doi: 10.1162/neco.2006.18.7.1527.

[22]L. Deng, G. Hinton and B. Kingsbury. New types of deep neural network learning for speech recognition and related applications: an overview. 2013 IEEE International Conference on Acoustics, Speech and Signal Processing. Vancouver, BC, Canada: IEEE; 2013. pp. 8599-8603. doi: 10.1109/ICASSP.2013.6639344.

[23]Lecun Y, Bottou L, Bengio Y, Haffner P. Gradient-based learning applied to document recognition. In: Proceedings of the IEEE. 1998; 86(11):2278-324. doi: 10.1109/5.726791.

[24]Krizhevsky A, Sutskever I, Hinton GE. ImageNet Classification with Deep Convolutional Neural Networks. Communications of the ACM. 2017; 60(6):84-90. doi: 10.1145/3065386.

[25]Esteva A, Chou K, Yeung S, Naik N, Madani A, Mottaghi A, et al. Deep learning-enabled medical computer vision. NPJ Digital Medicine. 2021; 4(1):5. doi: 10.1038/s41746-020-00376-2.

[26]Abnet CC, Arnold M, Wei WQ. Epidemiology of Esophageal Squamous Cell Carcinoma. Gastroenterology. 2018; 154(2):360-373. doi: 10.1053/j.gastro.2017.08.023.

[27]Coleman HG, Xie SH, Lagergren J. The Epidemiology of Esophageal Adenocarcinoma. Gastroenterology. 2018; 154(2):390-405. doi: 10.1053/j.gastro.2017.07.046.

[28]Yang XX, Li Z, Shao XJ, Ji R, Qu JY, Zheng MQ, et al. Real-time artificial intelligence for endoscopic diagnosis of early esophageal squamous cell cancer (with video). Digestive Endoscopy. 2021; 33(7):1075-1084. doi: 10.1111/den.13908.

[29]Guo L, Xiao X, Wu C, Zeng X, Zhang Y, Du J, et al. Real-time automated diagnosis of precancerous lesions and early esophageal squamous cell carcinoma using a deep learning model (with videos). Gastrointestinal Endoscopy. 2020; 91(1):41-51. Doi: 10.1016/j.gie.2019.08.018.

[30]Everson MA, Garcia-Peraza-Herrera L, Wang HP, Lee CT, Chung CS, Hsieh PH, et al. A clinically interpretable convolutional neural network for the real-time prediction of early squamous cell cancer of the esophagus: comparing diagnostic performance with a panel of expert European and Asian endoscopists. Gastrointestinal Endoscopy. 2021; 94(2):273-281. doi: 10.1016/j.gie.2021.01.043.

[31]Oyama T, Inoue H, Arima M, Momma K, Omori T, Ishihara R, et al. Prediction of the invasion depth of superficial squamous cell carcinoma based on microvessel morphology: magnifying endoscopic classification of the Japan Esophageal Society. Esophagus. 2017; 14(2):105-112. doi: 10.1007/s10388-016-0527-7.

[32]Sharma P, Shaheen NJ, Katzka D, Bergman J. AGA Clinical Practice Update on Endoscopic Treatment of Barrett's Esophagus With Dysplasia and/or Early Cancer: Expert Review. Gastroenterology. 2020; 158(3):760-769. doi: 10.1053/j.gastro.2019.09.051.

[33]Nakagawa K, Ishihara R, Aoyama K, Ohmori M, Nakahira H, Matsuura N, et al. Classification for invasion depth of esophageal squamous cell carcinoma using a deep neural network compared with experienced endoscopists. Gastrointestinal Endoscopy. 2019; 90(3):407-414. doi: 10.1016/j.gie.2019.04.245.

[34]Geboes K, Hoorens A. The cell of origin for Barrett's esophagus. Science. 2021; 373(6556):737-738. doi: 10.1126/science.abj9797.

[35]Kamboj AK, Leggett CL. Barrett's esophagus indefinite for dysplasia carries a definite risk of neoplasia. Gastrointestinal Endoscopy. 2021; 94(2):271-272. doi: 10.1016/j.gie.2021.03.019.

[36]Hashimoto R, Requa J, Dao T, Ninh A, Tran E, Mai D, et al. Artificial intelligence using convolutional neural networks for real-time detection of early esophageal neoplasia in Barrett's esophagus (with video). Gastrointestinal Endoscopy. 2020; 91(6):1264-71.e1. doi: 10.1016/j.gie.2019.12.049.

[37]Ebigbo A, Mendel R, Probst A, Manzeneder J, Prinz F, de Souza LA, Jr., et al. Real-time use of artificial intelligence in the evaluation of cancer in Barrett's oesophagus. Gut. 2020; 69(4):615-616. doi: 10.1136/gutjnl-2019-319460.

[38]Frei NF, Stachler MD, Bergman J. Risk stratification in Barrett's esophagus patients with diagnoses of indefinite for dysplasia: the definite silver bullet has not (yet) been found. Gastrointestinal Endoscopy. 2020; 91(1):11-3. doi: 10.1016/j.gie.2019.09.020.

[39]Ebigbo A, Mendel R, Rückert T, Schuster L, Probst A, Manzeneder J, et al. Endoscopic prediction of submucosal invasion in Barrett's cancer with the use of artificial intelligence: a pilot study. Endoscopy. 2021; 53(9):878-883. doi: 10.1055/a-1311-8570.

[40]Hoda KM, Rodriguez SA, Faigel DO. EUS-guided sampling of suspected GI stromal tumors. Gastrointestinal Endoscopy. 2009; 69(7):1218-23. doi: 10.1016/j.gie.2008.09.045.

[41]Kuroki K, Oka S, Tanaka S, Yorita N, Hata K, Kotachi T, et al. Clinical significance of endoscopic ultrasonography in diagnosing invasion depth of early gastric cancer prior to endoscopic submucosal dissection. Gastric Cancer. 2021; 24(1):145-155. doi: 10.1007/s10120-020-01100-5.

[42]Cassani L, Aihara H, Anand GS, Chahal P, Dacha S, Duloy A, et al. Core curriculum for EUS. Gastrointestinal Endoscopy. 2020; 92(3):469-473. doi: 10.1016/j.gie.2020.04.026.

[43]Zhang M, Zhu C, Wang Y, Kong Z, Hua Y, Zhang W, et al. Differential diagnosis for esophageal protruded lesions using a deep convolution neural network in endoscopic images. Gastrointestinal Endoscopy. 2021; 93(6):1261-1272.e2. doi: 10.1016/j.gie.2020.10.005.

[44]Dong TS, Kalani A, Aby ES, Le L, Luu K, Hauer M, et al. Machine Learning-based Development and Validation of a Scoring System for Screening High-Risk Esophageal Varices. Clinical Gastroenterology Hepatol. 2019; 17(9):1894-1901.e1. doi: 10.1016/j.cgh.2019.01.025.

[45]Garcia-Tsao G, Abraldes JG, Berzigotti A, Bosch J. Portal hypertensive bleeding in cirrhosis: Risk stratification, diagnosis, and management: 2016 practice guidance by the American Association for the study of liver diseases. Hepatology. 2017; 65(1):310-335. doi: 10.1002/hep.28906.

[46]Chen M, Wang J, Xiao Y, Wu L, Hu S, Chen S, et al. Automated and real-time validation of gastroesophageal varices under esophagogastroduodenoscopy using a deep convolutional neural network: a multicenter retrospective study (with video). Gastrointestinal Endoscopy. 2021; 93(2):422-32.e3. doi: 10.1016/j.gie.2020.06.058.

[47]Agarwal S, Sharma S, Kumar M, Venishetty S, Bhardwaj A, Kaushal K, et al. Development of a machine learning model to predict bleed in esophageal varices in compensated advanced chronic liver disease: A proof of concept. Journal of Gastroenterology Hepatol. 2021; 36(10):2935-2942. doi: 10.1111/jgh.15560.

[48]Wu L, Zhou W, Wan X, Zhang J, Shen L, Hu S, et al. A deep neural network improves endoscopic detection of early gastric cancer without blind spots. Endoscopy. 2019; 51(6):522-531. doi: 10.1055/a-0855-3532.

[49]Li L, Chen Y, Shen Z, Zhang X, Sang J, Ding Y, et al. Convolutional neural network for the diagnosis of early gastric cancer based on magnifying narrow band imaging. Gastric Cancer. 2020; 23(1):126-132. doi: 10.1007/s10120-019-00992-2.

[50]Tang D, Zhou J, Wang L, Ni M, Chen M, Hassan S, et al. A Novel Model Based on Deep Convolutional Neural Network Improves Diagnostic Accuracy of Intramucosal Gastric Cancer (With Video). Frontiers in Oncology. 2021; 11:622827. doi: 10.3389/fonc.2021.622827.

[51]Hanada Y, Wang KK. Safety and feasibility of same-day discharge after esophageal endoscopic submucosal dissection. Gastrointestinal Endoscopy. 2021; 93(4):853-860. doi: 10.1016/j.gie.2020.07.037.

[52]Draganov PV, Wang AY, Othman MO, Fukami N. AGA Institute Clinical Practice Update: Endoscopic Submucosal Dissection in the United States. Clinical Gastroenterologgy Hepatology. 2019; 17(1):16-25.e1. doi: 10.1016/j.cgh.2018.07.041.

[53]Japanese Gastric Cancer Association. Japanese gastric cancer treatment guidelines 2018 (5th edition). Gastric Cancer. 2021; 24(1):1-21. doi: 10.1007/s10120-020-01042-y.

[54]Zhu Y, Wang QC, Xu MD, Zhang Z, Cheng J, Zhong YS, et al. Application of convolutional neural network in the diagnosis of the invasion depth of gastric cancer based on conventional endoscopy. Gastrointestinal Endoscopy. 2019; 89(4):806-15.e1. doi: 10.1016/j.gie.2018.11.011.

[55]Ling T, Wu L, Fu Y, Xu Q, An P, Zhang J, et al. A deep learning-based system for identifying differentiation status and delineating the margins of early gastric cancer in magnifying narrow-band imaging endoscopy. Endoscopy. 2021; 53(5):469-477. doi: 10.1055/a-1229-0920.

[56]Wu L, Wang J, He X, Zhu Y, Jiang X, Chen Y, et al. Deep learning system compared with expert endoscopists in predicting early gastric cancer and its invasion depth and differentiation status (with videos). Gastrointestinal Endoscopy. 2022; 95(1):92-104.e3. doi: 10.1016/j.gie.2021.06.033.

[57]den Hollander WJ, Holster IL, den Hoed CM, Capelle LG, Tang TJ, Anten MP, et al. Surveillance of premalignant gastric lesions: a multicentre prospective cohort study from low incidence regions. Gut. 2019; 68(4):585-593. doi: 10.1136/gutjnl-2017-314498.

[58]Shichijo S, Hirata Y, Niikura R, Hayakawa Y, Yamada A, Ushiku T, et al. Histologic intestinal metaplasia and endoscopic atrophy are predictors of gastric cancer development after Helicobacter pylori eradication. Gastrointestinal Endoscopy. 2016; 84(4):618-24. doi: 10.1016/j.gie.2016.03.791.

[59]Zhang Y, Li F, Yuan F, Zhang K, Huo L, Dong Z, et al. Diagnosing chronic atrophic gastritis by gastroscopy using artificial intelligence. Digestive and Liver Disease. 2020; 52(5):566-572. doi: 10.1016/j.dld.2019.12.146.

[60]Xu M, Zhou W, Wu L, Zhang J, Wang J, Mu G, et al. Artificial intelligence in the diagnosis of gastric precancerous conditions by image-enhanced endoscopy: a multicenter, diagnostic study (with video). Gastrointestinal Endoscopy. 2021; 94(3):540-548.e4. doi: 10.1016/j.gie.2021.03.013.

[61]Leja M, Grinberga-Derica I, Bilgilier C, Steininger C. Review: Epidemiology of Helicobacter pylori infection. Helicobacter. 2019; 24 Suppl 1:e12635. doi: 10.1111/hel.12635.

[62]Samet JM, Chiu WA, Cogliano V, Jinot J, Kriebel D, Lunn RM, et al. The IARC Monographs: Updated Procedures for Modern and Transparent Evidence Synthesis in Cancer Hazard Identification. Journal of National Cancer Institute. 2020; 112(1):30-37. doi: 10.1093/jnci/djz169.

[63]Yoshii S, Mabe K, Watano K, Ohno M, Matsumoto M, Ono S, et al. Validity of endoscopic features for the diagnosis of Helicobacter pylori infection status based on the Kyoto classification of gastritis. Digestive Endoscopy. 2020; 32(1):74-83. doi: 10.1111/den.13486.

[64]Okamura T, Iwaya Y, Kitahara K, Suga T, Tanaka E. Accuracy of Endoscopic Diagnosis for Mild Atrophic Gastritis Infected with Helicobacter pylori. Clinical Endoscopy. 2018; 51(4):362-367. doi: 10.5946/ce.2017.177.

[65]Shichijo S, Nomura S, Aoyama K, Nishikawa Y, Miura M, Shinagawa T, et al. Application of Convolutional Neural Networks in the Diagnosis of Helicobacter pylori Infection Based on Endoscopic Images. EBioMedicine. 2017; 25:106-11. doi: 10.1016/j.ebiom.2017.10.014.

[66]Zheng W, Zhang X, Kim JJ, Zhu X, Ye G, Ye B, et al. High Accuracy of Convolutional Neural Network for Evaluation of Helicobacter pylori Infection Based on Endoscopic Images: Preliminary Experience. Clinical and Translational Gastroenterology. 2019; 10(12):e00109. doi: 10.14309/ctg.0000000000000109.

[67]Bisschops R, Areia M, Coron E, Dobru D, Kaskas B, Kuvaev R, et al. Performance measures for upper gastrointestinal endoscopy: a European Society of Gastrointestinal Endoscopy (ESGE) Quality Improvement Initiative. Endoscopy. 2016; 48(9):843-64. doi: 10.1055/s-0042-113128.

[68]Wu L, Zhang J, Zhou W, An P, Shen L, Liu J, et al. Randomised controlled trial of WISENSE, a real-time quality improving system for monitoring blind spots during esophagogastroduodenoscopy. Gut. 2019; 68(12):2161-2169. doi: 10.1136/gutjnl-2018-317366.

[69]Wu L, He X, Liu M, Xie H, An P, Zhang J, et al. Evaluation of the effects of an artificial intelligence system on endoscopy quality and preliminary testing of its performance in detecting early gastric cancer: a randomized controlled trial. Endoscopy. 2021; 53(12):1199-1207. doi: 10.1055/a-1350-5583.

[70]Iddan G, Meron G, Glukhovsky A, Swain P. Wireless capsule endoscopy. Nature. 2000; 405(6785):417. doi: 10.1038/35013140.

[71]Flemming J, Cameron S. Small bowel capsule endoscopy: Indications, results, and clinical benefit in a University environment. Medicine (Baltimore). 2018; 97(14):e0148. doi: 10.1097/MD.0000000000010148.

[72]Costamagna G, Shah SK, Riccioni ME, Foschia F, Mutignani M, Perri V, et al. A prospective trial comparing small bowel radiographs and video capsule endoscopy for suspected small bowel disease. Gastroenterology. 2002;123(4):999-1005. doi: 10.1053/gast.2002.35988.

[73]Aoki T, Yamada A, Aoyama K, Saito H, Tsuboi A, Nakada A, et al. Automatic detection of erosions and ulcerations in wireless capsule endoscopy images based on a deep convolutional neural network. Gastrointestinal Endoscopy. 2019; 89(2):357-363.e2. doi: 10.1016/j.gie.2018.10.027.

[74]Ding Z, Shi H, Zhang H, Meng L, Fan M, Han C, et al. Gastroenterologist-Level Identification of Small-Bowel Diseases and Normal Variants by Capsule Endoscopy Using a Deep-Learning Model. Gastroenterology. 2019; 157(4):1044-1054.e5. doi: 10.1053/j.gastro.2019.06.025.

[75]Yamada A, Niikura R, Otani K, Aoki T, Koike K. Automatic detection of colorectal neoplasia in wireless colon capsule endoscopic images using a deep convolutional neural network. Endoscopy. 2021; 53(8):832-836. doi: 10.1055/a-1266-1066.

[76]Leenhardt R, Souchaud M, Houist G, Le Mouel JP, Saurin JC, Cholet F, et al. A neural network-based algorithm for assessing the cleanliness of small bowel during capsule endoscopy. Endoscopy. 2021; 53(9):932-936. doi: 10.1055/a-1301-3841.

[77]Saraiva MM, Ferreira JPS, Cardoso H, Afonso J, Ribeiro T, Andrade P, et al. Artificial intelligence and colon capsule endoscopy: development of an automated diagnostic system of protruding lesions in colon capsule endoscopy. Techniques Coloproctology. 2021; 25(11):1243-1248. doi: 10.1007/s10151-021-02517-5.

[78]Araghi M, Soerjomataram I, Jenkins M, Brierley J, Morris E, Bray F, et al. Global trends in colorectal cancer mortality: projections to the year 2035. International Journal of Cancer. 2019; 144(12):2992-3000. doi: 10.1002/ijc.32055.

[79]Chen H, Li N, Ren J, Feng X, Lyu Z, Wei L, et al. Participation and yield of a population-based colorectal cancer screening programme in China. Gut. 2019; 68(8):1450-1457. doi: 10.1136/gutjnl-2018-317124.

[80]Hassan C, Spadaccini M, Iannone A, Maselli R, Jovani M, Chandrasekar VT, et al. Performance of artificial intelligence in colonoscopy for adenoma and polyp detection: a systematic review and meta-analysis. Gastrointestinal Endoscopy. 2021; 93(1):77-85.e6. doi: 10.1016/j.gie.2020.06.059.

[81]Barua I, Vinsard DG, Jodal HC, Løberg M, Kalager M, Holme Ø, et al. Artificial intelligence for polyp detection during colonoscopy: a systematic review and meta-analysis. Endoscopy. 2021; 53(3):277-284.doi: 10.1055/a-1201-7165.

[82]Luo X, Wang J, Han Z, Yu Y, Chen Z, Huang F, et al. Artificial intelligence-enhanced white-light colonoscopy with attention guidance predicts colorectal cancer invasion depth. Gastrointestinal Endoscopy. 2021; 94(3):627-638.e1. doi: 10.1016/j.gie.2021.03.936.

[83]Ichimasa K, Kudo SE, Mori Y, Misawa M, Matsudaira S, Kouyama Y, et al. Artificial intelligence may help in predicting the need for additional surgery after endoscopic resection of T1 colorectal cancer. Endoscopy. 2018; 50(3):230-240. doi: 10.1055/s-0043-122385.

[84]Click B, Pinsky PF, Hickey T, Doroudi M, Schoen RE. Association of Colonoscopy Adenoma Findings With Long-term Colorectal Cancer Incidence. JAMA. 2018; 319(19):2021-2031. doi: 10.1001/jama.2018.5809.

[85]Waldmann E, Kammerlander AA, Gessl I, Penz D, Majcher B, Hinterberger A, et al. Association of Adenoma Detection Rate and Adenoma Characteristics With Colorectal Cancer Mortality After Screening Colonoscopy. Clinical Gastroenterology and Hepatology. 2021; 19(9):1890-1898. doi: 10.1016/j.cgh.2021.04.023.

[86]Kaminski MF, Wieszczy P, Rupinski M, Wojciechowska U, Didkowska J, Kraszewska E, et al. Increased Rate of Adenoma Detection Associates With Reduced Risk of Colorectal Cancer and Death. Gastroenterology. 2017; 153(1):98-105. doi: 10.1053/j.gastro.2017.04.006.

[87]Zhao S, Wang S, Pan P, Xia T, Chang X, Yang X, et al. Magnitude, Risk Factors, and Factors Associated With Adenoma Miss Rate of Tandem Colonoscopy: A Systematic Review and Meta-analysis. Gastroenterology. 2019; 156(6):1661-1674.e11. doi: 10.1053/j.gastro.2019.01.260.

[88]Kaminski MF, Thomas-Gibson S, Bugajski M, Bretthauer M, Rees CJ, Dekker E, et al. Performance measures for lower gastrointestinal endoscopy: a European Society of Gastrointestinal Endoscopy (ESGE) Quality Improvement Initiative. Endoscopy. 2017; 49(4):378-397. doi: 10.1055/s-0043-103411.

[89]Kudo SE, Misawa M, Mori Y, Hotta K, Ohtsuka K, Ikematsu H, et al. Artificial Intelligence-assisted System Improves Endoscopic Identification of Colorectal Neoplasms. Clinical Gastroenterology and Hepatology. 2020; 18(8):1874-1881.e2. doi: 10.1016/j.cgh.2019.09.009.

[90]Wang P, Xiao X, Glissen Brown JR, Berzin TM, Tu M, Xiong F, et al. Development and validation of a deep-learning algorithm for the detection of polyps during colonoscopy. Nature Biomedical Engineering. 2018; 2(10):741-748. doi: 10.1038/s41551-018-0301-3.

[91]Wang P, Berzin TM, Glissen Brown JR, Bharadwaj S, Becq A, Xiao X, et al. Real-time automatic detection system increases colonoscopic polyp and adenoma detection rates: a prospective randomised controlled study. Gut. 2019; 68(10):1813-1819. doi: 10.1136/gutjnl-2018-317500.

[92]Wang P, Liu X, Berzin TM, Glissen Brown JR, Liu P, Zhou C, et al. Effect of a deep-learning computer-aided detection system on adenoma detection during colonoscopy (CADe-DB trial): a double-blind randomised study. The lancet. Gastroenterology & hepatology. 2020; 5(4):343-351. doi: 10.1016/S2468-1253(19)30411-X.

[93]Sinonquel P, Eelbode T, Hassan C, Antonelli G, Filosofi F, Neumann H, et al. Real-time unblinding for validation of a new CADe tool for colorectal polyp detection. Gut. 2021; 70(4):641-643. doi: 10.1136/gutjnl-2020-322491.

[94]Kaplan GG. The global burden of IBD: from 2015 to 2025. Nature reviews. Gastroenterology & Hepatology. 2015; 12(12):720-7. doi: 10.1038/nrgastro.2015.150.

[95]Lee DW, Koo JS, Choe JW, Suh SJ, Kim SY, Hyun JJ, et al. Diagnostic delay in inflammatory bowel disease increases the risk of intestinal surgery. World Journal of Gastroenterology. 2017; 23(35):6474-6481. doi: 10.3748/wjg.v23.i35.6474.

[96]Moon CM, Jung SA, Kim SE, Song HJ, Jung Y, Ye BD, et al. Clinical Factors and Disease Course Related to Diagnostic Delay in Korean Crohn's Disease Patients: Results from the Connect Study. PLoS One. 2015; 10(12):e0144390. doi: 10.1371/journal.pone.0144390.

[97]Frøslie KF, Jahnsen J, Moum BA, Vatn MH. Mucosal healing in inflammatory bowel disease: results from a Norwegian population-based cohort. Gastroenterology. 2007; 133(2):412-22. doi: 10.1053/j.gastro.2007.05.051.

[98]Osada T, Ohkusa T, Yokoyama T, Shibuya T, Sakamoto N, Beppu K, et al. Comparison of several activity indices for the evaluation of endoscopic activity in UC: inter- and intraobserver consistency. Inflammatory Bowel Diseases. 2010; 16(2):192-7. doi: 10.1002/ibd.21000.

[99]Takenaka K, Ohtsuka K, Fujii T, Negi M, Suzuki K, Shimizu H, et al. Development and Validation of a Deep Neural Network for Accurate Evaluation of Endoscopic Images From Patients With Ulcerative Colitis. Gastroenterology. 2020; 158(8):2150-2157. doi: 10.1053/j.gastro.2020.02.012.

[100]Bossuyt P, Nakase H, Vermeire S, de Hertogh G, Eelbode T, Ferrante M, et al. Automatic, computer-aided determination of endoscopic and histological inflammation in patients with mild to moderate ulcerative colitis based on red density. Gut. 2020; 69(10):1778-1786. doi: 10.1136/gutjnl-2019-320056.

[101]Gottlieb K, Requa J, Karnes W, Chandra Gudivada R, Shen J, Rael E, et al. Central Reading of Ulcerative Colitis Clinical Trial Videos Using Neural Networks. Gastroenterology. 2021; 160(3):710-719.e2. doi: 10.1053/j.gastro.2020.10.024.

[102]Soffer S, Kopylov U, Klang E. Artificial Intelligence for the Evaluation of Mucosal Healing in IBD: The Future is Here. Gastroenterology. 2021; 161(3):1073-1074. doi: 10.1053/j.gastro.2020.12.052.

[103]Zhou J, Wu L, Wan X, Shen L, Liu J, Zhang J, et al. A novel artificial intelligence system for the assessment of bowel preparation (with video). Gastrointestinal Endoscopy. 2020; 91(2):428-435.e2. doi: 10.1016/j.gie.2019.11.026.

[104]Zhou W, Yao L, Wu H, Zheng B, Hu S, Zhang L, et al. Multi-step validation of a deep learning-based system for the quantification of bowel preparation: a prospective, observational study. The Lancet. Digital Health. 2021; 3(11):e697-e706. doi: 10.1016/S2589-7500(21)00109-6.

[105]Shaukat A, Rector TS, Church TR, Lederle FA, Kim AS, Rank JM, et al. Longer Withdrawal Time Is Associated With a Reduced Incidence of Interval Cancer After Screening Colonoscopy. Gastroenterology. 2015; 149(4):952-7. doi: 10.1053/j.gastro.2015.06.044.

[106]Gong D, Wu L, Zhang J, Mu G, Shen L, Liu J, et al. Detection of colorectal adenomas with a real-time computer-aided system (ENDOANGEL): a randomised controlled study. The Lancet. Gastroenterology & Hepatology. 2020; 5(4):352-61. doi: 10.1016/S2468-1253(19)30413-3.

Download PDF

Download XML

How to Cite

Zhang, C., and L. Wu. “The Application of Artificial Intelligence in Gastrointestinal Endoscopy: A State-of-the-Art Review”. Journal of Digital Health, vol. 1, no. 1, Feb. 2022, pp. 3-18, doi:10.55976/jdh.120221423-18.

Download Citation

Issue

Journal of Digital Health, Volume 1 Issue 1, 2022

Section

Review

License

This work is licensed under a Creative Commons Attribution 4.0 International License.

Copyright licenses detail the rights for publication, distribution, and use of research. Open Access articles published by Luminescience do not require transfer of copyright, as the copyright remains with the author. In opting for open access, the author(s) should agree to publish the article under the CC BY license (Creative Commons Attribution 4.0 International License). The CC BY license allows for maximum dissemination and re-use of open access materials and is preferred by many research funding bodies. Under this license, users are free to share (copy, distribute and transmit) and remix (adapt) the contribution, including for commercial purposes, providing they attribute the contribution in the manner specified by the author or licensor.

Home /

Archives /

Journal of Digital Health, Volume 1 Issue 1, 2022 /

Review

The application of artificial intelligence in gastrointestinal endoscopy: a state-of-the-art review

Abstract

Introduction

How to Cite

Issue

Section

License

Follow Luminescience

Further Information

Home / Archives / Journal of Digital Health, Volume 1 Issue 1, 2022 / Review

The application of artificial intelligence in gastrointestinal endoscopy: a state-of-the-art review

Abstract

Introduction

How to Cite

Issue

Section

License

Follow Luminescience

Further Information

Home /

Archives /

Journal of Digital Health, Volume 1 Issue 1, 2022 /

Review