Baishideng Publishing Group Inc, 7041 Koll Center Parkway, Suite 160, Pleasanton, CA 94566, USA
Share the Article
Zhou NY, Luo Q, Liu W, Yuan XL, Jiang H, Hu B. Artificial intelligence in chronic atrophic gastritis: A comprehensive review. World J Gastrointest Endosc 2026; 18(11): 124991 [DOI: 10.4253/wjge.124991]
Nuo-Ya Zhou, Qi Luo, Wei Liu, Xiang-Lei Yuan, Huan Jiang, Bing Hu, Department of Gastroenterology and Hepatology/Medical Engineering Integration Laboratory of Digestive Endoscopy, West China Hospital, Sichuan University, Chengdu 610041, Sichuan Province, China
Author contributions: Zhou NY and Luo Q conducted a comprehensive literature review and composed the manuscript, both authors made crucial and indispensable contributions towards the completion of the project and thus qualified as the co-first authors of the paper; Liu W contributed invaluable critical insights and carried out substantial revisions; Yuan XL and Jiang H contributed significantly to the literature review process; Hu B meticulously revised the manuscript; the final version of the manuscript underwent thorough review and approval by all authors.
AI contribution statement: During the preparation of this manuscript, we used AI-assisted tools solely for language polishing and grammatical refinement to improve the readability and fluency of the text, and to assist in checking consistency among the main text, Table 1, and the reference list. Specifically, we used ChatGPT and DeepSeek for these purposes. All AI-generated suggestions were critically reviewed, revised as necessary, and verified by the authors against the original sources to ensure scientific accuracy, logical coherence, and conformity to the journal’s stylistic requirements. AI tools were not used for data analysis, interpretation, result generation, or any core scientific decision-making process. The authors assume full responsibility for the final content of this manuscript and its scientific integrity.
Supported by Natural Science Foundation of Sichuan Province, No. 2026YFHZ0074.
Conflict-of-interest statement: All authors have no conflict of interest to disclose.
Corresponding author: Bing Hu, Department of Gastroenterology and Hepatology/Medical Engineering Integration Laboratory of Digestive Endoscopy, West China Hospital, Sichuan University, No. 37 Guo Xue Alley, Wuhou District, Chengdu 610041, Sichuan Province, China. hubing@wchscu.edu.cn
Received: June 29, 2026 Revised: September 1, 2026 Accepted: October 20, 2026 Published online: November 16, 2026 Processing time: 138 Days and 23.6 Hours
Abstract
Chronic atrophic gastritis (CAG) is a significant precancerous condition associated with gastric cancer, and its accurate diagnosis and comprehensive management are of great importance for reducing the risk of gastric cancer. In recent years, artificial intelligence (AI) has advanced rapidly in the field of digestive diseases, offering new approaches for the intelligent diagnosis and treatment of CAG. Early studies suggest that AI may improve diagnostic performance and assist in risk stratification. However, whether these systems reduce clinician workload or improve patient outcomes remains to be established. Most evidence comes from retrospective diagnostic accuracy studies, and no AI system has yet been incorporated into clinical guidelines. This review summarizes the research progress of AI in CAG, focusing on five key aspects: Assisted diagnosis, pathological interpretation, Helicobacter pylori infection diagnosis, risk prediction, and patient management. It aims to provide a reference for research on CAG and to highlight priorities for future research and clinical translation.
Core Tip: Chronic atrophic gastritis (CAG) is a key precancerous condition associated with gastric cancer, necessitating accurate diagnosis and effective management. Artificial intelligence (AI) has emerged as a promising tool in digestive diseases, offering novel solutions for intelligent CAG care. Early evidence suggests that AI may improve diagnostic accuracy and assist in risk stratification. However, its impact on clinician workload and patient outcomes remains to be demonstrated, and no AI system described here is currently ready for guideline-based clinical use. This review synthesizes AI research in CAG across five domains: Assisted diagnosis, pathological interpretation, Helicobacter pylori infection diagnosis, risk prediction and patient management, with the aim of informing future research and clinical translation.
Citation: Zhou NY, Luo Q, Liu W, Yuan XL, Jiang H, Hu B. Artificial intelligence in chronic atrophic gastritis: A comprehensive review. World J Gastrointest Endosc 2026; 18(11): 124991
Chronic atrophic gastritis (CAG) is a chronic pathological condition characterized by the reduction or loss of native gastric glands, with or without intestinal metaplasia, typically induced by Helicobacter pylori (H. pylori) infection or autoimmune factors. As a key precancerous step in the development of intestinal-type gastric cancer, CAG lies at the early stage of the classic “Correa’s cascade” (normal mucosa → atrophic gastritis → intestinal metaplasia → dysplasia → gastric cancer)[1]. Therefore, its early recognition and precise management hold strategic significance for gastric cancer prevention.
In China, the prevalence of chronic gastritis diagnosed by endoscopy approaches 90%[2]. However, this remarkably high figure may partly reflect overdiagnosis. Although nationwide epidemiological data remain incomplete, one large-scale cross-sectional survey involving 8892 patients revealed that CAG and chronic non-atrophic gastritis accounted for 17.7% and 82.3% of cases, respectively[3]. Although the overall risk of progression to gastric cancer is low, the associated risk should not be overlooked. A large Swedish cohort study of 400000 patients reported that approximately one in 50 patients with CAG and one in 39 patients with gastric intestinal metaplasia progressed to gastric cancer within 20 years[4].
Current clinical management of CAG faces several core challenges: Endoscopic diagnosis is subjective, with conventional white-light imaging showing limited sensitivity and specificity for detecting gastric mucosal atrophy[5]; the gold standard, histopathological biopsy, is invasive, time-consuming, and prone to sampling bias; evidence on prognostic prediction in CAG patients is limited; patient management protocols are inconsistent; and patients often have insufficient understanding of the disease, resulting in unnecessary anxiety in those with mild disease or complete disregard in those with severe atrophy, leading to suboptimal H. pylori treatment and follow-up. In recent years, artificial intelligence (AI) based on deep learning has offered a new paradigm for objective and efficient diagnosis and treatment of CAG. AI has shown great potential in various aspects, including computer-aided diagnosis, image segmentation, pathological interpretation, risk prediction, and patient management.
This review aims to summarize the current landscape of AI applications in the diagnosis and management of CAG. We updated our literature search on June 20, 2026, using PubMed and Web of Science, covering records from database inception through June 20, 2026. The following search terms were used: “artificial intelligence”, “deep learning”, “convolutional neural network”, “computer-aided diagnosis”, “large language model”, “chronic atrophic gastritis”, “gastric atrophy”, “gastric intestinal metaplasia”, “Helicobacter pylori”, “H. pylori” and “gastric precancerous conditions”. The initial search identified 337 records. After removing duplicates, two authors independently screened titles and abstracts. Full texts of potentially eligible studies were reviewed. Disagreements were resolved by discussion with a third author. Studies were included if they evaluated AI applications in CAG or related gastric precancerous conditions using endoscopic images, pathological slides, or clinical data, and reported diagnostic performance, risk prediction, or patient management outcomes. We excluded studies focusing exclusively on gastric adenocarcinoma, non-English articles, conference abstracts without full data, and studies not directly relevant to CAG. As this is a narrative minireview, we prioritized original research articles, systematic reviews, and meta-analyses, with preference for multicenter, prospective, and externally validated studies. A total of 36 references were included.
AI-ASSISTED ENDOSCOPIC DIAGNOSIS OF CAG
Classification, detection, and segmentation of atrophy images
Endoscopic features of gastric mucosal atrophy include pale mucosal color, a predominantly white and red pattern (with white predominance), visibility of submucosal vessels, and flattening or disappearance of folds. Researchers have successively developed AI models utilizing deep learning techniques for the detection, classification, and segmentation of atrophy (Table 1).
Table 1 Artificial intelligence-assisted endoscopic detection of chronic atrophic gastritis.
For image classification and detection, Guimarães et al[6] developed a deep learning model (VGG16 architecture) for diagnosing atrophic gastritis using white-light endoscopic images. The model was trained on 200 images (100 atrophic, 100 non-atrophic) using 10-fold cross-validation and independently tested on a separate set of 70 images (30 atrophic, 40 non-atrophic). In the independent test set, it achieved an accuracy of 92.9%, sensitivity of 100%, specificity of 87.5%, and area under the curve (AUC) of 0.981, outperforming expert endoscopists in accuracy, balanced accuracy, and F-score. Zhang et al[7] collected 5470 gastric antrum images from 1699 patients, annotated based on pathological results, and developed a model for CAG diagnosis that achieved an accuracy, sensitivity, and specificity of 94.24%, 94.58%, and 94.01% respectively, outperforming three expert endoscopists. Lin et al[8] used TResNet to construct a multi-label classification model that achieved an AUC of 0.98 for identifying CAG. Mu et al[9] employed UNet++ for lesion segmentation and ResNet-50 to build a multi-level classification framework that achieved up to 95.00% and 92.86% accuracy in the video set for classifying CAG vs non-atrophic gastritis and atrophy vs intestinal metaplasia, respectively. Xu et al[10] included 6250 image-enhanced endoscopy (IEE) images and 98 IEE video clips. The model achieved 90.1% accuracy for atrophy diagnosis in the internal test set, 86.4% in the multi-center external test set, and 87.8% in the prospective video test set. Qu et al[11] developed an Image Recognition-based Automated Endoscopic Report Generation System using YOLO v3. The overall accuracy for identifying 10 common gastrointestinal diseases was 88.41% in the multicenter test set and 89.65% in the prospective test set. For CAG, the model achieved an accuracy of 98.04% (multicenter test set) and 98.10% (prospective test set). Luo et al[12] collected 10593 white-light endoscopic images from 4005 patients and built two ResNet50-based models. Model 1, which detected gastric antrum atrophy, achieved an accuracy of 90.2% (internal) and 89.0% (external), with sensitivities of 89.1% and 90.5%, and specificities of 91.5% and 89.0%, respectively. Model 2, which detected CAG across all gastric regions, achieved accuracies of 85.9% (internal), 85.4% (external, mixed sites), and 91.6% (external, non-antrum). The model’s ability to identify gastric antrum atrophy was comparable to that of expert endoscopists. These findings underscore the potential of AI as a valuable auxiliary diagnostic tool for endoscopists.
In the evolution towards multi-task and real-time systems, Zhao et al[13] developed a U-Net-based real-time video monitoring model for endoscopic diagnosis of CAG using 5290 endoscopic images from 1711 patients (4175 CAG and 1115 chronic nonatrophic gastritis), of which 70% (3703 images) were used for training with five-fold cross-validation and 30% (1587 images) for testing. In the static test set, the model achieved an accuracy of 92.63%, sensitivity of 92.73%, and specificity of 92.24%[13,14]. Subsequently, the research team conducted two prospective studies using this model. In a prospective paired-case cohort study involving 431 patients (of whom 268 patients were used for matched analysis), the model significantly increased the CAG detection rate from 24.6% to 35.8% (P = 0.005), increased the number of atrophic lesions detected (1.019 vs 0.522, P < 0.001), and reduced the number of biopsies (1.284 vs 3.243, P < 0.001), outperforming the detection results of endoscopists[14]. In a nested case-control study conducted within the same cohort, after propensity score matching (676 patients: 338 CAG vs 338 chronic nonatrophic gastritis), the model achieved an accuracy of 90.53%, a sensitivity of 84.02%, and a specificity of 97.04%, with an AUC of 0.909 (95%CI: 0.884-0.934), significantly outperforming endoscopists (accuracy of 72.34%, AUC of 0.740)[13]. Together, these two studies suggest that the same U-Net model may improve the clinical detection rate of CAG in real-time video monitoring and provide high diagnostic accuracy, with a high degree of agreement with pathological diagnosis results. Chen et al[15] developed and validated a novel endoscopic AI system, based on a multi-site feature fusion model. This system automatically classified five gastric regions (corpus, overturned corpus, fundus, angle, and antrum) and used ResNet-101 to enable differential diagnosis of autoimmune gastritis, H. pylori-associated atrophic gastritis (HpAG), and non-atrophic gastritis. The system maintained a sensitivity of 90.3% on external test sets and, notably, significantly enhanced the diagnostic ability of endoscopists (especially non-experts), increasing their sensitivity from 70.0% to 88.7%. These approaches highlight the potential clinical utility of AI-assisted endoscopic systems, particularly for improving diagnostic performance among less experienced endoscopists.
The integration of novel architectures like Vision Transformer (ViT) has also been explored. Yang et al[16] constructed a new dataset containing 21420 physician-annotated Linked Color Imaging and White Light Imaging images (including images from antrum, angle, cardia, fundus, and body), with 6651 images of atrophy. They developed a classification model for normal, atrophic, and metaplastic gastric mucosa, achieving 97.12% accuracy for atrophy classification, surpassing four physicians (whose accuracy ranged from 62% to 70%). Chen et al[17] developed and evaluated two ViT-based models to assist in the endoscopic diagnosis of CAG. A Swin Transformer was trained to classify CAG and chronic nonatrophic gastritis and to recognize five gastric anatomical structures, achieving an accuracy of 0.91 for CAG classification, with high concordance with senior endoscopists. In parallel, an SSFormer model was trained to segment atrophic lesions at the pixel level, achieving a mean Dice coefficient of 0.94 on the test set, indicating substantial overlap with expert-reviewed annotations. Chiang et al[18] developed a system for automatically diagnosing H. pylori infection and gastric precancerous lesions using routine upper gastrointestinal endoscopic images. This multi-step cascaded model performed a series of classification tasks: Excluding low-quality/non-gastric images, identifying gastric regions, delineating anatomical sites (cardia/fundus, body, antrum), and finally performing binary classification for active H. pylori infection, CAG, and intestinal metaplasia. The model achieved diagnostic accuracies of 92.8%, 88.6%, and 88.0% for H. pylori infection, CAG, and intestinal metaplasia, respectively.
Conventional endoscopic assessment of atrophy is operatordependent and may contribute to overdiagnosis, as suggested by the high reported prevalence rates. AIassisted diagnostic systems have been explored to improve the objectivity and consistency of endoscopic evaluation; however, the potential for overdiagnosis also warrants careful attention in AI models.
Grading of gastritis severity
Another significant application of AI in gastritis management is the automatic grading of endoscopic atrophy severity, aiming to combine visual features with standardized systems such as the Sydney Classification or the Kimura-Takemoto classification to achieve objective, quantitative risk stratification (Table 1).
Several studies have made notable progress in this direction. Zhao et al[19] developed a deep learning model based on the U-Net architecture, using 5290 high-quality endoscopic images from 1711 patients, where CAG images were sub-classified as mild, moderate, or severe according to the Sydney Classification. Prospective validation showed that the model’s diagnostic performance for CAG (sensitivity 89.31%, specificity 90.46%, AUC = 0.919) was significantly better than that of endoscopists, with high consistency with pathological operative link for gastritis assessment (OLGA) staging (Kappa = 0.816). It performed particularly well in identifying high-risk OLGA III/IV patients (sensitivity 93.97%, specificity 98.77%). These results indicate the model’s potential to effectively aid endoscopists in real-time diagnosis and precise risk stratification of CAG. Tao et al[20] developed a white-light endoscopy-based AI system for automatic recognition of gastric atrophy and risk stratification based on the Kimura-Takemoto classification. The system achieved stratification accuracies of 80.00%, 77.42%, 83.33%, and 85.71% for non-atrophy, mild, moderate, and severe atrophy, respectively, reaching the level of expert endoscopists and significantly outperforming junior physicians. Together, these studies demonstrate the potential of AI to effectively assist endoscopists in diagnosing and risk-stratifying gastric atrophy, particularly in improving diagnostic consistency among less experienced physicians, thereby providing a practical automated approach for gastric cancer risk monitoring.
Collectively, these findings illustrate the feasibility and promise of integrating AI with established classification systems such as the Sydney System and the KimuraTakemoto classification to enable objective, quantitative risk stratification in clinical practice.
AI-ASSISTED PATHOLOGY SLIDE ANALYSIS OF CAG
While endoscopic AI focuses on macroscopic mucosal patterns, histopathological assessment remains the diagnostic gold standard. The combination of AI and digital pathology is driving a shift towards objective, quantitative, and efficient histological assessment of gastric mucosal lesions.
Fang et al[21] developed and validated a semisupervised deep learning algorithm named GasMIL for diagnosing and grading gastric atrophy and intestinal metaplasia based on pathological images. Using 2725 whole slide images (WSI) from 545 patients, the model achieved AUCs of 0.884 and 0.877 for diagnosing intestinal metaplasia and atrophy, respectively, outperforming the ten pathologists involved in the study at the WSI level, and ranking second in OLGA and fifth in operative link for gastric intestinal metaplasia assessment (OLGIM) at the patient level. Its assistance significantly improved pathologists’ diagnostic performance for intestinal metaplasia (weighted kappa P = 0.016). Ma et al[22] proposed AMMNet, an attentionbased multiinstance multilabel learning network capable of simultaneous multilabel diagnosis of activity, atrophy, and intestinal metaplasia. Across 1096 patients from seven independent medical centers, AMMNet achieved AUCs of 0.93 for activity, 0.97 for atrophy, and 0.93 for intestinal metaplasia, while reducing WSI reading time from 5.46 minutes to 2.85 minutes per slide. Xia et al[23] developed GastritisMIL, an interpretable deep learning model for finegrading of inflammation, activity, atrophy, and intestinal metaplasia according to the updated Sydney System using retrospective data from 2744 patients. The model achieved AUCs of 0.971 for inflammation, 0.984 for activity, 0.974 for antrum atrophy, and 0.975 for intestinal metaplasia in the internal test set, demonstrating performance comparable to that of senior pathologists and superior to that of junior pathologists, while providing interpretable attention heatmaps to assist less experienced practitioners in localizing suspicious lesions. Khor et al[24] developed a twostage deep learning model for automated detection of both H. pylori and intestinal metaplasia on HEstained WSI, achieving a precision of 0.905 and F1score of 0.861 for intestinal metaplasia at the patch level, and precision of 1.0 and F1score of 0.96 at the WSI level. Fu et al[25] proposed a multitask learning framework utilizing the Virchow foundation model as a feature extractor to simultaneously predict four key pathological indicators-inflammation severity, atrophy, acute activity, and intestinal metaplasia-on gastric biopsy WSIs while accounting for their interrelationships. On a dataset of 984 WSIs, the multitask model consistently outperformed singletask learning across all four tasks, with particularly notable improvements in acute activity grading (accuracy: +6.09%) and intestinal metaplasia grading (accuracy: +5.29%).
For machine learningbased histomorphometric approaches, Lan et al[26] constructed a diagnostic model for gastric atrophy using HEstained sections from 289 gastric biopsy samples. A total of 464 pathological features were extracted via CellProfiler, and 10 key features were selected using LASSO. A logistic regression model achieved AUCs of 0.900, 0.901, and 0.818 in the training, test, and external validation sets, respectively. Two pathologists demonstrated substantial diagnostic agreement (Kappa = 0.68). These findings suggest that machine learning models based on pathological features can improve the diagnostic consistency of gastric atrophy.
Collectively, these studies demonstrate that AI can extract quantitative information from pathology slides that is difficult for the human eye to perceive, providing reproducible and objective evidence for the diagnosis of gastric atrophy and precancerous lesions, potentially addressing the subjective variability of traditional pathological diagnosis. It should be noted that some of these pathology studies assess general chronic gastritis or multiple indicators, and only the atrophy-specific results should be regarded as direct CAG evidence; findings on inflammation, activity, or intestinal metaplasia are indirect.
AI-ASSISTED ENDOSCOPIC DIAGNOSIS OF H. PYLORI INFECTION
H. pylori is the most important causative factor of CAG, inducing gastric mucosal atrophy and intestinal metaplasia. Endoscopic identification of H. pylori infection is a prerequisite for risk stratification in atrophic gastritis. Characteristic endoscopic signs of H. pylori infection include mucosal nodularity, diffuse redness, enlarged folds, and mucosal swelling. On the other hand, the presence of a regular arrangement of collecting venules (RAC) and fundic gland polyps serves as a reliable endoscopic marker for H. pylori negative status. In recent years, deep learningbased endoscopic AIassisted diagnostic systems have developed rapidly. Hu et al[27] conducted a multicenter study and developed the HOPE AI system, which was trained and validated on 308887 endoscopic images and 197 video clips. The system achieved an AUC of 0.932 in the internal validation set and AUCs ranging from 0.855 to 0.971 across seven external geographically diverse validation sets. In the video validation set, its diagnostic sensitivity reached 85.7%, significantly outperforming senior endoscopists (68.0%). Multiple systematic reviews and meta-analyses have confirmed that endoscopic AI systems exhibit excellent diagnostic performance in detecting H. pylori infection, with high pooled sensitivity and specificity[28,29]. Furthermore, Sun et al[30] developed MCCLIP, a visionlanguage foundation modelbased system for threecategory classification of H. pylori infection (current infection, past infection, and no infection) under magnetically controlled capsule endoscopy. Following largescale pretraining on over 2.42 million imagetext pairs and subsequent finetuning, the system achieved an accuracy of 89.6% in the internal validation set and 86.6% in the external validation set, with a sensitivity of 91.4% for current infection, significantly surpassing that of senior endoscopists (84.3%). AIassisted endoscopic diagnosis not only improves the detection rate of H. pylori infection but also provides technical support for the early identification of highrisk populations for atrophic gastritis. Although H. pylori infection is a major cause of CAG, studies on AI-assisted H. pylori diagnosis are not direct evidence for CAG diagnosis and should be considered indirect evidence.
AIASSISTED GASTRIC CANCER RISK PREDICTION INCORPORATING CAG
CAG, as a key precancerous stage in the development of gastric cancer, is commonly used as a risk factor for gastric cancer prediction. The predictive models have evolved from traditional machine learning approaches such as XGBoost to deep learning and other advanced techniques.
Taninaga et al[31], in a case-control study, used the XGBoost algorithm to build a machine learning model for gastric cancer risk prediction. The system integrated multi-source clinical data including endoscopic features, blood markers (e.g., HbA1c, mean corpuscular volume), demographic information, and medical history. Results showed that the model incorporating multi-dimensional information significantly outperformed models using single or few indicators, achieving an AUC of 0.899, demonstrating the critical role of multi-source data fusion in improving the accuracy of gastric cancer risk prediction in the context of CAG. In a study by Arai et al[32], a machine learning–based personalized prediction model was developed to estimate the cumulative incidence of gastric cancer in patients with chronic gastritis after their initial endoscopy. The retrospective cohort included 1099 patients. They compared a conventional Cox proportional hazards model with three machine learning algorithms-DeepSurv, gradientboosting decision tree (GBDT), and naive support vector machine-and comprehensively incorporated baseline characteristics, endoscopic atrophy (which was prospectively graded by endoscopists according to the KimuraTakemoto classification), and OLGIM/OLGA histologic staging as input variables. Over a mean followup of 5.63 years, 94 patients (8.55%) developed gastric cancer. The GBDT model achieved the best discriminative performance (Cindex = 0.84 in the test set) and significantly stratified patients into high, intermediate, and lowrisk groups (logrank P < 0.001). Age, OLGIM/OLGA stage, endoscopistassessed atrophy, and a history of nongastric malignancies were identified as key predictors. Importantly, the model generated personalized cumulative incidence curves for each patient, enabling dynamic, riskthreshold–based recommendations for surveillance intervals (e.g., 1.4 years for a highrisk patient vs 6.5 years for a lowrisk patient if a 5% incidence cutoff is applied). This model offers a practical tool for personalized endoscopic surveillance. Dilaghi et al[33] utilized AI to assess the risk of gastric neoplastic lesions (GNL) in patients with corpus-predominant CAG. The study included 103 CAG patients under long-term follow-up (22 of whom developed GNL over a median follow-up of 60 months). Using 30 clinical, endoscopic, and histological variables, the study employed Fisher scoring combined with a one-class support vector machine for variable selection and model construction. The model achieved sensitivity/specificity of 92.3%/77.9% for predicting epithelial GNL and 77.8%/92.6% for neuroendocrine tumors. Kirita et al[34] evaluated the value of combining AI-based endoscopic diagnosis with the Kimura-Takemoto classification for gastric cancer risk stratification following H. pylori eradication. The study designed a combined rule: Expert-based traditional classification alongside AI-based scoring of atrophy in key gastric regions; high risk was assigned only when experts diagnosed open-type atrophy and AI scores indicated moderate-to-severe atrophy. In a validation set of 72 patients, this combined approach achieved a predictive AUC of 0.674, showing better performance than traditional classification alone and more clearly distinguishing high-risk from low-risk groups (P = 0.014). Although the AUC and other evaluation metrics of these predictive models appear satisfactory, the low number of positive cases in which atrophic gastritis progresses to gastric cancer naturally leads to an overly optimistic assessment of AUC due to class imbalance. Therefore, future studies should also report metrics more robust to imbalanced data, such as the Matthews correlation coefficient and F1-score, along with the positive rate, to provide a fairer evaluation of the model’s predictive performance.
AI-ASSISTED PATIENT MANAGEMENT AND EDUCATION IN CAG
With the rise of large language models (LLMs), AI applications in doctor-patient communication and chronic disease management are being explored. Theoretically, intelligent question and answer systems based on LLMs can provide patients with 24/7 personalized answers regarding disease knowledge, dietary advice, medication instructions, and the importance of follow-up visits, alleviating the burden on outpatient clinics. Critically, these platforms are not passive monitors; they actively analyze patientinput logs to deliver personalized reminders for lifestyle modifications (e.g., diet and smoking cessation), track medication adherence (e.g., H. pylori eradication therapy compliance), and monitor followup appointment attendance. Such active engagement empowers patients to take ownership of their disease course, thus aiding effective long-term, out-of-hospital disease management. Fu et al[35] developed an endoscopic nursing chatbot named Endo-chat based on retrieval-augmented generation (RAG) technology, which integrated 17 relevant clinical guidelines and internal departmental data into its knowledge base to accurately answer various patient queries before and after endoscopic procedures. In a randomized questionnaire survey involving 200 patients, Endo-chat significantly outperformed the manual nurse group in response timeliness (91% instant replies vs 18%), information completeness (88% fully covered vs 71%), and patient satisfaction (54% very satisfied vs 41%) (all P < 0.05), while showing no statistical difference in answer accuracy (82% vs 91%, P = 0.085). For the specific issue of follow-up surveillance for atrophic gastritis, Endo-chat was able to provide detailed, stratified recommendations based on the latest guidelines. In contrast, the generic ChatGPT only gave vague and nonspecific responses, highlighting the marked advantage of RAG in acquiring domainspecific knowledge. All participating nurses believed that AI could reduce their workload, and most patients supported AIassisted services. This study confirms that a RAGenhanced LLM can effectively compensate for the shortcomings of generic models in terms of outdated medical knowledge and lack of personalization, thus serving as an efficient auxiliary tool for digestive endoscopic nursing. Wang and Ye[36] conducted a study aimed at exploring the application value of AI agents in postdischarge selfmanagement for patients with chronic gastritis. They built an agent framework based on LLM, incorporating RAG and search engine tools, and collected realworld patient questions from clinical settings. These questions were categorized by difficulty level and health management task scenarios, and the framework was evaluated using a dualevaluation system combining automated metrics and expert manual scoring by five clinical specialists, with comparisons against several generalpurpose and medicalspecific LLMs. The results showed that the AI agent significantly outperformed standalone LLMs in handling highcomplexity questions, achieved a physicianrated safety score of 4.98 out of 5, demonstrated minimal variability in stability tests over 30 repeated experiments, and surpassed stateoftheart medical models such as Baichuan14BM1 and MedGemma27B. However, all evaluations were conducted in an offline setting based solely on response quality, and none of the agentgenerated answers were actually deployed to real patients, nor were any clinical outcomes (e.g., symptom control, medication adherence, or changes in healthcareseeking behavior) measured. Therefore, the claimed clinical benefits-such as reducing physician workload and improving home care quality-remain theoretical and require validation through prospective clinical studies. These LLM-based systems are not CAG-specific and have not been deployed to real patients; their findings are therefore indirect and exploratory, and should not be interpreted as direct evidence for CAG management.
CONCLUSION
Existing deep learning-based AI-assisted diagnostic models have shown promising diagnostic performance in identifying, classifying, and grading atrophy and intestinal metaplasia on endoscopic images, sometimes matching or exceeding average endoscopists. These findings suggest that AI may help reduce subjective bias and diagnostic heterogeneity. These models may also offer potential clinical value, such as identifying low-risk lesions that could reduce unnecessary biopsies, or assisting in risk stratification that might help tailor surveillance intervals. It should be emphasized that these potential benefits have not yet been confirmed sufficiently in prospective clinical trials, and no AI system described here is ready for guideline-based use.
However, several challenges remain before these technologies can be widely adopted in clinical practice. First, model development heavily relies on largescale, highquality data with expertlevel precise annotation. Most existing studies are based on singlecenter, retrospective data, which limits model generalizability and may lead to performance degradation when facing realworld image artifacts or data from different medical centers. This data heterogeneity, coupled with the lack of standardized prospective validation across diverse populations and endoscopic platforms, remains a major barrier to clinical adoption. Therefore, there is an urgent need to establish standardized, multicenter shared image databases with pathological goldstandard labels. Second, most current models still have relatively limited functionality. Future efforts should integrate multiple tasks, including atrophy detection, grading, risk stratification, and H. pylori status identification, and explore the incorporation of patient management functions to build an integrated system covering the entire pathway of “screening, diagnosis, risk prediction and health management”. Third, the “black-box” nature of deep learning models remains a critical barrier to clinical trust; future developments must prioritize model interpretability-for example, through visualization of attention maps or lesion-saliency regions-so that clinicians can understand and verify AI-generated decisions. Fourth, existing models predominantly rely on single-modal endoscopic imaging; integrating multi-modal data, such as serological biomarkers (e.g., pepsinogen and gastrin), genomic or epigenetic profiles, and detailed patient histories, holds great promise for enabling holistic and precise risk assessment. Fifth, existing predictive models for gastric cancer risk are mostly based on retrospective datasets with small sample sizes (typically hundreds to thousands), incomplete data and narrow objectives (mainly focused on gastric cancer), offering limited practical value for the general population. Furthermore, in the domain of patient management, there is a lack of rigorous prospective, multicenter randomized controlled trials to verify whether AI assistance can truly improve patients’ disease awareness and health behaviors. Finally, issues concerning medical safety and liability require urgent clarification-specifically, how accountability should be assigned when AI predictions are erroneous or discordant with pathological outcomes, which demands clear regulatory and medicolegal frameworks.
Direct CAG evidence currently comes mainly from studies on endoscopic or pathological atrophy detection and grading, whereas evidence on H. pylori, intestinal metaplasia, general chronic gastritis, endoscopic nursing, and non-deployed LLM systems is indirect and should be interpreted with caution. Moving forward, rather than merely cataloging these limitations, future research should actively pivot toward establishing large-scale, publicly accessible, prospective multi-center cohorts with standardized imaging and pathological annotations, developing interpretable and generalizable foundation models tailored to gastrointestinal endoscopy, and conducting pragmatic randomized controlled trials that evaluate AI-assisted decision-making within real-world clinical workflows. Ultimately, the goal is to transition these systems from research prototypes to indispensable clinical decision-support tools that are seamlessly embedded into routine endoscopic practice, with robust governance frameworks ensuring patient safety, data privacy, and clear responsibility boundaries.
Du Y, Bai Y, Xie P, Fang J, Wang X, Hou X, Tian D, Wang C, Liu Y, Sha W, Wang B, Li Y, Zhang G, Li Y, Shi R, Xu J, Li Y, Huang M, Han S, Liu J, Ren X, Xie P, Wang Z, Cui L, Sheng J, Luo H, Wang Z, Zhao X, Dai N, Nie Y, Zou Y, Xia B, Fan Z, Chen Z, Lin S, Li ZS; Chinese Chronic Gastritis Research group. Chronic gastritis in China: a national multi-center survey.BMC Gastroenterol. 2014;14:21.
[RCA] [PubMed] [DOI] [Full Text][Cited by in Crossref: 131][Cited by in RCA: 123][Article Influence: 10.3][Reference Citation Analysis (1)]
Lin N, Yu T, Zheng W, Hu H, Xiang L, Ye G, Zhong X, Ye B, Wang R, Deng W, Li J, Wang X, Han F, Zhuang K, Zhang D, Xu H, Ding J, Zhang X, Shen Y, Lin H, Zhang Z, Kim JJ, Liu J, Hu W, Duan H, Si J. Simultaneous Recognition of Atrophic Gastritis and Intestinal Metaplasia on White Light Endoscopic Images Based on Convolutional Neural Networks: A Multicenter Study.Clin Transl Gastroenterol. 2021;12:e00385.
[RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)][Cited by in Crossref: 33][Cited by in RCA: 27][Article Influence: 5.4][Reference Citation Analysis (1)]
Chen S, Xu L, Yan L, Zhang J, Zhou X, Wang J, Yan T, Wang J, He X, Ma H, Zhang X, Zhu S, Zhang Y, Xu C, Gao J, Ji X, Bai D, Chen Y, Chen H, Ke Y, Li L, Yu C, Mao X, Li T, Chen Y. A novel endoscopic artificial intelligence system to assist in the diagnosis of autoimmune gastritis: a multicenter study.Endoscopy. 2025;57:299-309.
[RCA] [PubMed] [DOI] [Full Text][Cited by in Crossref: 18][Cited by in RCA: 23][Article Influence: 23.0][Reference Citation Analysis (0)]
Fang S, Liu Z, Qiu Q, Tang Z, Yang Y, Kuang Z, Du X, Xiao S, Liu Y, Luo Y, Gu L, Tian L, Liang X, Fan G, Zhang Y, Zhang P, Zhou W, Liu X, Tian J, Wei W. Diagnosing and grading gastric atrophy and intestinal metaplasia using semi-supervised deep learning on pathological images: development and validation study.Gastric Cancer. 2024;27:343-354.
[RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)][Cited by in Crossref: 26][Cited by in RCA: 28][Article Influence: 14.0][Reference Citation Analysis (1)]
Ma M, Zeng X, Qu L, Sheng X, Ren H, Chen W, Li B, You Q, Xiao L, Wang Y, Dai M, Zhang B, Lu C, Sheng W, Huang D. Advancing Automatic Gastritis Diagnosis: An Interpretable Multilabel Deep Learning Framework for the Simultaneous Assessment of Multiple Indicators.Am J Pathol. 2024;194:1538-1549.
[RCA] [PubMed] [DOI] [Full Text][Cited by in RCA: 3][Reference Citation Analysis (0)]
Khor LY, Neo CC, Prathaban K, Choa E, Quah WK, Lum EN, Chen R, Heng SY, Koh VC, Seow JX, Jegannathan N, Shi R, Loong S, Song LH, Natarajan A, Ravi S, Oh KS, Cheng CL. Deep learning model for automated detection of Helicobacter pylori and intestinal metaplasia on gastric biopsy digital whole slide images.Am J Clin Pathol. 2025;164:830-839.
[RCA] [PubMed] [DOI] [Full Text][Cited by in Crossref: 2][Cited by in RCA: 3][Article Influence: 3.0][Reference Citation Analysis (0)]
Sun X, Liu J, Wu L, Chen X, Ma X, Teng F, Zhang T, Su H, Fan X, Li J, Xu S, Jin P, Jiao H. AI-powered three-category Helicobacter pylori diagnosis via magnetic controlled capsule endoscopy: a multicenter validation of a vision-language model.Front Microbiol. 2025;16:1687021.
[RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)][Reference Citation Analysis (0)]
Kirita K, Futagami S, Nakamura K, Agawa S, Ueki N, Higuchi K, Habiro M, Kawawa R, Kato Y, Tada T, Iwakiri K. Combination of artificial intelligence endoscopic diagnosis and Kimura-Takemoto classification determined by endoscopic experts may effectively evaluate the stratification of gastric atrophy in post-eradication status.DEN Open. 2025;5:e70029.
[RCA] [PubMed] [DOI] [Full Text][Cited by in RCA: 4][Reference Citation Analysis (0)]
Creativity or innovation: Grade A, Grade A, Grade A
Scientific significance: Grade A, Grade A, Grade A
P-Reviewer: Baryshnikova NV, MD, PhD, Associate Professor, Russia; Seshadri PR, Associate Professor, India S-Editor: Liu H L-Editor: A P-Editor: Zhang YL
