Development and application of an artificial intelligence-assisted endoscopy system for diagnosis of Helicobacter pylori infection: a multicenter randomized controlled study

Development and application of an artificial intelligence-assisted endoscopy system for diagnosis of Helicobacter pylori infection: a multicenter randomized controlled study | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Development and application of an artificial intelligence-assisted endoscopy system for diagnosis of Helicobacter pylori infection: a multicenter randomized controlled study Pei-Ying Zou, Jian-Ru Zhu, Zhe Zhao, Hao Mei, Jing-Tao Zhao, Wen-Jing Sun, and 4 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-3747640/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 30 Sep, 2024 Read the published version in BMC Gastroenterology → Version 1 posted 3 You are reading this latest preprint version Abstract Background The early diagnosis and treatment of Heliobacter pylori gastrointestinal infection provide significant benefits to patients. We constructed a convolutional neural network (CNN) model based on an endoscopic system to diagnose H. pylori infection, and then examined the potential benefit of this model to endoscopists in their diagnosis of H. pylori infection. Materials and Methods A CNN neural network system for endoscopic diagnosis of H. pylori infection was established by collecting 7377 endoscopic images from 639 patients. The accuracy, sensitivity, and specificity were determined. Then, a randomized controlled study was used to compare the accuracy of diagnosis of H. pylori infection by endoscopists who were assisted or unassisted by this CNN model. Results The deep CNN model for diagnosis of H. pylori infection had an accuracy of 89.6%, a sensitivity of 90.9%, and a specificity of 88.9%. Relative to the group of endoscopists unassisted by AI, the AI-assisted group had better accuracy (92.8% [194/209; 95%CI: 89.3%, 96.4%] vs. 75.6% [158/209; 95%CI: 69.7%, 81.5%]), sensitivity (91.8% [67/73; 95%CI: 85.3%, 98.2%] vs. 78.6% [44/56; 95%CI: 67.5%, 89.7%]), and specificity (93.4% [127/136; 95%CI: 89.2%, 97.6%] vs. 74.5% [114/153; 95%CI: 67.5%, 81.5%]). All of these differences were statistically significant ( P < 0.05). Conclusion Our AI-assisted system for diagnosis of H. pylori infection has good diagnostic ability, and can improve the accuracy of endoscopists in gastroscopic diagnosis. Helicobacter pylori Endoscopy Artificial Intelligence Convolutional Neural Network Figures Figure 1 Figure 2 Figure 3 1. Introduction Helicobacter pylori is a microaerobic Gram-negative bacillus that can cause gastrointestinal diseases, such as chronic active gastritis, peptic ulcer, gastric mucosa-associated lymphoid tissue lymphoma, and gastric cancer. About 50% of people worldwide have H. pylori infections. Early diagnosis of H. pylori infection and eradication are important for treating chronic active gastritis and reducing the recurrence of peptic ulcer and the occurrence of gastric cancer [1–3] . Invasive and non-invasive methods are currently used to diagnose H. pylori infection. The main invasive methods are the Rapid Urease Test (RUT), histopathological diagnosis, endoscopic diagnosis, and molecular analysis; the main non-invasive methods are the Urea Breath Test (UBT), Stool Antigen Test, and serological examination [4–6] . In China, the cost of an endoscopic examination is relatively low and this procedure is widely accepted by patients. China also has a high rate of infection by H. pylori and a high incidence of gastric cancer, so if two tests can be combined and completed simultaneously using gastroscopy, H. pylori infection can be treated in a timely manner, precancerous lesions of gastric cancer and early gastric cancer can be detected, and medical costs can be reduced. Although invasive methods using gastroscopy provide these benefits, there are also has some shortcomings, in that the RUT is susceptible to false negatives depending on the load and focal distribution of H. pylori . However, histopathological and molecular biological tests cannot provide real-time results. After H. pylori infection, endoscopy can identify the unique characteristics of the gastric mucosa, and enable the diagnosis of H. pylori infection [8] . Gastric mucosa that is positive for H. pylori typically has diffuse redness, spotty redness, mucosal swelling, with or without white turbid mucus, atrophy, intestinal metaplasia, disappearance of regular arrangement of collecting venules (RAC), nodular changes, xanthoma, and hyperplastic polyps. Gastric mucosa that is negative for H. pylori typically has a RAC, and may have fundic gland polyposis, a red streak in the gastric antrum and gastric body, and attachment of a former bleeding spot. After H. pylori eradication, there are often localized red areas of various sizes on white mucosa, and redness after erosion healing appears as patchy redness. A previous study that used receiver operating characteristic (ROC) analysis found that use of conventional gastroscopy with white light imaging (WLI) for the diagnosis of H. pylori infection had an area under the curve (AUC) of 0.783, with a total coincidence rate of 78.46%, a sensitivity of 77.44%, a specificity of 79.31%, a positive predictive value (PPV) of 75.74%, and a negative predictive value (NPV) of 80.83% [9] . This previous study was based on the endoscopic observations of 2 experienced endoscopists (each with more than 5 years of endoscopy experience and examination of more than 5000 cases using gastroscopy). This previous study also reported that the accuracy of diagnosis depended on the experience of the endoscopist, and that diagnosis was time-consuming and adversely affected by fatigue. Therefore, objective, accurate, fast, and feasible endoscopic diagnosis of H. pylori is needed in clinical practice. Several studies reported that artificial intelligence (AI) provided accurate and standardized endoscopic diagnosis of H. pylori infection [9–12] . Convolutional neural networks (CNNs) are commonly used for analysis of medical images. CNNs have been widely applied in gastroenterology, and are currently the most widely used network architecture for medical image deep learning [13] . The present study established an endoscopic system for the diagnosis of H. pylori infection based on deep learning and applied it to a clinical setting. In particular, we compared the accuracy of gastroscopic diagnosis of H. pylori infection by endoscopists who were assisted by AI with diagnosis by endoscopists who were unassisted by AI, and then analyzed the possible clinical applications of this AI diagnostic system. 2. Methods 2.1 Study design and patients This study consisted of a diagnostic study and a multicenter randomized controlled study. Firstly, an endoscopic system for the diagnosis of H. pylori infection based on a deep learning CNN model was constructed and tested. Then, the diagnostic performance of AI-assisted and AI-unassisted endoscopists were compared. The main indicators were accuracy, sensitivity and specificity, and the secondary indicators were PPV and NPV (FIGURE 1 ). From September 2020 to July 2021, 1258 patients were screened at two institutions (Department of Gastroenterology of Chongqing Daping Hospital, Chongqing 13th People's Hospital and Chongqing Jiulongpo District Second People's Hospital). All included patients signed written informed consent documents. The inclusion criteria were: ( i ) age of 18 to 70 years old, male or female; ( ii ) completion of two or more diagnostic tests (UBT, RUT, H. pylori culture, H. pylori histology) with consistent results; and ( iii) more than one year after H. pylori treatment. The exclusion criteria were: ( i ) use of an antibiotic, bismuth, histamine H 2 -receptor antagonist, proton pump inhibitor (PPI), or probiotic during the 4 weeks before testing; ( ii ) women who were planning to become pregnant, or were pregnant or breastfeeding; ( iii ) use of an adrenocorticosteroid, nonsteroidal anti-inflammatory drug, or anticoagulant; ( iv ) pre-existing serious underlying disease (liver disease, cardiovascular disease, lung disease, kidney disease, metabolic disease, mental illness, malignant tumor, etc.) that made it difficult to tolerate gastroscopy; ( v ) previous history of gastric or duodenal ulcer, gastric malignant tumor, or other gastric organic lesions; ( vi ) participation in another clinical study within 3 months before participating in the present study; and ( vii ) difficulty in completing follow-up or other factors that affect compliance. 2.2 Endoscopy equipment Fourteen endoscopists performed the esophagogastroduodenoscopies (EGDs) using standard EGD equipment (CLV-290SL, CV-260SL, GIF-H260, GIF-H260Z, GIF-Q260J, GIF-Q260, GIF-H290, GIF-H290Z; Olympus Medical Systems). 2.3 Development data Two endoscopists collected gastric images of 639 patients, 260 who were H. pylori- positive and 379 who were H. pylori -negative (Table 1 ). For each patient, there was at least one image of each of the five anatomical sites in the stomach (upper body, middle body, lower body, lesser curvature, and antrum), and there were five or more total images per patient. Representative images were selected to establish the CNN training model, and the model was then trained and constructed. Finally, 7377 images from these 639 patients were used for model learning. 2.4 Test data To evaluate the diagnostic accuracy of the AI system, a separate set of test data was prepared. These test data consisted of 2080 images of 201 patients, 66 who were H. pylori- positive and 135 who were H. pylori- negative (Table 1 ). All of these patients received EGDs in the Department of Gastroenterology, Daping Hospital of Chongqing from January to March 2021. There was no overlap of patients between the development and test data sets. 2.5 Multicenter randomized controlled trial Endoscopists were randomized to an AI-assisted group or an AI-unassisted group, each with 14 endoscopists. Three endoscopists in each group were trained in the early detection of cancer at classes in the Shanghai Jiaotong University and obtained a training certificate; the 11 other endoscopists in each group were not trained in the early detection of cancer. Each group completed endoscopy independently in AI-assisted or AI-unassisted mode, as appropriate. After the examination, the AI-assisted group received the diagnostic results in real time, and had the opportunity to make a final judgment based on these AI results. Endoscopists in the AI-unassisted group made judgments of H. pylori infection after examination. There was no overlap of patients at any stage. This study was approved by the Ethics Committee of Daping Hospital and was registered with the Chinese Clinical Trial Registration Center ( www.chictr.org.cn ; registration number: ChiCTR2000037801). 2.6 Training algorithm To establish a system with high accuracy and diagnostic performance for the detection of H. pylori infection, EfficientNet ( https://arxiv.org/pdf/1905.11946.pdf ), a CNN developed by Google Research, was used. Several previously proposed dimensions of model zooming were investigated: network depth, network width, and image resolution. Previous research mostly enlarged one of these dimensions to achieve high accuracy. For example, ResNet-18 and ResNet-152 improve accuracy by increasing the network depth. Through the optimal combination of scaling these three dimensions, an EfficientNet neural network model can consider speed and accuracy. A series of EfficientNet models were obtained by amplifying the basic model. This series of models out-performed all previous CNN models in efficiency and accuracy. A state-of-the-art deep neural network architecture, EfficientNet-B0, was used for this procedure. The model had 237 layers and 5,330,564 parameters, and 5,288,548 parameters were needed for gradient descent in training. The core structure of the network was the Mobile Inverted Bottleneck Convolution (MBConvBlock) module, which also uses the Squeeze-and-Excitation Network (SENet). When SENet was proposed, it achieved the highest accuracy on Imagenet data at that time. The deep CNN uses back propagation to train the model. All layers of the network were fine-tuned using AdamW ( https://arxiV.org/pdf/1711.05101.pdf ), a method for stochastic optimization with a global learning rate of 0.01 and decay to one-tenth every 30 epochs. To optimize images for EfficientNet-B0, they were resized to 512×512 pixels. 2.7 Statistical analysis Estimation of sample size The purpose was to verify the noninferiority of AI-assisted endoscopists compared with AI-unassisted endoscopists in the diagnosis of H. pylori infection. In the test data, AI diagnostic accuracy was 89.6%, and the accuracy of the endoscopists was 82.4% [14] . Based on a noninferiority margin of 10%, an α of 0.05, and a 1 − β of 0.90, the estimated sample size was 84 per group. Noninferiority would be inferred if the 95% lower confidence boundary for the difference between the two groups in accuracy was more than 10%. The sample size was expanded using PASS version 11,with a total of 418 patients in the final two groups. Analysis of outcomes Numerical data were expressed as means ± SDs. The accuracy, specificity, sensitivity, PPV, and NPV from the different analyses were compared using the chi-square test or Fisher's exact test, as appropriate. All statistical calculations were performed using IBM SPSS version 23. A p-value below 0.05 was considered statistically significant. The trained CNN generated a continuous number for the probability of H. pylori infection (range: 0 to 1.0), and this was used as a continuous variable in the ROC analysis. The AUC values of the H. pylori -positive and H. pylori- negative groups were compared and plotted using R language version 4.01. 3. Results 3.1 Development data From September 2020 to December 2020, 639 patients (260 H. pylori- positive, 379 H. pylori- negative; 289 males, 350 females) received EGDs in the endoscopy center of the Department of Gastroenterology, Daping Hospital of Chongqing (Table 1 ). For every patient, there was at least one image from each of five sites in the stomach (upper body, middle body, lower body, lesser curvature, and antrum). We used 7377 images from 639 patients (development data set) to construct the CNN. 3.2 Performance of the CNN From January 2021 to March 2021, 201 different patients (66 H. pylori -positive, 135 H. pylori- negative; 90 males, 111 females) received EGDs at the same institution and these images were used for the test data set (Table 1 ). For the diagnosis of H. pylori infection, the CNN model had an accuracy of 89.6%, a sensitivity of 90.9%, and a specificity of 88.9%. The AUC for H. pylori positivity was 84.1% (95%CI: 73.0%, 95.2%) and the AUC for H. pylori negativity was 90.3% (95%CI: 82.2%, 98.4%; FIGURE 2 ). We also performed separate analyses of diagnosis based on the gastric body alone (upper body, middle body, lower body, and lesser curvature) and the antrum alone (Table 2 ). Analysis of the gastric body provided significantly greater sensitivity (86.2% vs. 19.8%) and accuracy (76.4% vs. 68.4%), but analysis of the antrum provided significantly greater specificity (93.3% vs. 71.7%); analysis of the antrum provided a significantly lower false-positive rate (6.7% vs. 28.3%), but analysis of the body provided a significantly lower false-negative rate (13.8% vs. 80.2%). Combining data from the body and antrum provided the best diagnostic performance. Table 2 Sensitivity, specificity, accuracy, false-positive rate, and false-negative rate for diagnosis of H. pylori infection based on examination of the body and the antrum. Diagnostic parameter Body* Antrum P value Sensitivity 86.2% 19.8% < 0.001 Specificity 71.7% 93.3% < 0.001 Accuracy 76.4% 68.4% < 0.001 False positive rate 28.3% 6.7% < 0.001 False negative rate 13.8% 80.2% < 0.001 *Stomach body consists of the upper body, middle body, lower body, and lesser curvature. 3.3 Multicenter randomized controlled study From April 2021 to July 2021, we enrolled 418 patients from three institutions (Department of Gastroenterology of Chongqing Daping Hospital, The 13th People's Hospital of Chongqing, and The Second People's Hospital of Jiulongpo District of Chongqing) for a multicenter randomized controlled trial of endoscopists. There were 209 patients in the AI-assisted group and 209 other patients in the AI-unassisted group (Table 3 ), and these two groups had no significant differences at baseline. Table 3 Baseline demographic and clinical characteristics of patients examined in the multicenter randomized controlled study of endoscopists. Characteristic AI-assisted AI-unassisted p -value Sex Male, n (%) Female, n (%) 92/209(44.0) 117/209(56.0) 81/209(38.7) 128/209(61.3) 0.275 Mean age, years (SD) 46.3 ± 12.1 45.7 ± 12.4 0.652 Mean BMI (SD) 22.7 ± 3.4 23.2 ± 2.9 0.134 H. pylori status, n (%) Positive Negative 73/209(34.9) 136/209(65.0) 56/209(26.7) 153/209(73.2) 0.072 Cigarette smoking, n (%) 33/209(15.7) 28/209(13.3) 0.488 Alcohol drinking, n (%) 53/209(25.3) 47/209(22.4) 0.492 Family history of esophageal or gastric carcinoma, n (%) 12/209(5.7) 18/209(8.6) 0.256 Biopsy cases, n (%) 26/209(12.4) 31/209(14.8) 0.476 Atrophy cases, n (%) 61/209(29.1) 64/209(30.6) 0.749 Comparisons of all 14 endoscopists in the AI-assisted and AI-unassisted groups indicated the AI-assisted group had significantly better accuracy (92.8% vs. 75.6%, P < 0.001), sensitivity (91.8% vs. 78.6%, P = 0.032), and specificity (93.4% vs. 74.5%, P < 0.001; Table 4 ). Table 4 Sensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV) of endoscopists with and without AI assistance in the diagnosis of H. pylori infection. AI-assisted Total AI-unassisted Total Diagnostic parameter Untrained* (n = 11) Trained** (n = 3) 14 Untrained* (n = 11) Trained** (n = 3) 14 P value *** Sensitivity 91.6%(44/48) 92.0%(23/25) 91.8%(67/73) 76.3%(29/38) 83.3%(15/18) 78.6%(44/56) 0.032 Specificity 91.9%(103/112) 100%(24/24) 93.4%(127/136) 70.6%(82/116) 86.4%(32/37) 74.5%(114/153) < 0.001 Accuracy 91.9%(147/160) 95.9%(47/49) 92.8%(194/209) 72.1%(111/154) 85.5%(47/55) 75.6%(158/209) < 0.001 PPV 83.0%(44/53) 100%(23/23) 88.2%(67/76) 46.0%(29/63) 75.0%(15/20) 53.0%(44/83) < 0.001 NPV 96.2%(103/107) 92.3%(24/26) 95.5%(127/133) 90.1%(82/91) 91.4%(32/35) 90.5%(114/126) 0.113 * Not trained for early cancer screening ** Trained for early cancer screening *** Comparison of all AI-assisted vs. all AI-unassisted We then determined the accuracy of endoscopists in the AI-unassisted group who had different levels of experience, using two different metrics to define “junior” and “senior” status (FIGURE 3 ). The accuracy of diagnosis was 85.5% (47/55) for senior endoscopists who were trained in early cancer detection, and was 72.1% (111/154) for junior endoscopists who did not have this training ( P = 0.047). The accuracy of diagnosis was 77.3% (75/97) for senior endoscopists who had 5 years of experience or examination of more than 5000 cases, and was 73.2% (82/112) for junior endoscopists who did not have this experience ( P = 0.494). 4. Discussion The development of endoscopic technology has made it possible to identify H. pylori using EGD. However, some current guidelines do not recommend endoscopy for the routine diagnosis of H. pylori infection because these methods require special equipment, endoscopists must receive relevant training, and the accuracy and specificity may differ among endoscopists [23] . Notably, developments in AI may resolve some of the problems related to the endoscopic diagnosis of H. pylori infection. AI methods can learn rules from sample data, and can recognize different types of data, such as text, images, and sounds. The image recognition used in the present study is based on deep learning and is a key technology for AI-assisted endoscopic diagnosis of H.pylori infection [12, 13] . Recent studies showed that AI endoscopic diagnosis of H.pylori infection is possible. In particular, a meta-analysis of AI-based diagnosis of H.pylori infection from endoscopy showed that the sensitivity was 0.87 (95%CI: 0.72, 0.94), the specificity was 0.86 (95%CI: 0.77, 0.92), and the AUC was 0.92 (95%CI: 0.90, 0.94) [15] . However, most studies of this topic were retrospective and diagnostic, and there have been no clinical prospective studies [15,19–22] . In the present study, we combined a diagnostic study and multicenter randomized controlled study of endoscopists. Our results confirmed the clinical value of AI-assisted endoscopic diagnosis of H. pylori infection. In the present study, the endoscopic diagnostic model of H. pylori infection was based on deep learning and was constructed using endoscopic images of five stomach regions. In the test data, AI diagnosis that was based on images of the gastric body (upper body, middle body, lower body, and lesser curvature) provided better accuracy, sensitivity, and PPV than AI diagnosis based on images of the antrum; however, images of the antrum provided better specificity and NPV than images of the body. The combined use of body and antrum images had an accuracy of 89.6%, a sensitivity of 90.9%, and a specificity of 88.9%. A 2020 Japanese study found that diagnostic accuracy was greater when using the lesser curvature of the middle-upper body than the fornix and the greater curvature of the middle-upper body [17,18] . A 2019 study in China found that the sensitivity, specificity, and accuracy from using multiple gastric images (average 8.3 ± 3.3 per patient) for AI diagnosis of H. pylori infection were greater than those from using a single image [18] . This led us to use an AI-assisted endoscopy diagnostic system for H. pylori infection that is based on images of multiple sites for model establishment. Most previous endoscopy studies have used WLI, and only a few have used newer endoscopic techniques. A 2018 Japanese study compared the effectiveness of using AI with WLI, Blue Laser Imaging (BLI), and Linked Color Imaging (LCI) for the diagnosis of H. pylori infection, and found that the AUC was higher for BLI (0.96) and LCI (0.95) than for WLI (0.66, both P < 0.01) [16] . Although these new techniques seem to have advantages over WLI in the diagnosis of H. pylori infection, they have not been widely adopted because they require additional expensive equipment and professionally trained endoscopists. Therefore, an H. pylori diagnostic system developed for WLI is more suitable for widespread acceptance, especially in regions where resources are limited. In addition, some of the limitations of WLI can be overcome by use of AI learning from images at multiple sites in the stomach. The present study was the first to examine an AI system for the endoscopic diagnosis of H. pylori infection in a multicenter randomized controlled clinical trial of endoscopists. We found that endoscopists with AI assistance had significantly better accuracy, sensitivity, and specificity than unassisted endoscopists (all P < 0.05). This indicates that our AI endoscopic system for the diagnosis of H. pylori infection significantly improved the diagnostic ability of endoscopists when using gastroscopy. Our analysis of endoscopists also found the number of working years and the number of procedures that were performed had no significant effect on their diagnostic performance. However, endoscopists who were trained in early cancer detection had greater accuracy than endoscopists without this training (85.5% [47/55] vs. 72.1% [111/154], P = 0.047). This may be because in the professional training for early cancer detection, the recognition of H. pylori status of gastric mucosa using endoscopy is very important for determining the mucosal background of gastric cancer, so these specially trained endoscopists are familiar with the manifestations of H. pylori infection and pay close attention to these observations. Endoscopists not trained in early cancer detection may have lower accuracy, even though they are skilled in endoscopy procedures. This further emphasizes the clinical value of the using an AI system for the endoscopic diagnosis of H. pylori infection. The UBT is the most common method for the clinical detection of H. pylori infection. However, when the results of this test are close to the test threshold, diagnosis can be difficult. The present study showed that an AI endoscopy system can assist doctors whose UBT results are near the threshold. In particular, among the 418 patients, 6 patients had UBT results near the critical value, and 5 of them were accurately assessed by the AI system, corresponding to an accuracy of 83.3% and a sensitivity of 100% (data not shown). However, this result is from only 6 patients, so a larger sample size is needed for further study of this topic. The status of stomach infection by H. pylori may be classified as “infected”, “uninfected”, or “eradicated”, with the latter two states considered negative. We found it was difficult to identify the eradicated state using endoscopy, and identification of this state was also affected by the time since eradication. In particular, among the 418 patients, 369 were not treated for H. pylori infection, and the accuracy of diagnosis in these patients was 85.9%, greater than the accuracy in patients treated for H. pylori infection (71.4%, data not shown). This result indicates that prior treatment for H. pylori infection adversely affected diagnosis by doctors alone and by doctors using AI. This may be because of decreased gastric mucosal inflammation and atypical gastric mucosal appearance after the use of antibiotics and PPIs. We also found better accuracy (81.0% vs. 65.4%), specificity (83.3% vs. 68.1%), and sensitivity (66.7% vs. 50.0%) in patients who received treatment more than two years ago rather than less than 2 years ago (data not shown). We speculate that as the time after treatment increases, there were greater declines in inflammation of the gastric mucosa,so endoscopic gastric mucosa becomes more typical [24–27] . For example, the diffuse redness from the gastric body to the gastric fundus disappears, and some of these individuals even show RAC again, so the accuracy of endoscopic diagnosis improved accordingly. There are some limitations in this study. First, our development and test data sets were all from a single center. Second, our sample size was rather small, and a larger sample should be used in subsequent studies. Finally, this study did not use the classification system of “positive”, “negative”, and “negative after eradication”, a topic that is also worthy of further study. In conclusion, we established an AI-assisted endoscopy system for the diagnosis of H. pylori infection that was based on AI learning of images from five sites in the stomach, and applied this system to the first clinical trial of this topic in China. The results of this multicenter randomized controlled trial verified that the AI system described here improved the ability of endoscopists to diagnose H. pylori infection using gastroscopy. Because this system provided greater improvements to endoscopists who were not trained in early cancer detection, we believe it may be beneficial for geographic regions that have limited resources, high incidences of gastric cancer, and common use and acceptance of gastroscopy. Declarations CONFLICT OF INTEREST There are no conflicts of interest to declare. AUTHOR CONTRIBUTIONS All authors contributed to the study conception and design. The first three authors contributed equally to this work. Material preparation, data collection, and data analysis were performed by Pei-Ying Zou, Jian-Ru Zhu, Zhe Zhao, Hao Mei, Jing-Tao Zhao, Wen-Jing Sun, Li-Lin Fan, and Guo-Hua Wang. The first draft of the manuscript was written by Pei-Ying Zou. Chun-Hui Lan critically reviewed the manuscript. All the authors commented on previous versions of the manuscript and approved the final manuscript. CONSENT FOR PULICATION Not applicable. ACKNOWLEDGEMENTS This study was supported by grants from the Science and Technology Innovation Enhancement Project of Army Medical University (No. 20I9CXLCB003) and the National Science Foundation of China (No. 82072253). This study was approved by the Ethics Committee of Daping Hospital (10/07/2020)（No.89,2020）and was registered with the Chinese Clinical Trial Registration Center (02/09/2020) (www.chictr.org.cn; registration number: ChiCTR2000037801). VAILABILITY OF DATA AND MATERIALS All data generated or analysed during this study are included in this published article. References Hooi, J.K.Y., et al., Global Prevalence of Helicobacter pylori Infection: Systematic Review and Meta-Analysis. Gastroenterology, 2017. 153(2): p. 420–429. Savoldi, A., et al., Prevalence of Antibiotic Resistance in Helicobacter pylori: A Systematic Review and Meta-analysis in World Health Organization Regions. Gastroenterology, 2018. 155(5): p. 1372–1382.e17. Ndip, R.N., et al., Helicobacter pylori isolates recovered from gastric biopsies of patients with gastro-duodenal pathologies in Cameroon: current status of antibiogram. Tropical Medicine & International Health, 2008. 13(6): p. 848–854. Huh, C.W. and B.W. Kim, [Diagnosis of Helicobacter pylori Infection]. Korean J Gastroenterol, 2018. 72(5): p. 229–236. Wang, Y.K., et al., Diagnosis of Helicobacter pylori infection: Current options and developments. World J Gastroenterol, 2015. 21(40): p. 11221-35. Makristathis, A., et al., Review: Diagnosis of Helicobacter pylori infection. Helicobacter, 2019. 24 Suppl 1: p. e12641. Chen Huang wei., et al., The application value of Kyoto gastritis classification in the direct judgment of Helicobacter pylori infection under white light gastroscopy. New Medicine, 2019. 50(6): 457–462 Kato Motoji et al., Classification of gastritis by Kyoto. Stomach and Intestine, 2019. 54(5):616–619. Kato T,Yagi N,Kamada T,Shimbo T,Watanabe H,Ida K;Diagnosis of Helicobacter pylori infection in gastric mucosa by endoscopic features:a multicenter prospective Endosc,2013,25(5):508–518. Cho JH,Chang YW,Jang JY,Shim.JJ,Lee CK,Dong SH,Kim HJ,Kim BH,Lee TH,Cho JY.Close observation of gastric mucosal pattern by standard endoscopy can predict Helicobacter pylori infection status.J Gastroenterol Hepatol,2013,28(2 4. Watanabe K,Nagata N,Nakashima R,Furuhata E,Shimbo T,Kobayakawa M,Sakurai T,Imbe K,Niikura R,Yokoi C J,Uemura N.Predictive findings for Helicobacter pylori-uninfected,-infected and-eradicated gastric mucosa:validation orld J Gastroenterol,2013,19(27):4374–4379. Esteva, A., et al., Dermatologist-level classification of skin cancer with deep neural networks. Nature, 2017. 542(7639): p. 115–118. Ribeiro E, Uhl A, Wimmer G, Häfner M. Exploring deep learning and transfer learning for colonic polyp classification. Comput Math Methods Med 2016;2016;6584725. Shichijo, S., et al., Application of Convolutional Neural Networks in the Diagnosis of Helicobacter pylori Infection Based on Endoscopic Images. EBioMedicine, 2017. 25: p. 106–111. Bang, C.S., J.J. Lee and G.H. Baik, Artificial Intelligence for the Prediction of Helicobacter Pylori Infection in Endoscopic Images: Systematic Review and Meta-Analysis Of Diagnostic Test Accuracy. J Med Internet Res, 2020. 22(9): p. e21983. Nakashima, H., et al., Artificial intelligence diagnosis of Helicobacter pylori infection using blue laser imaging-bright and linked color imaging: a single-center prospective study. Ann Gastroenterol, 2018. 31(4): p. 462–468. Yasuda, T., et al., Potential of automatic diagnosis system with linked color imaging for diagnosis of Helicobacter pylori infection. Dig Endosc, 2020. 32(3): p. 373–381. Zheng, W., et al., High Accuracy of Convolutional Neural Network for Evaluation of Helicobacter pylori Infection Based on Endoscopic Images: Preliminary Experience. Clin Transl Gastroenterol, 2019. 10(12): p. e00109. Shichijo S, Endo Y, Aoyama K, Takeuchi Y, Ozawa T, Takiyama H, et al. Application of convolutional neural networks for evaluating Helicobacter pylori infection status on the basis of endoscopic images. Scand J Gastroenterol 2019 Feb;54(2):158–163. Itoh T, Kawahira H, Nakashima H, Yata N. Deep learning analyzes Helicobacter pylori infection by upper gastrointestinal endoscopy images. Endosc Int Open 2018 Feb;6(2):E139-E144 [FREE Full text] [doi: 10.1055/s-0043-120830] [Medline: 29399610]. Huang C, Chung P, Sheu B, Kuo H, Popper M. Helicobacter pylori-related gastric histology classification using support-vector-machine-based feature selection. IEEE Trans Inf Technol Biomed 2008 Jul;12(4):523–531. Huang C, Sheu B, Chung P, Yang H. Computerized diagnosis of Helicobacter pylori infection and associated gastric inflammation from endoscopic images by refined feature selection using a neural network. Endoscopy 2004 Jul;36(7):601–608. Liu Wenzhong et al., The fifth National Consensus report on the management of Helicobacter pylori infection. Gastroenterology, 2017. 22(06): 346–360. Yumin Zhang, Dadao Jing and Yulan Qiu, Effects of Helicobacter pylori eradication on gastric mucosal precancerous lesions: a meta-analysis. Journal of Clinical Gastroenterology, 2009. 21(05): 268–272. Sung JJ‚ Lin SR‚Ching JY‚et al༎ Atrophy and intestinal metaplasia one year after cure of H.pylori infection: a prospective‚ randomized study༎Gastroenterology‚2000‚119༚7–14༎ 4 Correa P‚Fontham ET‚Bravo JC‚et al༎Chemoprevention of gastric dysplasia: randomized trial of antioxidants supplements and ant-i Helicobacter pylori therapy༎J Natl Cancer Inst‚2000‚92༚ 1881–1888༎ Khan, M.Y., et al., Effectiveness of Helicobacter pylori eradication in preventing metachronous gastric cancer and preneoplastic lesions. A systematic review and meta-analysis. Eur J Gastroenterol Hepatol, 2020. 32(6): p. 686–694. Tables Table 1 is available in the Supplementary Files section. Additional Declarations No competing interests reported. Supplementary Files Table1.docx Cite Share Download PDF Status: Published Journal Publication published 30 Sep, 2024 Read the published version in BMC Gastroenterology → Version 1 posted Editor invited by journal 16 Jan, 2024 Submission checks completed at journal 16 Jan, 2024 First submitted to journal 13 Dec, 2023 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-3747640","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":267364405,"identity":"82af620a-f7e3-4c9c-91bc-4191b86e72e4","order_by":0,"name":"Pei-Ying Zou","email":"","orcid":"","institution":"Daping Hospital","correspondingAuthor":false,"prefix":"","firstName":"Pei-Ying","middleName":"","lastName":"Zou","suffix":""},{"id":267364406,"identity":"8de1f340-4b46-435d-a186-69a5b7803a4f","order_by":1,"name":"Jian-Ru Zhu","email":"","orcid":"","institution":"Daping Hospital","correspondingAuthor":false,"prefix":"","firstName":"Jian-Ru","middleName":"","lastName":"Zhu","suffix":""},{"id":267364407,"identity":"fd66226b-9ea8-4e46-ab83-25725e5ad3e8","order_by":2,"name":"Zhe Zhao","email":"","orcid":"","institution":"Daping Hospital","correspondingAuthor":false,"prefix":"","firstName":"Zhe","middleName":"","lastName":"Zhao","suffix":""},{"id":267364408,"identity":"c441a225-2d7d-4522-a876-0fd3205e4e7a","order_by":3,"name":"Hao Mei","email":"","orcid":"","institution":"Daping Hospital","correspondingAuthor":false,"prefix":"","firstName":"Hao","middleName":"","lastName":"Mei","suffix":""},{"id":267364409,"identity":"5cc09847-bcf6-4cbb-b71b-ee8dfe260485","order_by":4,"name":"Jing-Tao Zhao","email":"","orcid":"","institution":"Daping Hospital","correspondingAuthor":false,"prefix":"","firstName":"Jing-Tao","middleName":"","lastName":"Zhao","suffix":""},{"id":267364410,"identity":"bb04b6d1-b5fa-4f78-a17f-0000bcee45b2","order_by":5,"name":"Wen-Jing Sun","email":"","orcid":"","institution":"Chongqing 13th People's Hospital","correspondingAuthor":false,"prefix":"","firstName":"Wen-Jing","middleName":"","lastName":"Sun","suffix":""},{"id":267364411,"identity":"3b79c96f-e9d0-4c04-978a-514af1061496","order_by":6,"name":"Guo-Hua Wang","email":"","orcid":"","institution":"ChongqingSkyforbioCo.","correspondingAuthor":false,"prefix":"","firstName":"Guo-Hua","middleName":"","lastName":"Wang","suffix":""},{"id":267364412,"identity":"33c2c8d4-1018-41b8-a545-522f1b536c7b","order_by":7,"name":"Dong-Feng Chen","email":"","orcid":"","institution":"Daping Hospital","correspondingAuthor":false,"prefix":"","firstName":"Dong-Feng","middleName":"","lastName":"Chen","suffix":""},{"id":267364413,"identity":"74d169b2-f841-427d-a0bf-fe68f52b7e9f","order_by":8,"name":"Li-Lin Fan","email":"","orcid":"","institution":"Chongqing Jiulongpo District Second People's Hospital","correspondingAuthor":false,"prefix":"","firstName":"Li-Lin","middleName":"","lastName":"Fan","suffix":""},{"id":267364414,"identity":"34e23643-7a86-4ca9-9485-2462e9458c8a","order_by":9,"name":"Chun-Hui lan","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA1ElEQVRIiWNgGAWjYBACPiBmBuIEBmbmA1CxBPxa2BBa2BIbSNTCwGNIpBb2swc/F7bdyTM4zvP90c2cwwz87DkGDD934NHCk5csPbPtWbHBYd6NzbnbDjNI9rwxYOw9g89hOWbMvG2HEzfAtBjcyDFgZmzDo4X/DUwLz0OwFnuCWiTgtvAwQmyRIKjljbE0z7nDiTMPsxnOzt2WziNx5lnBwV48Wvj5cww/85QdTuw7f/jB59xt1nL87ckbH/zEowUD8ICIAyRoGAWjYBSMglGABQAA661PRAGvakUAAAAASUVORK5CYII=","orcid":"","institution":"Daping Hospital","correspondingAuthor":true,"prefix":"","firstName":"Chun-Hui","middleName":"","lastName":"lan","suffix":""}],"badges":[],"createdAt":"2023-12-13 09:29:21","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-3747640/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-3747640/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1186/s12876-024-03389-3","type":"published","date":"2024-09-30T15:58:08+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":49889483,"identity":"64dbb514-005e-4ac6-9d9e-f14e8a81a7e1","added_by":"auto","created_at":"2024-01-19 19:48:47","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":48959,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePatient selection and study design\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-3747640/v1/fbdc0fe215a38d0cf7a4e582.png"},{"id":49889484,"identity":"f3ee6c5f-96c8-48c7-9387-2875fc61d54d","added_by":"auto","created_at":"2024-01-19 19:48:47","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":103854,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePositive diagnosis (left) and negative diagnosis (right) of \u003c/strong\u003e\u003cem\u003e\u003cstrong\u003eH. pylori\u003c/strong\u003e\u003c/em\u003e\u003cstrong\u003e infection using the CNN model.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-3747640/v1/cc4209616c858bf5fa93d654.png"},{"id":49889486,"identity":"e0d649ac-6d80-4e7d-956a-09788d80703a","added_by":"auto","created_at":"2024-01-19 19:48:47","extension":"jpeg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":211249,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAccuracy of endoscopists diagnosis of \u003c/strong\u003e\u003cem\u003e\u003cstrong\u003eH. pylori\u003c/strong\u003e\u003c/em\u003e\u003cstrong\u003e infection when using different criteria to define “junior” and “senior” endoscopists.\u003c/strong\u003e\u003c/p\u003e","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-3747640/v1/f3e1f5d9caa9cafd55ed3eb8.jpeg"},{"id":66096955,"identity":"05d9c086-f085-4c39-9200-61f76ec96eab","added_by":"auto","created_at":"2024-10-07 16:12:07","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1019714,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-3747640/v1/711e875c-d5ad-47ba-b532-986b1ba8e876.pdf"},{"id":49889886,"identity":"9fd6562e-0ad0-4cbc-a0ea-572f0ed25822","added_by":"auto","created_at":"2024-01-19 19:56:47","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":13400,"visible":true,"origin":"","legend":"","description":"","filename":"Table1.docx","url":"https://assets-eu.researchsquare.com/files/rs-3747640/v1/18c39876353a2f39f706bba8.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Development and application of an artificial intelligence-assisted endoscopy system for diagnosis of Helicobacter pylori infection: a multicenter randomized controlled study","fulltext":[{"header":"1. Introduction","content":"\u003cp\u003e \u003cem\u003eHelicobacter pylori\u003c/em\u003e is a microaerobic Gram-negative bacillus that can cause gastrointestinal diseases, such as chronic active gastritis, peptic ulcer, gastric mucosa-associated lymphoid tissue lymphoma, and gastric cancer. About 50% of people worldwide have \u003cem\u003eH. pylori\u003c/em\u003e infections. Early diagnosis of \u003cem\u003eH. pylori\u003c/em\u003e infection and eradication are important for treating chronic active gastritis and reducing the recurrence of peptic ulcer and the occurrence of gastric cancer\u003csup\u003e[1\u0026ndash;3]\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eInvasive and non-invasive methods are currently used to diagnose \u003cem\u003eH. pylori\u003c/em\u003e infection. The main invasive methods are the Rapid Urease Test (RUT), histopathological diagnosis, endoscopic diagnosis, and molecular analysis; the main non-invasive methods are the Urea Breath Test (UBT), Stool Antigen Test, and serological examination\u003csup\u003e[4\u0026ndash;6]\u003c/sup\u003e. In China, the cost of an endoscopic examination is relatively low and this procedure is widely accepted by patients. China also has a high rate of infection by \u003cem\u003eH. pylori\u003c/em\u003e and a high incidence of gastric cancer, so if two tests can be combined and completed simultaneously using gastroscopy, \u003cem\u003eH. pylori\u003c/em\u003e infection can be treated in a timely manner, precancerous lesions of gastric cancer and early gastric cancer can be detected, and medical costs can be reduced. Although invasive methods using gastroscopy provide these benefits, there are also has some shortcomings, in that the RUT is susceptible to false negatives depending on the load and focal distribution of \u003cem\u003eH. pylori\u003c/em\u003e. However, histopathological and molecular biological tests cannot provide real-time results.\u003c/p\u003e \u003cp\u003eAfter \u003cem\u003eH. pylori\u003c/em\u003e infection, endoscopy can identify the unique characteristics of the gastric mucosa, and enable the diagnosis of \u003cem\u003eH. pylori\u003c/em\u003e infection\u003csup\u003e[8]\u003c/sup\u003e. Gastric mucosa that is positive for \u003cem\u003eH. pylori\u003c/em\u003e typically has diffuse redness, spotty redness, mucosal swelling, with or without white turbid mucus, atrophy, intestinal metaplasia, disappearance of regular arrangement of collecting venules (RAC), nodular changes, xanthoma, and hyperplastic polyps. Gastric mucosa that is negative for \u003cem\u003eH. pylori\u003c/em\u003e typically has a RAC, and may have fundic gland polyposis, a red streak in the gastric antrum and gastric body, and attachment of a former bleeding spot. After \u003cem\u003eH. pylori\u003c/em\u003e eradication, there are often localized red areas of various sizes on white mucosa, and redness after erosion healing appears as patchy redness.\u003c/p\u003e \u003cp\u003eA previous study that used receiver operating characteristic (ROC) analysis found that use of conventional gastroscopy with white light imaging (WLI) for the diagnosis of \u003cem\u003eH. pylori\u003c/em\u003e infection had an area under the curve (AUC) of 0.783, with a total coincidence rate of 78.46%, a sensitivity of 77.44%, a specificity of 79.31%, a positive predictive value (PPV) of 75.74%, and a negative predictive value (NPV) of 80.83%\u003csup\u003e[9]\u003c/sup\u003e. This previous study was based on the endoscopic observations of 2 experienced endoscopists (each with more than 5 years of endoscopy experience and examination of more than 5000 cases using gastroscopy). This previous study also reported that the accuracy of diagnosis depended on the experience of the endoscopist, and that diagnosis was time-consuming and adversely affected by fatigue. Therefore, objective, accurate, fast, and feasible endoscopic diagnosis of \u003cem\u003eH. pylori\u003c/em\u003e is needed in clinical practice. Several studies reported that artificial intelligence (AI) provided accurate and standardized endoscopic diagnosis of \u003cem\u003eH. pylori\u003c/em\u003e infection\u003csup\u003e[9\u0026ndash;12]\u003c/sup\u003e. Convolutional neural networks (CNNs) are commonly used for analysis of medical images. CNNs have been widely applied in gastroenterology, and are currently the most widely used network architecture for medical image deep learning\u003csup\u003e[13]\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eThe present study established an endoscopic system for the diagnosis of \u003cem\u003eH. pylori\u003c/em\u003e infection based on deep learning and applied it to a clinical setting. In particular, we compared the accuracy of gastroscopic diagnosis of \u003cem\u003eH. pylori\u003c/em\u003e infection by endoscopists who were assisted by AI with diagnosis by endoscopists who were unassisted by AI, and then analyzed the possible clinical applications of this AI diagnostic system.\u003c/p\u003e"},{"header":"2. Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Study design and patients\u003c/h2\u003e \u003cp\u003eThis study consisted of a diagnostic study and a multicenter randomized controlled study. Firstly, an endoscopic system for the diagnosis of \u003cem\u003eH. pylori\u003c/em\u003e infection based on a deep learning CNN model was constructed and tested. Then, the diagnostic performance of AI-assisted and AI-unassisted endoscopists were compared. The main indicators were accuracy, sensitivity and specificity, and the secondary indicators were PPV and NPV (FIGURE \u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eFrom September 2020 to July 2021, 1258 patients were screened at two institutions (Department of Gastroenterology of Chongqing Daping Hospital, Chongqing 13th People's Hospital and Chongqing Jiulongpo District Second People's Hospital). All included patients signed written informed consent documents. The inclusion criteria were: (\u003cem\u003ei\u003c/em\u003e) age of 18 to 70 years old, male or female; (\u003cem\u003eii\u003c/em\u003e) completion of two or more diagnostic tests (UBT, RUT, \u003cem\u003eH. pylori\u003c/em\u003e culture, \u003cem\u003eH. pylori\u003c/em\u003e histology) with consistent results; and (\u003cem\u003eiii)\u003c/em\u003e more than one year after \u003cem\u003eH. pylori\u003c/em\u003e treatment. The exclusion criteria were: (\u003cem\u003ei\u003c/em\u003e) use of an antibiotic, bismuth, histamine H\u003csub\u003e2\u003c/sub\u003e-receptor antagonist, proton pump inhibitor (PPI), or probiotic during the 4 weeks before testing; (\u003cem\u003eii\u003c/em\u003e) women who were planning to become pregnant, or were pregnant or breastfeeding; (\u003cem\u003eiii\u003c/em\u003e) use of an adrenocorticosteroid, nonsteroidal anti-inflammatory drug, or anticoagulant; (\u003cem\u003eiv\u003c/em\u003e) pre-existing serious underlying disease (liver disease, cardiovascular disease, lung disease, kidney disease, metabolic disease, mental illness, malignant tumor, etc.) that made it difficult to tolerate gastroscopy; (\u003cem\u003ev\u003c/em\u003e) previous history of gastric or duodenal ulcer, gastric malignant tumor, or other gastric organic lesions; (\u003cem\u003evi\u003c/em\u003e) participation in another clinical study within 3 months before participating in the present study; and (\u003cem\u003evii\u003c/em\u003e) difficulty in completing follow-up or other factors that affect compliance.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Endoscopy equipment\u003c/h2\u003e \u003cp\u003eFourteen endoscopists performed the esophagogastroduodenoscopies (EGDs) using standard EGD equipment (CLV-290SL, CV-260SL, GIF-H260, GIF-H260Z, GIF-Q260J, GIF-Q260, GIF-H290, GIF-H290Z; Olympus Medical Systems).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Development data\u003c/h2\u003e \u003cp\u003eTwo endoscopists collected gastric images of 639 patients, 260 who were \u003cem\u003eH. pylori-\u003c/em\u003epositive and 379 who were \u003cem\u003eH. pylori\u003c/em\u003e-negative (Table \u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). For each patient, there was at least one image of each of the five anatomical sites in the stomach (upper body, middle body, lower body, lesser curvature, and antrum), and there were five or more total images per patient. Representative images were selected to establish the CNN training model, and the model was then trained and constructed. Finally, 7377 images from these 639 patients were used for model learning.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 Test data\u003c/h2\u003e \u003cp\u003eTo evaluate the diagnostic accuracy of the AI system, a separate set of test data was prepared. These test data consisted of 2080 images of 201 patients, 66 who were \u003cem\u003eH. pylori-\u003c/em\u003epositive and 135 who were \u003cem\u003eH. pylori-\u003c/em\u003enegative (Table \u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). All of these patients received EGDs in the Department of Gastroenterology, Daping Hospital of Chongqing from January to March 2021. There was no overlap of patients between the development and test data sets.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.5 Multicenter randomized controlled trial\u003c/h2\u003e \u003cp\u003eEndoscopists were randomized to an AI-assisted group or an AI-unassisted group, each with 14 endoscopists. Three endoscopists in each group were trained in the early detection of cancer at classes in the Shanghai Jiaotong University and obtained a training certificate; the 11 other endoscopists in each group were not trained in the early detection of cancer. Each group completed endoscopy independently in AI-assisted or AI-unassisted mode, as appropriate. After the examination, the AI-assisted group received the diagnostic results in real time, and had the opportunity to make a final judgment based on these AI results. Endoscopists in the AI-unassisted group made judgments of \u003cem\u003eH. pylori\u003c/em\u003e infection after examination. There was no overlap of patients at any stage. This study was approved by the Ethics Committee of Daping Hospital and was registered with the Chinese Clinical Trial Registration Center (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ewww.chictr.org.cn\u003c/a\u003e\u003c/span\u003e\u003cspan address=\"http://www.chictr.org.cn\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e; registration number: ChiCTR2000037801).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003e2.6 Training algorithm\u003c/h2\u003e \u003cp\u003eTo establish a system with high accuracy and diagnostic performance for the detection of \u003cem\u003eH. pylori\u003c/em\u003e infection, EfficientNet (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://arxiv.org/pdf/1905.11946.pdf\u003c/span\u003e\u003cspan address=\"https://arxiv.org/pdf/1905.11946.pdf\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e), a CNN developed by Google Research, was used. Several previously proposed dimensions of model zooming were investigated: network depth, network width, and image resolution. Previous research mostly enlarged one of these dimensions to achieve high accuracy. For example, ResNet-18 and ResNet-152 improve accuracy by increasing the network depth. Through the optimal combination of scaling these three dimensions, an EfficientNet neural network model can consider speed and accuracy.\u003c/p\u003e \u003cp\u003eA series of EfficientNet models were obtained by amplifying the basic model. This series of models out-performed all previous CNN models in efficiency and accuracy. A state-of-the-art deep neural network architecture, EfficientNet-B0, was used for this procedure. The model had 237 layers and 5,330,564 parameters, and 5,288,548 parameters were needed for gradient descent in training. The core structure of the network was the Mobile Inverted Bottleneck Convolution (MBConvBlock) module, which also uses the Squeeze-and-Excitation Network (SENet). When SENet was proposed, it achieved the highest accuracy on Imagenet data at that time. The deep CNN uses back propagation to train the model. All layers of the network were fine-tuned using AdamW (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://arxiV.org/pdf/1711.05101.pdf\u003c/span\u003e\u003cspan address=\"https://arxiV.org/pdf/1711.05101.pdf\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e), a method for stochastic optimization with a global learning rate of 0.01 and decay to one-tenth every 30 epochs. To optimize images for EfficientNet-B0, they were resized to 512\u0026times;512 pixels.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e2.7 Statistical analysis\u003c/h2\u003e \u003cp\u003e \u003cem\u003eEstimation of sample size\u003c/em\u003e \u003c/p\u003e \u003cp\u003eThe purpose was to verify the noninferiority of AI-assisted endoscopists compared with AI-unassisted endoscopists in the diagnosis of \u003cem\u003eH. pylori\u003c/em\u003e infection. In the test data, AI diagnostic accuracy was 89.6%, and the accuracy of the endoscopists was 82.4%\u003csup\u003e[14]\u003c/sup\u003e. Based on a noninferiority margin of 10%, an α of 0.05, and a 1\u0026thinsp;\u0026minus;\u0026thinsp;β of 0.90, the estimated sample size was 84 per group. Noninferiority would be inferred if the 95% lower confidence boundary for the difference between the two groups in accuracy was more than 10%. The sample size was expanded using PASS version 11,with a total of 418 patients in the final two groups.\u003c/p\u003e \u003cp\u003e \u003cem\u003eAnalysis of outcomes\u003c/em\u003e \u003c/p\u003e \u003cp\u003eNumerical data were expressed as means\u0026thinsp;\u0026plusmn;\u0026thinsp;SDs. The accuracy, specificity, sensitivity, PPV, and NPV from the different analyses were compared using the chi-square test or Fisher's exact test, as appropriate. All statistical calculations were performed using IBM SPSS version 23. A p-value below 0.05 was considered statistically significant. The trained CNN generated a continuous number for the probability of \u003cem\u003eH. pylori\u003c/em\u003e infection (range: 0 to 1.0), and this was used as a continuous variable in the ROC analysis. The AUC values of the \u003cem\u003eH. pylori\u003c/em\u003e-positive and \u003cem\u003eH. pylori-\u003c/em\u003enegative groups were compared and plotted using R language version 4.01.\u003c/p\u003e \u003c/div\u003e"},{"header":"3. Results","content":"\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Development data\u003c/h2\u003e \u003cp\u003eFrom September 2020 to December 2020, 639 patients (260 \u003cem\u003eH. pylori-\u003c/em\u003epositive, 379 \u003cem\u003eH. pylori-\u003c/em\u003enegative; 289 males, 350 females) received EGDs in the endoscopy center of the Department of Gastroenterology, Daping Hospital of Chongqing (Table \u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). For every patient, there was at least one image from each of five sites in the stomach (upper body, middle body, lower body, lesser curvature, and antrum). We used 7377 images from 639 patients (development data set) to construct the CNN.\u003c/p\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.2 Performance of the CNN\u003c/h2\u003e \u003cp\u003eFrom January 2021 to March 2021, 201 different patients (66 \u003cem\u003eH. pylori\u003c/em\u003e-positive, 135 \u003cem\u003eH. pylori-\u003c/em\u003enegative; 90 males, 111 females) received EGDs at the same institution and these images were used for the test data set (Table \u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). For the diagnosis of \u003cem\u003eH. pylori\u003c/em\u003e infection, the CNN model had an accuracy of 89.6%, a sensitivity of 90.9%, and a specificity of 88.9%. The AUC for \u003cem\u003eH. pylori\u003c/em\u003e positivity was 84.1% (95%CI: 73.0%, 95.2%) and the AUC for \u003cem\u003eH. pylori\u003c/em\u003e negativity was 90.3% (95%CI: 82.2%, 98.4%; FIGURE \u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eWe also performed separate analyses of diagnosis based on the gastric body alone (upper body, middle body, lower body, and lesser curvature) and the antrum alone (Table \u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). Analysis of the gastric body provided significantly greater sensitivity (86.2% \u003cem\u003evs.\u003c/em\u003e 19.8%) and accuracy (76.4% \u003cem\u003evs.\u003c/em\u003e 68.4%), but analysis of the antrum provided significantly greater specificity (93.3% \u003cem\u003evs.\u003c/em\u003e 71.7%); analysis of the antrum provided a significantly lower false-positive rate (6.7% \u003cem\u003evs.\u003c/em\u003e 28.3%), but analysis of the body provided a significantly lower false-negative rate (13.8% \u003cem\u003evs.\u003c/em\u003e 80.2%). Combining data from the body and antrum provided the best diagnostic performance.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eSensitivity, specificity, accuracy, false-positive rate, and false-negative rate for diagnosis of \u003cem\u003eH. pylori\u003c/em\u003e infection based on examination of the body and the antrum.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDiagnostic parameter\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eBody*\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAntrum\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cem\u003eP\u003c/em\u003e value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSensitivity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e86.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e19.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e71.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e93.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e76.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e68.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFalse positive rate\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e28.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e6.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFalse negative rate\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e13.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e80.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e*Stomach body consists of the upper body, middle body, lower body, and lesser curvature.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e3.3 Multicenter randomized controlled study\u003c/h2\u003e \u003cp\u003eFrom April 2021 to July 2021, we enrolled 418 patients from three institutions (Department of Gastroenterology of Chongqing Daping Hospital, The 13th People's Hospital of Chongqing, and The Second People's Hospital of Jiulongpo District of Chongqing) for a multicenter randomized controlled trial of endoscopists. There were 209 patients in the AI-assisted group and 209 other patients in the AI-unassisted group (Table \u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e), and these two groups had no significant differences at baseline.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eBaseline demographic and clinical characteristics of patients examined in the multicenter randomized controlled study of endoscopists.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCharacteristic\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAI-assisted\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAI-unassisted\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cem\u003ep\u003c/em\u003e-value\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSex\u003c/p\u003e \u003cp\u003eMale, n (%)\u003c/p\u003e \u003cp\u003eFemale, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e92/209(44.0)\u003c/p\u003e \u003cp\u003e117/209(56.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e81/209(38.7)\u003c/p\u003e \u003cp\u003e128/209(61.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.275\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMean age, years (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e46.3\u0026thinsp;\u0026plusmn;\u0026thinsp;12.1\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e45.7\u0026thinsp;\u0026plusmn;\u0026thinsp;12.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.652\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eMean BMI (SD)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e22.7\u0026thinsp;\u0026plusmn;\u0026thinsp;3.4\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e23.2\u0026thinsp;\u0026plusmn;\u0026thinsp;2.9\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.134\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cem\u003eH. pylori\u003c/em\u003e status, n (%)\u003c/p\u003e \u003cp\u003ePositive\u003c/p\u003e \u003cp\u003eNegative\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e73/209(34.9)\u003c/p\u003e \u003cp\u003e136/209(65.0)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e56/209(26.7)\u003c/p\u003e \u003cp\u003e153/209(73.2)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.072\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCigarette smoking, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e33/209(15.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e28/209(13.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.488\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAlcohol drinking, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e53/209(25.3)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e47/209(22.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.492\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eFamily history of esophageal or gastric carcinoma, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e12/209(5.7)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e18/209(8.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.256\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBiopsy cases, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e26/209(12.4)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e31/209(14.8)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.476\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAtrophy cases, n (%)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e61/209(29.1)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e64/209(30.6)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.749\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eComparisons of all 14 endoscopists in the AI-assisted and AI-unassisted groups indicated the AI-assisted group had significantly better accuracy (92.8% \u003cem\u003evs.\u003c/em\u003e 75.6%, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001), sensitivity (91.8% \u003cem\u003evs.\u003c/em\u003e 78.6%, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.032), and specificity (93.4% \u003cem\u003evs.\u003c/em\u003e 74.5%, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001; Table \u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eSensitivity, specificity, accuracy, positive predictive value (PPV), and negative predictive value (NPV) of endoscopists with and without AI assistance in the diagnosis of \u003cem\u003eH. pylori\u003c/em\u003e infection.\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c3\" namest=\"c2\"\u003e \u003cp\u003eAI-assisted\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e \u003cp\u003eAI-unassisted\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eTotal\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eDiagnostic parameter\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eUntrained*\u003c/p\u003e \u003cp\u003e(n\u0026thinsp;=\u0026thinsp;11)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eTrained**\u003c/p\u003e \u003cp\u003e(n\u0026thinsp;=\u0026thinsp;3)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e14\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eUntrained*\u003c/p\u003e \u003cp\u003e(n\u0026thinsp;=\u0026thinsp;11)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eTrained**\u003c/p\u003e \u003cp\u003e(n\u0026thinsp;=\u0026thinsp;3)\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003e14\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003e\u003cem\u003eP\u003c/em\u003e value\u003csup\u003e***\u003c/sup\u003e\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSensitivity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e91.6%(44/48)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e92.0%(23/25)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e91.8%(67/73)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e76.3%(29/38)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e83.3%(15/18)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e78.6%(44/56)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e\u003cb\u003e0.032\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSpecificity\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e91.9%(103/112)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e100%(24/24)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e93.4%(127/136)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e70.6%(82/116)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e86.4%(32/37)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e74.5%(114/153)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e\u003cb\u003e\u0026lt;\u0026thinsp;0.001\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e91.9%(147/160)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e95.9%(47/49)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e92.8%(194/209)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e72.1%(111/154)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e85.5%(47/55)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e75.6%(158/209)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e\u003cb\u003e\u0026lt;\u0026thinsp;0.001\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003ePPV\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e83.0%(44/53)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e100%(23/23)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e88.2%(67/76)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e46.0%(29/63)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e75.0%(15/20)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e53.0%(44/83)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e\u003cb\u003e\u0026lt;\u0026thinsp;0.001\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNPV\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e96.2%(103/107)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e92.3%(24/26)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e95.5%(127/133)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e90.1%(82/91)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e91.4%(32/35)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e90.5%(114/126)\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e\u003cb\u003e0.113\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003ctfoot\u003e \u003ctr\u003e\u003ctd colspan=\"8\"\u003e\u003csup\u003e*\u003c/sup\u003eNot trained for early cancer screening\u003c/td\u003e\u003c/tr\u003e \u003ctr\u003e\u003ctd colspan=\"8\"\u003e\u003csup\u003e**\u003c/sup\u003eTrained for early cancer screening\u003c/td\u003e\u003c/tr\u003e \u003ctr\u003e\u003ctd colspan=\"8\"\u003e\u003csup\u003e***\u003c/sup\u003eComparison of all AI-assisted \u003cem\u003evs.\u003c/em\u003e all AI-unassisted\u003c/td\u003e\u003c/tr\u003e \u003c/tfoot\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eWe then determined the accuracy of endoscopists in the AI-unassisted group who had different levels of experience, using two different metrics to define \u0026ldquo;junior\u0026rdquo; and \u0026ldquo;senior\u0026rdquo; status (FIGURE \u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). The accuracy of diagnosis was 85.5% (47/55) for senior endoscopists who were trained in early cancer detection, and was 72.1% (111/154) for junior endoscopists who did not have this training (\u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.047). The accuracy of diagnosis was 77.3% (75/97) for senior endoscopists who had 5 years of experience or examination of more than 5000 cases, and was 73.2% (82/112) for junior endoscopists who did not have this experience (\u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.494).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"4. Discussion","content":"\u003cp\u003eThe development of endoscopic technology has made it possible to identify \u003cem\u003eH. pylori\u003c/em\u003e using EGD. However, some current guidelines do not recommend endoscopy for the routine diagnosis of \u003cem\u003eH. pylori\u003c/em\u003e infection because these methods require special equipment, endoscopists must receive relevant training, and the accuracy and specificity may differ among endoscopists\u003csup\u003e[23]\u003c/sup\u003e. Notably, developments in AI may resolve some of the problems related to the endoscopic diagnosis of \u003cem\u003eH. pylori\u003c/em\u003e infection. AI methods can learn rules from sample data, and can recognize different types of data, such as text, images, and sounds. The image recognition used in the present study is based on deep learning and is a key technology for AI-assisted endoscopic diagnosis of \u003cem\u003eH.pylori\u003c/em\u003e infection\u003csup\u003e[12, 13]\u003c/sup\u003e. Recent studies showed that AI endoscopic diagnosis of \u003cem\u003eH.pylori\u003c/em\u003e infection is possible. In particular, a meta-analysis of AI-based diagnosis of \u003cem\u003eH.pylori\u003c/em\u003e infection from endoscopy showed that the sensitivity was 0.87 (95%CI: 0.72, 0.94), the specificity was 0.86 (95%CI: 0.77, 0.92), and the AUC was 0.92 (95%CI: 0.90, 0.94)\u003csup\u003e[15]\u003c/sup\u003e. However, most studies of this topic were retrospective and diagnostic, and there have been no clinical prospective studies\u003csup\u003e[15,19\u0026ndash;22]\u003c/sup\u003e. In the present study, we combined a diagnostic study and multicenter randomized controlled study of endoscopists. Our results confirmed the clinical value of AI-assisted endoscopic diagnosis of \u003cem\u003eH. pylori\u003c/em\u003e infection.\u003c/p\u003e \u003cp\u003eIn the present study, the endoscopic diagnostic model of \u003cem\u003eH. pylori\u003c/em\u003e infection was based on deep learning and was constructed using endoscopic images of five stomach regions. In the test data, AI diagnosis that was based on images of the gastric body (upper body, middle body, lower body, and lesser curvature) provided better accuracy, sensitivity, and PPV than AI diagnosis based on images of the antrum; however, images of the antrum provided better specificity and NPV than images of the body. The combined use of body and antrum images had an accuracy of 89.6%, a sensitivity of 90.9%, and a specificity of 88.9%. A 2020 Japanese study found that diagnostic accuracy was greater when using the lesser curvature of the middle-upper body than the fornix and the greater curvature of the middle-upper body\u003csup\u003e[17,18]\u003c/sup\u003e. A 2019 study in China found that the sensitivity, specificity, and accuracy from using multiple gastric images (average 8.3\u0026thinsp;\u0026plusmn;\u0026thinsp;3.3 per patient) for AI diagnosis of \u003cem\u003eH. pylori\u003c/em\u003e infection were greater than those from using a single image\u003csup\u003e[18]\u003c/sup\u003e. This led us to use an AI-assisted endoscopy diagnostic system for \u003cem\u003eH. pylori\u003c/em\u003e infection that is based on images of multiple sites for model establishment.\u003c/p\u003e \u003cp\u003eMost previous endoscopy studies have used WLI, and only a few have used newer endoscopic techniques. A 2018 Japanese study compared the effectiveness of using AI with WLI, Blue Laser Imaging (BLI), and Linked Color Imaging (LCI) for the diagnosis of \u003cem\u003eH. pylori\u003c/em\u003e infection, and found that the AUC was higher for BLI (0.96) and LCI (0.95) than for WLI (0.66, both \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.01)\u003csup\u003e[16]\u003c/sup\u003e. Although these new techniques seem to have advantages over WLI in the diagnosis of \u003cem\u003eH. pylori\u003c/em\u003e infection, they have not been widely adopted because they require additional expensive equipment and professionally trained endoscopists. Therefore, an \u003cem\u003eH. pylori\u003c/em\u003e diagnostic system developed for WLI is more suitable for widespread acceptance, especially in regions where resources are limited. In addition, some of the limitations of WLI can be overcome by use of AI learning from images at multiple sites in the stomach.\u003c/p\u003e \u003cp\u003eThe present study was the first to examine an AI system for the endoscopic diagnosis of \u003cem\u003eH. pylori\u003c/em\u003e infection in a multicenter randomized controlled clinical trial of endoscopists. We found that endoscopists with AI assistance had significantly better accuracy, sensitivity, and specificity than unassisted endoscopists (all \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05). This indicates that our AI endoscopic system for the diagnosis of \u003cem\u003eH. pylori\u003c/em\u003e infection significantly improved the diagnostic ability of endoscopists when using gastroscopy.\u003c/p\u003e \u003cp\u003eOur analysis of endoscopists also found the number of working years and the number of procedures that were performed had no significant effect on their diagnostic performance. However, endoscopists who were trained in early cancer detection had greater accuracy than endoscopists without this training (85.5% [47/55] \u003cem\u003evs.\u003c/em\u003e 72.1% [111/154], P\u0026thinsp;=\u0026thinsp;0.047). This may be because in the professional training for early cancer detection, the recognition of \u003cem\u003eH. pylori\u003c/em\u003e status of gastric mucosa using endoscopy is very important for determining the mucosal background of gastric cancer, so these specially trained endoscopists are familiar with the manifestations of \u003cem\u003eH. pylori\u003c/em\u003e infection and pay close attention to these observations. Endoscopists not trained in early cancer detection may have lower accuracy, even though they are skilled in endoscopy procedures. This further emphasizes the clinical value of the using an AI system for the endoscopic diagnosis of \u003cem\u003eH. pylori\u003c/em\u003e infection.\u003c/p\u003e \u003cp\u003eThe UBT is the most common method for the clinical detection of \u003cem\u003eH. pylori\u003c/em\u003e infection. However, when the results of this test are close to the test threshold, diagnosis can be difficult. The present study showed that an AI endoscopy system can assist doctors whose UBT results are near the threshold. In particular, among the 418 patients, 6 patients had UBT results near the critical value, and 5 of them were accurately assessed by the AI system, corresponding to an accuracy of 83.3% and a sensitivity of 100% (data not shown). However, this result is from only 6 patients, so a larger sample size is needed for further study of this topic.\u003c/p\u003e \u003cp\u003eThe status of stomach infection by \u003cem\u003eH. pylori\u003c/em\u003e may be classified as \u0026ldquo;infected\u0026rdquo;, \u0026ldquo;uninfected\u0026rdquo;, or \u0026ldquo;eradicated\u0026rdquo;, with the latter two states considered negative. We found it was difficult to identify the eradicated state using endoscopy, and identification of this state was also affected by the time since eradication. In particular, among the 418 patients, 369 were not treated for \u003cem\u003eH. pylori\u003c/em\u003e infection, and the accuracy of diagnosis in these patients was 85.9%, greater than the accuracy in patients treated for \u003cem\u003eH. pylori\u003c/em\u003e infection (71.4%, data not shown). This result indicates that prior treatment for \u003cem\u003eH. pylori\u003c/em\u003e infection adversely affected diagnosis by doctors alone and by doctors using AI. This may be because of decreased gastric mucosal inflammation and atypical gastric mucosal appearance after the use of antibiotics and PPIs. We also found better accuracy (81.0% \u003cem\u003evs.\u003c/em\u003e 65.4%), specificity (83.3% \u003cem\u003evs.\u003c/em\u003e 68.1%), and sensitivity (66.7% \u003cem\u003evs.\u003c/em\u003e 50.0%) in patients who received treatment more than two years ago rather than less than 2 years ago (data not shown). We speculate that as the time after treatment increases, there were greater declines in inflammation of the gastric mucosa,so endoscopic gastric mucosa becomes more typical\u003csup\u003e[24\u0026ndash;27]\u003c/sup\u003e. For example, the diffuse redness from the gastric body to the gastric fundus disappears, and some of these individuals even show RAC again, so the accuracy of endoscopic diagnosis improved accordingly.\u003c/p\u003e \u003cp\u003eThere are some limitations in this study. First, our development and test data sets were all from a single center. Second, our sample size was rather small, and a larger sample should be used in subsequent studies. Finally, this study did not use the classification system of \u0026ldquo;positive\u0026rdquo;, \u0026ldquo;negative\u0026rdquo;, and \u0026ldquo;negative after eradication\u0026rdquo;, a topic that is also worthy of further study.\u003c/p\u003e \u003cp\u003eIn conclusion, we established an AI-assisted endoscopy system for the diagnosis of \u003cem\u003eH. pylori\u003c/em\u003e infection that was based on AI learning of images from five sites in the stomach, and applied this system to the first clinical trial of this topic in China. The results of this multicenter randomized controlled trial verified that the AI system described here improved the ability of endoscopists to diagnose \u003cem\u003eH. pylori\u003c/em\u003e infection using gastroscopy. Because this system provided greater improvements to endoscopists who were not trained in early cancer detection, we believe it may be beneficial for geographic regions that have limited resources, high incidences of gastric cancer, and common use and acceptance of gastroscopy.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eCONFLICT OF INTEREST\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThere are no conflicts of interest to declare.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAUTHOR CONTRIBUTIONS\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll authors contributed to the study conception and design. The first three authors contributed equally to this work. Material preparation, data collection, and data analysis were performed by Pei-Ying Zou, Jian-Ru Zhu, Zhe Zhao, Hao Mei, Jing-Tao Zhao, Wen-Jing Sun, Li-Lin Fan, and Guo-Hua Wang. The first draft of the manuscript was written by Pei-Ying Zou. Chun-Hui Lan critically reviewed the manuscript. All the authors commented on previous versions of the manuscript and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCONSENT FOR PULICATION\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eACKNOWLEDGEMENTS\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was supported by grants from the Science and Technology Innovation Enhancement Project of Army Medical University (No. 20I9CXLCB003) and the National Science Foundation of China (No. 82072253).\u003c/p\u003e\n\u003cp\u003eThis study was approved by the Ethics Committee of Daping Hospital (10/07/2020)（No.89,2020）and was registered with the Chinese Clinical Trial Registration Center (02/09/2020) (www.chictr.org.cn; registration number: ChiCTR2000037801).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eVAILABILITY OF DATA \u0026nbsp;AND MATERIALS\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll data generated or analysed during this study are included in this published article.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n \u003cli\u003e\u003cspan\u003eHooi, J.K.Y., et al., Global Prevalence of Helicobacter pylori Infection: Systematic Review and Meta-Analysis. Gastroenterology, 2017. 153(2): p. 420\u0026ndash;429.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eSavoldi, A., et al., Prevalence of Antibiotic Resistance in Helicobacter pylori: A Systematic Review and Meta-analysis in World Health Organization Regions. Gastroenterology, 2018. 155(5): p. 1372\u0026ndash;1382.e17.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eNdip, R.N., et al., Helicobacter pylori isolates recovered from gastric biopsies of patients with gastro-duodenal pathologies in Cameroon: current status of antibiogram. Tropical Medicine \u0026amp; International Health, 2008. 13(6): p. 848\u0026ndash;854.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eHuh, C.W. and B.W. Kim, [Diagnosis of Helicobacter pylori Infection]. Korean J Gastroenterol, 2018. 72(5): p. 229\u0026ndash;236.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eWang, Y.K., et al., Diagnosis of Helicobacter pylori infection: Current options and developments. World J Gastroenterol, 2015. 21(40): p. 11221-35.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eMakristathis, A., et al., Review: Diagnosis of Helicobacter pylori infection. Helicobacter, 2019. 24 Suppl 1: p. e12641.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eChen Huang wei., et al., The application value of Kyoto gastritis classification in the direct judgment of Helicobacter pylori infection under white light gastroscopy. New Medicine, 2019. 50(6): 457\u0026ndash;462\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eKato Motoji et al., Classification of gastritis by Kyoto. Stomach and Intestine, 2019. 54(5):616\u0026ndash;619.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eKato T,Yagi N,Kamada T,Shimbo T,Watanabe H,Ida K;Diagnosis of Helicobacter pylori infection in gastric mucosa by endoscopic features:a multicenter prospective Endosc,2013,25(5):508\u0026ndash;518.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eCho JH,Chang YW,Jang JY,Shim.JJ,Lee CK,Dong SH,Kim HJ,Kim BH,Lee TH,Cho JY.Close observation of gastric mucosal pattern by standard endoscopy can predict Helicobacter pylori infection status.J Gastroenterol Hepatol,2013,28(2 4.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eWatanabe K,Nagata N,Nakashima R,Furuhata E,Shimbo T,Kobayakawa M,Sakurai T,Imbe K,Niikura R,Yokoi C J,Uemura N.Predictive findings for Helicobacter pylori-uninfected,-infected and-eradicated gastric mucosa:validation orld J Gastroenterol,2013,19(27):4374\u0026ndash;4379.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eEsteva, A., et al., Dermatologist-level classification of skin cancer with deep neural networks. Nature, 2017. 542(7639): p. 115\u0026ndash;118.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eRibeiro E, Uhl A, Wimmer G, H\u0026auml;fner M. Exploring deep learning and transfer learning for colonic polyp classification. Comput Math Methods Med 2016;2016;6584725.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eShichijo, S., et al., Application of Convolutional Neural Networks in the Diagnosis of Helicobacter pylori Infection Based on Endoscopic Images. EBioMedicine, 2017. 25: p. 106\u0026ndash;111.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eBang, C.S., J.J. Lee and G.H. Baik, Artificial Intelligence for the Prediction of Helicobacter Pylori Infection in Endoscopic Images: Systematic Review and Meta-Analysis Of Diagnostic Test Accuracy. J Med Internet Res, 2020. 22(9): p. e21983.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eNakashima, H., et al., Artificial intelligence diagnosis of Helicobacter pylori infection using blue laser imaging-bright and linked color imaging: a single-center prospective study. Ann Gastroenterol, 2018. 31(4): p. 462\u0026ndash;468.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eYasuda, T., et al., Potential of automatic diagnosis system with linked color imaging for diagnosis of Helicobacter pylori infection. Dig Endosc, 2020. 32(3): p. 373\u0026ndash;381.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eZheng, W., et al., High Accuracy of Convolutional Neural Network for Evaluation of Helicobacter pylori Infection Based on Endoscopic Images: Preliminary Experience. Clin Transl Gastroenterol, 2019. 10(12): p. e00109.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eShichijo S, Endo Y, Aoyama K, Takeuchi Y, Ozawa T, Takiyama H, et al. Application of convolutional neural networks for evaluating Helicobacter pylori infection status on the basis of endoscopic images. Scand J Gastroenterol 2019 Feb;54(2):158\u0026ndash;163.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eItoh T, Kawahira H, Nakashima H, Yata N. Deep learning analyzes Helicobacter pylori infection by upper gastrointestinal endoscopy images. Endosc Int Open 2018 Feb;6(2):E139-E144 [FREE Full text] [doi: 10.1055/s-0043-120830] [Medline: 29399610].\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eHuang C, Chung P, Sheu B, Kuo H, Popper M. Helicobacter pylori-related gastric histology classification using support-vector-machine-based feature selection. IEEE Trans Inf Technol Biomed 2008 Jul;12(4):523\u0026ndash;531.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eHuang C, Sheu B, Chung P, Yang H. Computerized diagnosis of Helicobacter pylori infection and associated gastric inflammation from endoscopic images by refined feature selection using a neural network. Endoscopy 2004 Jul;36(7):601\u0026ndash;608.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eLiu Wenzhong et al., The fifth National Consensus report on the management of Helicobacter pylori infection. Gastroenterology, 2017. 22(06): 346\u0026ndash;360.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eYumin Zhang, Dadao Jing and Yulan Qiu, Effects of Helicobacter pylori eradication on gastric mucosal precancerous lesions: a meta-analysis. Journal of Clinical Gastroenterology, 2009. 21(05): 268\u0026ndash;272.\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eSung JJ\u0026sbquo; Lin SR\u0026sbquo;Ching JY\u0026sbquo;et al༎ Atrophy and intestinal metaplasia one year after cure of H.pylori infection: a prospective\u0026sbquo; randomized study༎Gastroenterology\u0026sbquo;2000\u0026sbquo;119༚7\u0026ndash;14༎ 4 Correa P\u0026sbquo;Fontham ET\u0026sbquo;Bravo\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eJC\u0026sbquo;et al༎Chemoprevention of gastric dysplasia: randomized trial of antioxidants supplements and ant-i Helicobacter pylori therapy༎J Natl Cancer Inst\u0026sbquo;2000\u0026sbquo;92༚ 1881\u0026ndash;1888༎\u003c/span\u003e\u003c/li\u003e\n \u003cli\u003e\u003cspan\u003eKhan, M.Y., et al., Effectiveness of Helicobacter pylori eradication in preventing metachronous gastric cancer and preneoplastic lesions. A systematic review and meta-analysis. Eur J Gastroenterol Hepatol, 2020. 32(6): p. 686\u0026ndash;694.\u003c/span\u003e\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003eTable 1 is available in the Supplementary Files section.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-gastroenterology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmge","sideBox":"Learn more about [BMC Gastroenterology](http://bmcgastroenterol.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmge/default.aspx","title":"BMC Gastroenterology","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Helicobacter pylori, Endoscopy, Artificial Intelligence, Convolutional Neural Network","lastPublishedDoi":"10.21203/rs.3.rs-3747640/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-3747640/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eThe early diagnosis and treatment of \u003cem\u003eHeliobacter pylori\u003c/em\u003e gastrointestinal infection provide significant benefits to patients. We constructed a convolutional neural network (CNN) model based on an endoscopic system to diagnose \u003cem\u003eH. pylori\u003c/em\u003e infection, and then examined the potential benefit of this model to endoscopists in their diagnosis of \u003cem\u003eH. pylori\u003c/em\u003e infection.\u003c/p\u003e\u003ch2\u003eMaterials and Methods\u003c/h2\u003e \u003cp\u003eA CNN neural network system for endoscopic diagnosis of \u003cem\u003eH. pylori\u003c/em\u003e infection was established by collecting 7377 endoscopic images from 639 patients. The accuracy, sensitivity, and specificity were determined. Then, a randomized controlled study was used to compare the accuracy of diagnosis of \u003cem\u003eH. pylori\u003c/em\u003e infection by endoscopists who were assisted or unassisted by this CNN model.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eThe deep CNN model for diagnosis of \u003cem\u003eH. pylori\u003c/em\u003e infection had an accuracy of 89.6%, a sensitivity of 90.9%, and a specificity of 88.9%. Relative to the group of endoscopists unassisted by AI, the AI-assisted group had better accuracy (92.8% [194/209; 95%CI: 89.3%, 96.4%] \u003cem\u003evs.\u003c/em\u003e 75.6% [158/209; 95%CI: 69.7%, 81.5%]), sensitivity (91.8% [67/73; 95%CI: 85.3%, 98.2%] \u003cem\u003evs.\u003c/em\u003e 78.6% [44/56; 95%CI: 67.5%, 89.7%]), and specificity (93.4% [127/136; 95%CI: 89.2%, 97.6%] \u003cem\u003evs.\u003c/em\u003e 74.5% [114/153; 95%CI: 67.5%, 81.5%]). All of these differences were statistically significant (\u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05).\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e \u003cp\u003eOur AI-assisted system for diagnosis of \u003cem\u003eH. pylori\u003c/em\u003e infection has good diagnostic ability, and can improve the accuracy of endoscopists in gastroscopic diagnosis.\u003c/p\u003e","manuscriptTitle":"Development and application of an artificial intelligence-assisted endoscopy system for diagnosis of Helicobacter pylori infection: a multicenter randomized controlled study","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2024-01-19 19:48:43","doi":"10.21203/rs.3.rs-3747640/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"editorInvited","content":"","date":"2024-01-16T09:31:43+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2024-01-16T09:29:10+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Gastroenterology","date":"2023-12-13T09:26:05+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-gastroenterology","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bmge","sideBox":"Learn more about [BMC Gastroenterology](http://bmcgastroenterol.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bmge/default.aspx","title":"BMC Gastroenterology","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"8aa426e8-b2f0-4554-8f6b-e61dd099d4c9","owner":[],"postedDate":"January 19th, 2024","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2024-10-07T16:04:34+00:00","versionOfRecord":{"articleIdentity":"rs-3747640","link":"https://doi.org/10.1186/s12876-024-03389-3","journal":{"identity":"bmc-gastroenterology","isVorOnly":false,"title":"BMC Gastroenterology"},"publishedOn":"2024-09-30 15:58:08","publishedOnDateReadable":"September 30th, 2024"},"versionCreatedAt":"2024-01-19 19:48:43","video":"","vorDoi":"10.1186/s12876-024-03389-3","vorDoiUrl":"https://doi.org/10.1186/s12876-024-03389-3","workflowStages":[]},"version":"v1","identity":"rs-3747640","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-3747640","identity":"rs-3747640","version":["v1"]},"buildId":"qtupq5eGEP_6zYnWcrvyt","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}