Ying JX, Yan SY, Fu XY, Zhou YJ, Zhou JJ, Yang Y, Zhou XB, Wang ZZ, Li SW, Fang LN, Mao XL. Artificial intelligence-assisted endoscopists improve the detection rate of high-risk gastric lesions: A propensity score-matched retrospective study. World J Gastroenterol 2026; 32(21): 117299 [DOI: 10.3748/wjg.v32.i21.117299]
Corresponding Author of This Article
Xin-Li Mao, MD, PhD, Department of Gastroenterology, Taizhou Hospital of Zhejiang Province Affiliated to Wenzhou Medical University, No. 150 Ximen Street, Linhai 317000, Zhejiang Province, China. maoxl@enzemed.com
Research Domain of This Article
Gastroenterology & Hepatology
Article-Type of This Article
research-article
Open-Access Policy of This Article
This article is an open-access article which was selected by an in-house editor and fully peer-reviewed by external reviewers. It is distributed in accordance with the Creative Commons Attribution Non Commercial (CC BY-NC 4.0) license, which permits others to distribute, remix, adapt, build upon this work non-commercially, and license their derivative works on different terms, provided the original work is properly cited and the use is non-commercial. See: http://creativecommons.org/licenses/by-nc/4.0/
Baishideng Publishing Group Inc, 7041 Koll Center Parkway, Suite 160, Pleasanton, CA 94566, USA
Share the Article
Ying JX, Yan SY, Fu XY, Zhou YJ, Zhou JJ, Yang Y, Zhou XB, Wang ZZ, Li SW, Fang LN, Mao XL. Artificial intelligence-assisted endoscopists improve the detection rate of high-risk gastric lesions: A propensity score-matched retrospective study. World J Gastroenterol 2026; 32(21): 117299 [DOI: 10.3748/wjg.v32.i21.117299]
Jia-Xiu Ying, Xin-Li Mao, School of Medicine, Shaoxing University, Shaoxing 312000, Zhejiang Province, China
Jia-Xiu Ying, Xin-Li Mao, Taizhou Hospital of Zhejiang Province, Shaoxing University, Linhai 317000, Zhejiang Province, China
Si-Yan Yan, Department of Gastroenterology, The Affiliated Hospital of Jiaxing University, Jiaxing 314000, Zhejiang Province, China
Xin-Yu Fu, Department of Gastroenterology, Dalian Medical University, Dalian 116000, Liaoning Province, China
Yi-Jing Zhou, Jing-Jing Zhou, Xian-Bin Zhou, Zhen-Zhen Wang, Shao-Wei Li, Xin-Li Mao, Department of Gastroenterology, Taizhou Hospital of Zhejiang Province Affiliated to Wenzhou Medical University, Linhai 317000, Zhejiang Province, China
Jing-Jing Zhou, Xian-Bin Zhou, Zhen-Zhen Wang, Shao-Wei Li, Xin-Li Mao, Key Laboratory of Minimally Invasive Techniques and Rapid Rehabilitation of Digestive System Tumor of Zhejiang Province, Taizhou Hospital Affiliated to Wenzhou Medical University, Linhai 317000, Zhejiang Province, China
Jing-Jing Zhou, Xian-Bin Zhou, Zhen-Zhen Wang, Shao-Wei Li, Xin-Li Mao, Institute of Digestive Disease, Taizhou Hospital Affiliated to Wenzhou Medical University, Linhai 317000, Zhejiang Province, China
Yan Yang, Taizhou Hospital of Zhejiang Province, Zhejiang University, Linhai 317000, Zhejiang Province, China
Li-Na Fang, Endoscopy Center, Taizhou Hospital of Zhejiang Province Affiliated to Wenzhou Medical University, Linhai 317000, Zhejiang Province, China
Co-corresponding authors: Li-Na Fang and Xin-Li Mao.
Author contributions: Ying JX and Yan SY contributed equally to this article, they are the co-first authors of this manuscript; Ying JX, Yan SY, Fu XY, and Mao XL conceived and designed the study; Ying JX, Yan SY, Fu XY, Zhou YJ, and Yang Y participated in the data analysis, writing, and editing; Ying JX, Yan SY, Fu XY, Zhou YJ, Zhou JJ, Yang Y, Zhou XB, Wang ZZ, Li SW, Fang LN, and Mao XL performed validation and editing; Fang LN and Mao XL contributed equally to this article, they are the co-corresponding authors of this manuscript; and all authors have read and approved the final manuscript.
AI contribution statement: We used DeepL to help with translation and language polishing. We wrote every section ourselves. For parts of the Abstract, Introduction, Discussion, and Conclusion, we first drafted the content in Chinese and then used DeepL to help translate and polish the language. After that, we carefully reviewed and revised all the AI-assisted text to make sure the scientific content was accurate. We used DeepL for translation and language polishing in some parts of the manuscript. However, we did not use any AI tools for data analysis, and DeepL was only used to refine the language, not to help with writing the content itself. We designed the study, analyzed the data, and interpreted the results entirely on our own without any AI assistance. We created all the images and figures ourselves, without using any AI generation tools.
Supported by the “Pioneer” and “Leading Goose” R&D Program of Zhejiang, No. 2025C02139; the Medical Science and Technology Project of Zhejiang Province, No. 2024KY1788; the Program of Taizhou Science and Technology Grant, No. 22ywb09, No. 23ywa33, No. 23ywa35, and No. 23ywb03; and the Open Project Program of Key Laboratory of Minimally Invasive Techniques and Rapid Rehabilitation of Digestive System Tumor of Zhejiang Province, No. 21SZDSYS01.
Institutional review board statement: This study was approved by the Medical Ethics Committee of Taizhou Hospital of Zhejiang Province, approval No. K20240936.
Informed consent statement: The ethics committee has exempted the informed consent of the patients.
Conflict-of-interest statement: All the authors report no relevant conflicts of interest for this article.
Data sharing statement: Statistical code, and dataset available from the corresponding author at maoxl@enzemed.com. Participants was not obtained consent but the presented data are anonymized and risk of identification is low.
Corresponding author: Xin-Li Mao, MD, PhD, Department of Gastroenterology, Taizhou Hospital of Zhejiang Province Affiliated to Wenzhou Medical University, No. 150 Ximen Street, Linhai 317000, Zhejiang Province, China. maoxl@enzemed.com
Received: December 8, 2025 Revised: January 29, 2026 Accepted: February 26, 2026 Published online: June 7, 2026 Processing time: 171 Days and 21.4 Hours
Abstract
BACKGROUND
Early detection of high-risk gastric lesions (HrGLs) during esophagogastroduodenoscopy (EGD) is challenging, and artificial intelligence (AI) may enhance its diagnostic accuracy.
AIM
To evaluate the efficacy of AI in improving HrGLs detection during EGD.
METHODS
We included patients who underwent EGD examinations at the Taizhou Hospital between August 1, 2023, and July 31, 2024. The primary outcome was the detection rate of HrGLs between the AI-assisted and non-AI-assisted groups. The secondary objective was to explore the impact of AI on various aspects of the gastroscopy process in clinical practice.
RESULTS
Through propensity score matching, 15528 patients (7764 in the AI-assisted group and 7764 in the non-AI-assisted group) were included in the final analysis. A total of 110 HrGLs (AI: 73 vs non-AI: 37) were detected. The detection rate of HrGLs in the AI-assisted group was significantly higher than that in the non-AI-assisted group (0.94% vs 0.48%, odds ratio = 1.98, 95% confidence interval: 1.33-2.95, P = 0.001). This enhancement was consistently observed across all HrGL subtypes, including high-grade intraepithelial neoplasia, low-grade intraepithelial neoplasia, and early gastric cancer. Notably, AI assistance was of the greatest benefit when combined with experienced endoscopists and during procedures performed under anesthesia.
CONCLUSION
AI significantly improved HrGLs detection, highlighting its value as a clinical adjunct.
Core Tip: Early detection of high-risk gastric lesions remains challenging during routine esophagogastroduodenoscopy. In this large, propensity score-matched cohort study including 15528 patients, we demonstrated that artificial intelligence (AI) significantly improves the detection rate of high-risk gastric lesions, particularly high-grade intraepithelial neoplasia, low-grade intraepithelial neoplasia, and early gastric cancer. Notably, AI assistance was most beneficial when combined with experienced endoscopists and procedures performed under anesthesia. These findings highlight the clinical value of AI as an effective adjunct for enhancing diagnostic accuracy and optimizing endoscopic practice.
Citation: Ying JX, Yan SY, Fu XY, Zhou YJ, Zhou JJ, Yang Y, Zhou XB, Wang ZZ, Li SW, Fang LN, Mao XL. Artificial intelligence-assisted endoscopists improve the detection rate of high-risk gastric lesions: A propensity score-matched retrospective study. World J Gastroenterol 2026; 32(21): 117299
Gastric cancer is one of the most common malignant tumors worldwide and ranks fifth in terms of incidence and mortality. According to the latest statistics from the International Agency for Research on Cancer, an estimated 1.089 million new cases of gastric cancer and 768000 related deaths occurred worldwide in 2020[1], with nearly three-quarters of cases reported in Asian countries. China, a high-risk country for gastric cancer, carries a substantial disease burden. Data from the World Health Organization indicate 478000 new cases of malignant gastric tumors in China in 2020, accounting for 43.9% of the global gastric cancer incidence. Additionally, 373000 gastric cancer-related deaths were reported in China[2], representing 48.5% of global cancer mortality. These figures highlight the urgent need for preventive measures, early diagnosis, and timely treatment.
Neoplastic transformation of the gastric mucosa progresses through a precancerous cascade, beginning with atrophic gastritis and advancing through intestinal metaplasia, low-grade intraepithelial neoplasia (LGIN), and high-grade intraepithelial neoplasia (HGIN), before culminating in gastric cancer[3]. Early gastric cancer (EGC) is defined as a lesion confined to the mucosa or submucosa, regardless of the presence of regional lymph node metastases[4]. Because these stages are typically asymptomatic or associated with nonspecific manifestations, patients often remain undiagnosed until later stages. Therefore, countries with a high incidence of gastric cancer, such as Japan and South Korea, have established nationwide screening programs that recommend regular endoscopic examinations for individuals aged 50 and older. Large-scale cohort studies have demonstrated that systematic screening strategies can reduce gastric cancer mortality by up to 40%[5,6]. Patients with advanced gastric cancer have a poor prognosis, with a median overall survival of less than 12 months. In contrast, patients with EGC treated with endoscopic submucosal dissection achieve a 5-year survival rate exceeding 90%[7,8]. These insights highlight the importance of early detection and timely treatment to improve patient outcomes.
White light endoscopy (WLE) is the most commonly used method for gastric cancer screening[9]. Nevertheless, its diagnostic performance remains suboptimal. Prior studies have demonstrated a low concordance rate between WLE and histopathological findings for precancerous gastric lesions, ranging from 46.8% to 67.0%[10]. In recent years, advanced endoscopic imaging techniques, such as chromoendoscopy and narrow-band imaging (NBI), have improved the detection rate of gastric cancer[11]. Nevertheless, subtle precancerous lesions can still be overlooked. A retrospective study from the United Kingdom revealed that 8.3% of gastric cancer cases were missed on endoscopy within the three years preceding diagnosis[12]. Factors contributing to missed diagnoses may include small lesion size, flat or depressed appearance, anatomical location (lesser curvature or posterior wall), inadequate observation time, and insufficient biopsy sampling[13,14]. Furthermore, a multicenter cohort study showed that the missed diagnosis of gastric cancer is independently associated with proton pump inhibitor therapy, previous Billroth II anastomosis, and absence of alarm symptoms[15]. Additionally, the quality of endoscopy is influenced by the endoscopist’s skill level and subjective judgment, which may result in certain lesion areas being overlooked and lead to a missed diagnosis of EGC[16,17]. Therefore, improving the endoscopy quality is crucial.
Recently, the application of artificial intelligence (AI) in the medical field, particularly in the diagnosis of gastrointestinal diseases, has garnered widespread attention. Neural network technologies, represented by Convolutional Neural Networks (CNNs), have made significant progress in image recognition, enhancing the diagnostic capabilities of endoscopists[18]. CNNs are trained using large volumes of annotated images and videos, meticulously labeled for tasks such as image classification and segmentation. In gastric image analysis, the process begins with the acquisition and preprocessing of gastric images. These images are then annotated by medical experts to delineate cancerous regions. The labeled images are used to train the AI model, enabling it to learn discriminative features of the tumor tissue. Once trained, the model can accurately segment tumor regions in new, unseen images to assist in diagnostic procedures[19]. A meta-analysis of 17 studies comprising 51446 images and 174 videos involving 5539 patients found that the pooled sensitivity and specificity of CNNs in predicting the infiltration depth of gastric cancer were 82% and 90%, respectively, outperforming expert endoscopists[20]. Wu et al[21] developed an AI system for real-time detection of gastric neoplasms and validated its potential to reduce the rate of missed diagnoses in multicenter studies. However, these studies did not delve deeply into the impact of this system on endoscopists. Accordingly, this study employed a propensity score matching (PSM)-based retrospective design to evaluate the performance of an AI-assisted system in improving the detection of high-risk gastric lesions (HrGLs), with a particular focus on its impact on endoscopist ability to identify these lesions. Additionally, the influence of AI on key procedural aspects of gastroscopy in a clinical setting was assessed to provide a more comprehensive scientific foundation for the application of AI technology in the field of endoscopy.
MATERIALS AND METHODS
Patients
Patients undergoing EGD examinations at Taizhou Hospital from August 1, 2023 to July 31, 2024, were enrolled in our study. The Ethics Committee of Taizhou Hospital of Zhejiang Province approved this study, No. K20240936. The primary objective of EGD screening at our institution is to detect early upper gastrointestinal cancer. All patients, regardless of group assignment, adhered to the same standardized institutional protocols for preoperative preparation (including fasting requirements) and peri-procedural management. The decision to perform biopsies, the number of biopsies, and specific sampling sites were left to the discretion of the performing endoscopist, based on their visual findings and, in the AI-assisted group, the AI system’s prompts. Next, the patients were assigned to examination rooms, some of which were equipped with the AI system, in a quasi-random manner based on room availability.
Inclusion criteria: Patients who underwent EGD examination at Taizhou Hospital.
Exclusion criteria: Patients who underwent emergency EGDs, underwent endoscopic treatments, did not undergo biopsies, had a history of upper gastrointestinal malignancy, had previous upper gastrointestinal surgery, had multiple gastrointestinal endoscopies (for whom only the first-time data were extracted), or had incomplete data.
AI-assisted system
We utilized an AI-assisted system called ENDOANGEL, which is based on deep learning. ENDOANGEL-ELD integrates two main functions and five deep CNN modules. It identifies anatomical sites during endoscopy and detects gastric lesions, categorizing them as high-risk or low-risk under WLE and NBI modes. The system displays predictive results in real-time on a monitor, highlighting lesion sites with red or blue boxes to indicate high- and low-risk lesions, respectively. The effectiveness of this system has been validated in several clinical studies[21]. Figure 1 shows the clinical application of ENDOANGEL-ELD.
Figure 1
Clinical application diagram of ENDOANGEL-ELD.
Data collection
We collected relevant data from the endoscopy system, including age, sex, examination room, anesthesia use, biopsy details, endoscopist details, insertion and withdrawal times, endoscopic reports, and pathological diagnoses. The endoscopists were divided into AI-assisted and non-AI-assisted groups. The procedures were conducted by 25 endoscopists, each with at least one year of experience. Senior endoscopists had over 10 years of experience with at least 3000 EGD procedures, intermediate endoscopists had 5-10 years of experience, and junior endoscopists had 1-5 years of experience, conducting at least 130 EGD procedures. The biopsy number represents the total number of biopsies taken from the stomach during each EGD. The anatomical locations of HrGLs were categorized into the following regions: The cardia, fundus, body, angle, antrum, and pylorus. HrGLs include intraepithelial neoplasia (e.g., LGIN and HGIN) and EGC. Postoperative histopathology is considered the gold standard for determining the nature of lesions. For patients who did not undergo endoscopic or surgical resection, pathological classification was based on biopsy pathology. Certain HGIN lesions, owing to their similar biological behavior to that of EGC, are usually classified together with EGC. Accordingly, in this study, EGC was defined as HGIN or as a tumor confined to the gastric mucosa or submucosa according to postoperative pathology[22]. To avoid diagnostic bias, the pathologists did not know which group (AI-assisted or non-AI-assisted) each specimen came from. They also did not see the endoscopic findings when evaluating the tissue samples. Pathological analyses were performed by two experienced pathologists, with postoperative or biopsy pathology serving as the gold standard.
Outcomes
The primary goal was to compare the detection rates of HrGLs between the AI-assisted and non-AI-assisted groups using PSM. The secondary outcomes assessed the impact of endoscopist seniority, anesthesia use, and biopsy number on AI-assisted detection rates.
Statistical analysis
Statistical analyses were conducted using R software (version 4.5.0). PSM was performed using the “MatchIt” package with the genetic matching algorithm (method = “genetic”). Matching covariates included age, sex, anesthesia use, and endoscopist experience. Given the clinical importance of endoscopist experience as a potential confounder and its categorical nature (junior, medium, senior), exact matching was enforced for this variable to ensure perfect balance between groups. Covariate balance was assessed using standardized mean differences (SMDs) for all covariates. An SMD below 0.1 was considered indicative of adequate balance. Both pre- and post-matching SMDs were calculated to evaluate matching effectiveness. The cobalt package in R was used for balance diagnostics and SMD computation. We performed a formal causal mediation analysis using the mediation package in R. We used simple mediation models, including paths a, b, c and c’ (shown in Figure 2). Linear regression models were used to explore the relationship between AI assistance and biopsy number. Logistic regression model estimated the effect of biopsy number on HrGLs detection and the direct effect of AI, with both models adjusted for age, sex, anesthesia use, and endoscopist experience. Path a assessed the effect of AI assistance and biopsy number (mediator). Path b evaluated the association between biopsy number and HrGLs detection. Path c (direct effect) provided an estimate of the direct effect of biopsy number on the association between AI assistance and HrGLs detection. The total, direct, and indirect effects of AI assistance on HrGLs detection are reported as risk differences with 95% confidence intervals (CI) on the probability scale (0-1). Path coefficients (a-path and b-path) are reported as estimates with their standard errors and P values. The proportion mediated was calculated as (indirect effect/total effect) × 100%. An indirect effect with a 95%CI excluding zero was considered evidence of mediation. An unadjusted analysis was also conducted as a sensitivity check. Continuous variables are summarized as mean ± SD or median and interquartile range based on distribution. Student’s t-test, Mann-Whitney U test, χ² test, and Fisher’s exact test were applied as appropriate. Logistic regression was used to assess the correlation between AI use and endoscopist experience. Statistical significance was set at P < 0.05.
Figure 2 Path diagram of the mediation analysis models.
AI: Artificial intelligence; HrGLs: High-risk gastric lesions; CI: Confidence interval.
RESULTS
Baseline characteristics of the study population and PSM
This study analyzed 40735 EGD cases, excluding 261 procedures performed by less experienced endoscopists, resulting in a final sample of 40474 cases. The study population comprised 19519 men (48.23%) and 20955 women (51.77%), with an average age of 52.6 years and a median age of 54 years. Twenty-five endoscopists participated: 15 senior endoscopists performed 20629 procedures, 3 intermediate endoscopists completed 13265 procedures, and 7 junior endoscopists conducted 6580 procedures. Of the total procedures, 10532 were performed with AI assistance and 29942 without it, while 29600 were conducted under anesthesia and 10874 without anesthesia. The median number of biopsies performed per procedure was two. In total, 255 HrGLs were detected: 88 in the AI-assisted group and 167 in the non-AI group. These included 152 LGIN (55 vs 97), 66 HGIN (21 vs 45), and 104 EGC (39 vs 65). In terms of endoscopist experience, senior endoscopists performed 20629 procedures, mid-career endoscopists performed 13265 procedures, and junior endoscopists performed 6580 procedures (Table 1). Figure 3 shows the study flowchart.
PSM was employed to address potential biases arising from differences in the number of procedures between groups. After matching, 15528 procedures were analyzed [7841 men (48.18%), 8047 women (51.82%); mean age 51.7 years, median age 53 years]. A total of 13196 procedures were performed under anesthesia. A total of 110 HrGLs were detected: 73 in the AI-assisted group and 37 in the non-AI group, comprising 73 LGIN (47 vs 26), 23 HGIN (17 vs 6), and 43 EGC (31 vs 12). In terms of procedure volume, senior endoscopists performed 7516 procedures, mid-career endoscopists performed 6260 procedures, and junior endoscopists performed 1752 procedures. Covariate balance was quantitatively assessed using SMDs. After PSM, excellent balance was achieved for all matching covariates, with all SMDs below the threshold of 0.1 (Table 1).
Effect of AI assistance on HrGLs detection rate
Differences in endoscopic diagnostic outcomes between the AI-assisted and non-AI groups are presented in Table 2. Univariate logistic regression analysis demonstrated that AI assistance significantly improved the detection rates of subtle HrGLs [odds ratio (OR) = 1.98, 95%CI: 1.33-2.95, P = 0.001], HGIN (OR = 2.84, 95%CI: 1.12-7.20, P = 0.028), LGIN (OR = 1.81, 95%CI: 1.12-2.93, P = 0.015), and EGC (OR = 2.59, 95%CI: 1.33-5.05, P = 0.005).
Table 2 The impact of artificial intelligence assistance on the detection rate of high-risk gastric lesions in propensity score matching groups, n (%).
The influence of AI assistance on the detection of HrGLs by endoscopic physicians with different seniority
In exploratory subgroup analyses stratified by endoscopist experience, AI assistance was associated with significantly improved detection rates among senior endoscopists for HrGLs (OR = 2.23, 95%CI: 1.37-3.61, P = 0.001), HGIN (OR = 3.67, 95%CI: 1.02-13.18, P = 0.046), LGIN (OR = 1.96, 95%CI: 1.21-3.41, P = 0.018), and EGC (OR = 3.35, 95%CI: 1.34-8.34, P = 0.010). For intermediate and junior endoscopists, point estimates also suggested potential benefits (ORs > 1.0), but these did not reach statistical significance (Table 3).
Table 3 The impact of artificial intelligence assistance on the detection of high-risk gastric lesions by endoscopists of different levels of experience in propensity score matching groups.
Influence of AI assistance on detection rate of HrGLs in anesthetized patients
In procedures under anesthesia, AI assistance improved detection rates of HrGLs (OR = 2.04, 95%CI: 1.31-3.19, P = 0.002), LGIN (OR = 2.18, 95%CI: 1.27-3.75, P = 0.005), and EGC (OR = 2.17, 95%CI: 1.03-4.58, P = 0.043), but not HGIN (P = 0.052; Table 4). Moreover, a strong correlation was observed between the number of biopsies and the detection rates. An increase in biopsies led to significantly higher detection rates for HrGLs (OR = 1.89, 95%CI: 1.71-2.08, P < 0.001), LGIN (OR = 1.67, 95%CI: 1.49-1.87, P < 0.001), HGIN (OR = 1.85, 95%CI: 1.58-2.17, P < 0.001), and EGC (OR = 1.96, 95%CI: 1.72-2.23, P < 0.001).
Table 4 The impact of artificial intelligence assistance on the detection rate of high-risk gastric lesions in anesthetized patients of propensity score matching groups, n (%).
The influence of AI assistance on the detection rate of HrGLs in different parts of the stomach
A total of 115 HrGLs were detected in 110 patients, 5 of whom had composite lesions. The anatomical distribution of the lesions was as follows: Cardia, 17 (0.11%); gastric fundus, 4 (0.03%); gastric body, 16 (0.10%); gastric angle, 14 (0.09%); gastric antrum, 58 (0.37%); and gastric pylorus, 6 (0.04%; Table 5). Most high-risk lesions were detected in the gastric antrum in both groups. AI assistance significantly improved HrGLs detection in the gastric angle (0.14% vs 0.04%; OR = 3.67; 95%CI: 1.02-13.16; P = 0.046), whereas no differences were observed at other anatomical sites.
Table 5 Influence of artificial intelligence assistance on detection rate of high-risk gastric lesions in different parts of stomach in propensity score matching groups, n (%).
Logistic regression analysis of HrGLs detection in the PSM group
Univariate logistic regression indicated that the age and number of biopsies were significant predictors of HrGLs detection. Each additional year of age increased the detection odds by 8% (OR = 1.08, 95%CI: 1.06-1.09, P < 0.001), and each additional biopsy increased odds by 89% (OR = 1.89, 95%CI: 1.71-2.08, P < 0.001). AI significantly enhanced the detection rates, showing 98% higher odds compared to the non-AI group (OR = 1.98, 95%CI: 1.33-2.95, P = 0.001). The detection rate of subtle HrGLs was higher among the high-seniority endoscopists than that among the low-seniority endoscopists (OR = 3.62, 95%CI: 1.46-8.95, P = 0.005). However, no significant increase was observed in medium-seniority endoscopists (OR = 1.57, 95%CI: 0.61-4.07, P = 0.354). The use of anesthesia did not significantly improve the detection of subtle HrGLs (OR = 0.75, 95%CI: 0.46-1.21, P = 0.232, Table 6).
Table 6 Logistic regression analysis for detection rate of high-risk gastric lesions in propensity score matching groups.
Mediation and stratified analysis regarding biopsy number
To address whether the improvement in HrGLs detection with AI assistance was mediated through an increase in biopsy number, we performed both stratified and formal mediation analyses. Stratified analysis by biopsy count (Table 7) revealed that AI assistance was most beneficial when only one biopsy was taken (OR = 5.53, 95%CI: 1.46-35.99, P = 0.027), with diminishing effects as biopsy number increased (2 biopsies: OR = 2.08, P = 0.138; ≥ 3 biopsies: OR = 1.60, P = 0.145).
Table 7 Stratified analysis by biopsy number for detection rate of high-risk gastric lesions in propensity score matching groups.
Formal mediation analysis (Table 8) showed that in the covariate-adjusted model, the total effect of AI assistance on the HrGLs detection rate was 0.0044 (95%CI: 0.0020-0.0069). The direct effect of AI, independent of biopsy number, was 0.0039 (95%CI: 0.0017-0.0064), accounting for 95.1% of the total effect. The indirect effect mediated through increased biopsy number was 0.0002 (95%CI: 0.0001-0.0003), representing only 4.9% of the total effect. AI was associated with a modest increase in biopsy number (a-path: 0.107, P < 0.001), and each additional biopsy was associated with higher detection rates (b-path: 0.480, P < 0.001).
Table 8 Mediation analysis evaluating biopsy number as a potential mediator between artificial intelligence assistance and high-risk gastric lesions detection.
By leveraging a large-sample retrospective dataset and employing PSM, this study investigated the effectiveness of AI in assisting endoscopists to improve HrGLs detection. AI assistance significantly enhanced the detection of HrGLs, including LGIN, HGLN, and EGC. Notably, AI assistance provided greater benefits to senior endoscopists and yielded more pronounced improvements under anesthesia-assisted conditions. These findings highlight the potential value of integrating AI into routine clinical practice.
Gastric cancer represents a critical public health challenge, both globally and in China, where early diagnosis and treatment are essential to reduce mortality and improve the quality of life of patients. Endoscopy is the primary method for detecting HrGLs, and its quality directly influences the detection rate. However, variability in medical resources and endoscopist expertise across regions leads to inconsistent endoscopic quality, contributing to missed diagnoses of high-risk lesions or misinterpretation of suspicious findings. While advanced endoscopic techniques such as NBI and magnifying chromoendoscopy have enhanced detection rates, their widespread adoption remains limited owing to high costs and the extensive training they necessitate for endoscopists. Currently, WLE is the most commonly used method. Nevertheless, HrGLs typically present with subtle features, such as mild mucosal erythema, that are easily missed by less experienced endoscopists. Estimates indicate that 5%-10% of patients subsequently diagnosed with gastric cancer had undergone endoscopy yielding negative results within the prior three years, highlighting a considerable incidence of missed EGC[23].
Improving the quality of endoscopic examinations is crucial to reduce missed diagnoses of EGC. AI, with its substantial potential in image recognition, offers a novel approach for endoscopic diagnosis of HrGLs. The 2023 Wuhan Consensus recommends that computer-aided examinations are essential for quality control during upper gastrointestinal endoscopy[24]. Xu et al[25] conducted a retrospective analysis of endoscopic images from five Chinese hospitals and found that a computer-aided detection system achieved high accuracy (86.4%-90.8%) in diagnosing gastric precancerous lesions, comparable to expert endoscopists, and significantly superior to non-experts. According to Ikenoyama et al[26], a CNN trained on 13584 endoscopic images representing 2639 gastric lesions required an average diagnosis time of 45.5 ± 1.8 seconds for analyzing 2940 test images, whereas endoscopists required 173.0 ± 66.0 minutes. CNNs exhibit a sensitivity of 58.4%, specificity of 87.3%, positive predictive value of 26.0%, and negative predictive value of 96.5%. Notably, their sensitivity was significantly higher than that of endoscopists, facilitating the identification of more EGC cases in less time[26]. In a large-scale prospective study, Wu et al[27] developed a deep learning system under white light, achieving sensitivity and specificity rates of 91.8% and 92.4%, respectively, for diagnosing gastric neoplastic lesions. A subsequent randomized controlled tandem trial based on this system further revealed that the AI-assisted group exhibited a significantly lower rate of missed diagnosis of neoplastic lesions than did the conventional group (6.1% vs 27.3%)[21]. However, most of these studies used high-quality endoscopic images, which may introduce certain limitations. In real-world clinical settings, image quality is often suboptimal because of factors such as mucosal cleanliness and patient cooperation. Therefore, this study aimed to evaluate the effectiveness of AI-assisted technology for improving the detection rate of HrGLs in an actual clinical environment.
The results demonstrated that within the PSM cohort, AI assistance significantly improved the detection rates of HrGLs, EGC, HGIN, and LGIN. Previous studies have reported that 60%-85% of patients with HGIN progress to malignancy within a median follow-up period of 4-48 months, imposing a significantly higher risk than that of LGIN[28]. Furthermore, owing to the challenges in histopathological differentiation between HGIN and EGC, especially in small biopsy samples, clinical guidelines explicitly recommend managing HGIN and EGC using the same prevention and treatment strategies[22]. Therefore, the ability of the AI system to enhance HGIN detection not only facilitates the early identification of high-risk lesions but also provides critical support for timely therapeutic intervention, thereby significantly improving patient prognosis and reducing the risk of disease progression.
In contrast, LGIN often presents with subtle endoscopic features and mild mucosal changes, which increase the diagnostic challenge for endoscopists. In clinical practice, detecting these minimal lesions requires intense concentration, which is difficult to maintain. This is particularly relevant in China, where endoscopists perform numerous daily procedures. Under sustained high workloads, visual fatigue and mental exhaustion are common, leading to decreased attention and potentially higher rates of overlooked lesions. AI serves as a tireless second observer, providing real-time assistance, especially when endoscopists experience fatigue, thereby helping to identify subtle and easily missed lesions. Moreover, the diagnostic capability of AI continues to improve with accumulating exposure, further enhancing its accuracy over time. Thus, AI serves an important role in real-time alerting of endoscopists to suspicious areas, which helps reduce missed diagnoses of LGIN. Although LGIN carries a relatively low risk of progression to cancer, improving its detection rate remains clinically significant. A meta-analysis indicated that 25% of LGIN cases were upgraded to HGIN after endoscopic submucosal dissection, and 7% progressed to EGC[29]. Previous studies have shown that 38%-75% of LGIN lesions regress spontaneously, whereas 19%-50% persist, and of these, 23% progress to malignancy within 10-48 months[30]. A United States follow-up study further revealed that patients with LGIN have a 25.6-fold higher risk of developing gastric cancer than that of the general population, with a median time to progression of 2.6 years[31]. Consequently, the American Society for Gastrointestinal Endoscopy guidelines recommend the endoscopic resection of neoplastic lesions. When LGIN is not resected, surveillance protocols must be more rigorous than those for non-neoplastic lesions to mitigate the risk of progression to advanced cancer. Accordingly, accurate diagnosis of LGIN is pivotal in determining both treatment and follow-up approaches. By improving the LGIN detection rate, AI-assisted endoscopy enables timely intervention, which may prevent long-term adverse outcomes.
The detection of HrGLs under WLE relies heavily on the comprehensive knowledge and extensive experience of the endoscopist. Compared to expert endoscopists, non-expert endoscopists exhibit a 14% reduction in the absolute detection rate of gastric neoplasia[17]. In subgroup analyses, AI assistance was associated with a significant increase in the detection rate of HrGLs by senior endoscopists. However, our confounding model showed no significant interaction between AI assistance and endoscopist’s (P = 0.77). This seemingly contradictory finding stems from their different purposes and differences in statistical efficacy between subgroups. The interaction test indicates that the effect of AI on detection rates did not differ statistically across experience levels. In other words, the direction and magnitude of the effect were similar among groups. In contrast, the subgroup analysis tests whether the effect is significant within each group. Only the senior group reached statistical significance, largely because it had the largest sample size (n = 7516) and thus greater statistical power. The intermediate (n = 6260) and junior (n = 1752) groups had smaller samples and failed to reach the level of statistical significance. Even though both of them had an outperformance ratio greater than 1.0, which showed a consistent trend of benefit. This lack of statistical significance due to limited sample size is a common challenge in subgroup analysis[32]. Importantly, data from all three subgroups showed a trend of benefit in exactly the same direction (ORs for detection of HrGLs: 2.23 for the senior group, 1.55 for the intermediate group, and 1.50 for the junior group). This consistency supports the conclusion of the interaction test that the adjunctive effect of AI is not inherently physician-specific. On the other hand, differences in self-awareness among endoscopists with varying seniority may also play a role. This observation may be partly explained by the Dunning-Kruger effect[33], a cognitive bias in which individuals with limited expertise often overestimate their performance. Less experienced endoscopists may lack insights into their own performance, leading to overconfidence and a focus on procedure quantity over quality[34]. Regarding the hypothesis that less experienced endoscopists might derive limited benefit due to cognitive biases or lower trust in AI prompts, our study design did not capture objective metrics such as AI alert adoption rates or biopsy compliance stratified by endoscopist experience. However, this remains a plausible explanation worthy of further investigation. From a clinical perspective, the effective integration of AI into medical practice depends on the clinician’s ability to interpret and trust its output[35]. Although the ENDOANGEL system is capable of real-time lesion identification and prompting, it does not provide explanatory information regarding alerts. Senior physicians, due to their greater clinical experience, may be able to integrate and validate AI cues more effectively, thus exhibiting a more significant statistical effect in the current study. However, for less experienced physicians, AI also shows a potential trend of increasing detection rates. Therefore, we think that the significant differences between subgroups mainly reflect limitations in sample size and statistical testing capabilities, rather than fundamental differences in the clinical utility of AI. The core conclusion is that the benefits of as a “second observer” may be universally applicable to physicians at all levels. Our study design did not capture objective metrics of AI-human interaction, such as alert adoption rates or time to decision stratified by experience, which limits our ability to elucidate these behavioral mechanisms. Future research should incorporate such metrics to better understand how to tailor AI implementation and training for endoscopists at different stages of expertise.
This study evaluated the anatomical distribution of HrGLs, with the highest detection rate observed in the gastric antrum (0.37%), which is logical as it being a high-risk zone for gastric pathology due to anatomical and physiological factors. The gastric angle poses a particular diagnostic challenge owing to its complex anatomy, including prominent mucosal folds and acute angulation, in which lesions are often subtle and easily overlooked during conventional endoscopy. While AI assistance markedly increased HrGLs detection in the gastric angle, no significant improvements were observed in the cardia, gastric fundus, body, antrum, or pylorus. This absence of effect may be explained by two factors: First, the antrum is a well-established high-risk region where endoscopists already maintain a high level of vigilance and diagnostic proficiency, thereby limiting the incremental value of AI. Second, the relatively low lesion prevalence in other gastric regions, coupled with a limited sample size, may have reduced the statistical power to detect modest but potentially meaningful differences.
While anatomical location influences lesion detectability, the practical considerations of how AI assistance is integrated into clinical workflow is equally important. In our real-world study setting, endoscopists maintained full clinical discretion regarding biopsy decisions when AI highlighted suspicious areas. The AI system served as a decision-support tool rather than a prescriptive directive, reflecting typical clinical practice where physician judgment remains paramount. This approach is evidenced by the modest increase in biopsy number associated with AI assistance (a-path = 0.107), suggesting selective rather than automatic adherence to AI alerts. While preserving clinical autonomy, this real-world design limited our ability to systematically document false-positive rates or analyze the specific criteria endoscopists used to accept or reject AI suggestions. The absence of mandatory biopsy protocols for AI-highlighted areas may have reduced unnecessary procedures but precluded comprehensive characterization of the system's false-positive profile. Future studies should implement more detailed documentation of AI-human interactions, including recording of all AI alerts and corresponding clinical responses, to better understand decision-making patterns and optimize AI systems for seamless clinical workflow integration.
The detection rate of HrGLs is highly dependent on the quality of the endoscopic examination, which, in turn, is closely associated with patient cooperation. The 2019 Asian Consensus Recommendations advocate sedation to enhance the detection rates of superficial tumors, suggesting possible benefits from prolonged observation and detailed examination[36]. Prior research has demonstrated that sedation prolongs examination time, facilitates more comprehensive mucosal inspection, and increases biopsy sampling across different gastric regions[37]. Sedation has been shown to improve various aspects of endoscopic examination, including patient tolerance, mucosal exposure time, and visualization of specific anatomical regions such as the esophagogastric junction[38]. In our studies, in sedated patients, AI assistance significantly increased the detection rates of HrGLs, LGIN, and EGC. This may be attributed to improved patient stability and cooperation during sedation, which allowed for higher-quality image acquisition and more effective AI-based analyses. Adequate gastric distension is particularly important for visualizing areas such as the greater curvature, where prominent folds require sufficient air insufflation for full exposure and evaluation[39]. A well-distended gastric mucosa provides clearer images, enhancing the accuracy of AI interpretation. Although our retrospective design did not include objective metrics of image quality (e.g., clarity scores, distension ratings), the established benefits of sedation on endoscopic procedural quality provide a plausible mechanistic basis for our findings. In contrast, unsedated patients may experience discomfort during air insufflation, resulting in reduced cooperation, suboptimal mucosal visualization, and impaired AI performance. Future prospective studies should incorporate standardized image quality assessments to directly evaluate how sedation-mediated improvements in visualization contribute to AI performance.
Performing adequate biopsies on suspicious lesions during endoscopy is essential for achieving accurate pathological diagnoses[40]. Consistent with previous research[41], our study showed that a higher number of biopsies was associated with increased detection rates of HrGLs. Our mediation analysis provides mechanistic insights into how AI improves HrGLs detection. Contrary to the hypothesis that AI might work primarily by prompting additional biopsies, we found that only 4.9% of its total effect was mediated through increased biopsy number. The vast majority (95.1%) represented a direct effect, likely attributable to enhanced visualization and more accurate targeting of suspicious areas. This finding is consistent with our stratified analysis showing the strongest AI benefit in cases with only one biopsy (OR = 5.53), where increased biopsy sampling cannot explain the improved detection. This suggests that AI may enable more precise lesion localization, allowing for targeted biopsies that could reduce unnecessary procedures, minimize patient burden, and enhance diagnostic accuracy. While AI was associated with a modest increase in biopsy count (a-path = 0.107), this likely reflects a rational clinical response to AI-highlighted areas rather than indiscriminate sampling.
Endoscopy is crucial for detecting lesions and enabling timely treatment. While advanced gastric cancer carries high treatment costs and a five-year survival rate of less than 20%, early detection can improve the five-year survival rate to over 90%. AI systems support timely lesion detection without increasing patient costs, allowing earlier intervention and better outcomes. Our study found that AI assistance significantly increased the detection rate of EGC (0.40% vs 0.15%). Based on an annual estimate of 40000 EGD procedures at our hospital, AI assistance could enable the detection of 100 additional EGCs per year. According to the World Health Organization cost-effectiveness criteria and the 2021 per capita GDP (11382 dollars), curing early-stage disease could provide 10 additional healthy life-years, translating to an averted economic loss of about 11.38 million dollars for 100 patients. Considering treatment costs and hidden family expenses of 250000 yuan per case, the total potential savings for 100 patients would amount to 105.97 million yuan. Meanwhile, the purchase and maintenance cost of a single endoscopic AI system over a lifespan of 10 years is only 1.2 million yuan, highlighting its significant economic and social benefits. Although the risk of LGIN progression to cancer is low, there is still a 0%-23% chance of malignant transformation within 10-48 months[30]. AI assistance significantly improved the detection rate of LGINs, which are often easily overlooked. Thus, the integration of AI not only provides patients with high-quality endoscopic examinations and detects more subtle HrGLs but also leads to substantial cost savings for families and society.
This observational and retrospective study, which involved a large sample size, controlled for confounding factors using PSM, a rigorous approach that approximates the effects of a randomized trial. The PSM-matched cohort was well balanced, avoiding biases arising from differences in incidence rates between the cohorts. Despite these advantages, this study has some limitations. First, being conducted at a single center, it is subject to inherent selection biases. Therefore, further large-scale multicenter studies are required to validate our findings. Second, its retrospective design limited the availability of some key variables, including smoking and alcohol history, dietary and exercise habits, family history of gastric malignancies, proton pump inhibitor usage, and Helicobacter pylori infection status, which may have introduced unmeasured bias. Future prospective randomized controlled trials are warranted to address these limitations. Third, complications from EGD examinations are part of the quality control standard for endoscopy. This study did not record any adverse events during the examinations, which resulted in a lack of essential safety data for AI-assisted endoscopy. However, as an enhancement technology, AI-assisted endoscopy mainly improves inspection quality and accuracy through more precise observations and auxiliary diagnoses. Because AI systems do not directly interact with patients’ bodies, it is hypothesized that their safety is comparable to that of traditional endoscopy. Previous prospective randomized controlled trials indicated that AI does not increase the incidence of adverse events, including perforation, bleeding, and infection[42]. Fourth, our study included only one year of endoscopic data from our center, without subsequent follow-up. A meta-analysis reported that the missed diagnosis rate of upper gastrointestinal cancers during a single endoscopic examination was 11.3%[43]. Future research should extend the observation period and compare missed cancer diagnosis rates between AI-assisted and non-AI-assisted scenarios. Fifth, our study lacked information regarding the use of other endoscopic imaging technologies, potentially limiting the interpretation of the results. Additionally, while our database recorded insertion and extraction times, these intervals also encompassed biopsy and esophageal examinations, making it impossible to determine the actual time spent observing the stomach. Because inspection time is closely linked to detection rates, future studies should use larger cohorts and improved recording methods to obtain more reliable measures. Finally, interpreting AI decisions remains challenging. Although AI performs exceptionally well in diagnosis, its ability to provide explanations for diagnoses significantly enhances the trust and acceptance of healthcare providers. Therefore, interpretable AI technologies should address these limitations. Sixth, the retrospective real-world design of our study imposed limitations on capturing two interconnected dimensions crucial for understanding AI-assisted endoscopy: Objective assessments of endoscopic image quality and detailed patterns of human-AI interaction. We lacked systematic metrics for image clarity, luminal distension, or motion artifacts. We also lacked the documented AI alerts and corresponding endoscopist responses hindered our ability to discern whether biopsies were driven by AI prompts or operator initiative. Future prospective studies should incorporate standardized image quality assessments alongside systematic recording of AI-generated alerts and clinician actions to elucidate how procedural conditions and interactive factors jointly shape AI’s diagnostic performance.
CONCLUSION
Our study demonstrates that AI assistance during EGD significantly enhances the detection of HrGLs, including HGIN, LGIN, and EGC. By improving recognition of subtle mucosal changes, AI has the potential to reduce missed diagnoses and facilitate earlier intervention, ultimately contributing to improved patient outcomes. Importantly, the benefits of AI were most pronounced when applied alongside experienced endoscopists and during procedures performed under anesthesia, underscoring its role as a supportive rather than replacement tool. These findings suggest that integrating AI into routine endoscopic practice may enhance diagnostic quality and standardize performance across different clinical settings.
Young E, Philpott H, Singh R. Endoscopic diagnosis and treatment of gastric dysplasia and early cancer: Current evidence and what the future may hold.World J Gastroenterol. 2021;27:5126-5151.
[PubMed] [DOI] [Full Text]
Zhang Z, Gao N, Liu K, Ni M, Zhang X, Yan P, Chen M, Dou X, Guo H, Yang T, Ding X, Xu G, Tang D, Wang L, Zou X. Risk factors of missed early gastric cancer in endoscopic resected population: a retrospective, case-control study.Surg Endosc. 2024;38:4380-4389.
[RCA] [PubMed] [DOI] [Full Text][Cited by in RCA: 6][Reference Citation Analysis (0)]
Hernanz N, Rodríguez de Santiago E, Marcos Prieto HM, Jorge Turrión MÁ, Barreiro Alonso E, Rodríguez Escaja C, Jiménez Jurado A, Sierra M, Pérez Valle I, Volpato N, García Prada M, Nuñez-Gómez L, Ríos-León R, Castaño García A, García de Paredes AG, Aicart M, Caminoa A, Peñas Parcía B, Vázquez-Sequeiros E, Albillos A. Characteristics and consequences of missed gastric cancer: A multicentric cohort study.Dig Liver Dis. 2019;51:894-900.
[RCA] [PubMed] [DOI] [Full Text][Cited by in Crossref: 15][Cited by in RCA: 29][Article Influence: 4.1][Reference Citation Analysis (0)]
Hosokawa O, Hattori M, Douden K, Hayashi H, Ohta K, Kaizaki Y. Difference in accuracy between gastroscopy and colonoscopy for detection of cancer.Hepatogastroenterology. 2007;54:442-444.
[PubMed] [DOI]
Liu L, Dong Z, Cheng J, Bu X, Qiu K, Yang C, Wang J, Niu W, Wu X, Xu J, Mao T, Lu L, Wan X, Zhou H. Diagnosis and segmentation effect of the ME-NBI-based deep learning model on gastric neoplasms in patients with suspected superficial lesions - a multicenter study.Front Oncol. 2022;12:1075578.
[RCA] [PubMed] [DOI] [Full Text][Cited by in RCA: 12][Reference Citation Analysis (0)]
Wu L, Shang R, Sharma P, Zhou W, Liu J, Yao L, Dong Z, Yuan J, Zeng Z, Yu Y, He C, Xiong Q, Li Y, Deng Y, Cao Z, Huang C, Zhou R, Li H, Hu G, Chen Y, Wang Y, He X, Zhu Y, Yu H. Effect of a deep learning-based system on the miss rate of gastric neoplasms during upper gastrointestinal endoscopy: a single-centre, tandem, randomised controlled trial.Lancet Gastroenterol Hepatol. 2021;6:700-708.
[RCA] [PubMed] [DOI] [Full Text][Cited by in Crossref: 120][Cited by in RCA: 104][Article Influence: 20.8][Reference Citation Analysis (0)]
National Clinical Research Center for Digestive System Diseases (Shanghai), National Alliance for Early Cancer Prevention and Treatment of Digestive Tract (GECA); Chinese Society of Gastroenterology Helicobacter Pylori Study Group; Chinese Society of Health Management; Chinese Society of Endoscopic Physicians Digestive Endoscopy Professional Committee; Chinese Cancer Society Tumor Endoscopy Professional Committee. [Chinese consensus on management of gastric mucosal precancerous conditions and lesions (2020)].Zhonghua Xiaohua Neijing Zazhi. 2020;37:769-780.
[PubMed] [DOI] [Full Text]
Big Data Collaboration Group of the Digestive Endoscopy Branch of the Chinese Medical Association. [Expert consensus on the clinical application of artificial intelligence system to upper gastrointestinal endoscopy (2023, Wuhan)].Zhonghua Xiaohua Neijing Zazhi. 2024;41:85-96.
[PubMed] [DOI] [Full Text]
Banks M, Graham D, Jansen M, Gotoda T, Coda S, di Pietro M, Uedo N, Bhandari P, Pritchard DM, Kuipers EJ, Rodriguez-Justo M, Novelli MR, Ragunath K, Shepherd N, Dinis-Ribeiro M. British Society of Gastroenterology guidelines on the diagnosis and management of patients at risk of gastric adenocarcinoma.Gut. 2019;68:1545-1575.
[RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)][Cited by in Crossref: 524][Cited by in RCA: 471][Article Influence: 67.3][Reference Citation Analysis (11)]
Luo X, Yao K, Lin X, Lin B, Zhu C, Huang S, Chen Z, Li A, Wang J, Huang Y, Li Z, Liu S, Han Z. Intensive Systematic "Train-the-Trainer" Course as an Effective Strategy to Improve Detection of Early Gastric Cancer: A Multicenter Retrospective Study.J Gastrointest Surg. 2023;27:1303-1312.
[RCA] [PubMed] [DOI] [Full Text][Cited by in RCA: 3][Reference Citation Analysis (0)]
Kruger J, Dunning D. Unskilled and unaware of it: how difficulties in recognizing one's own incompetence lead to inflated self-assessments.J Pers Soc Psychol. 1999;77:1121-1134.
[PubMed] [DOI] [Full Text]
Jin EH, Lee D, Bae JH, Kang HY, Kwak MS, Seo JY, Yang JI, Yang SY, Lim SH, Yim JY, Lim JH, Chung GE, Chung SJ, Choi JM, Han YM, Kang SJ, Lee J, Chan Kim H, Kim JS. Improved Accuracy in Optical Diagnosis of Colorectal Polyps Using Convolutional Neural Networks with Visual Explanations.Gastroenterology. 2020;158:2169-2179.e8.
[RCA] [PubMed] [DOI] [Full Text][Cited by in Crossref: 132][Cited by in RCA: 102][Article Influence: 17.0][Reference Citation Analysis (5)]
Beg S, Ragunath K, Wyman A, Banks M, Trudgill N, Pritchard DM, Riley S, Anderson J, Griffiths H, Bhandari P, Kaye P, Veitch A. Quality standards in upper gastrointestinal endoscopy: a position statement of the British Society of Gastroenterology (BSG) and Association of Upper Gastrointestinal Surgeons of Great Britain and Ireland (AUGIS).Gut. 2017;66:1886-1899.
[RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)][Cited by in Crossref: 283][Cited by in RCA: 249][Article Influence: 27.7][Reference Citation Analysis (0)]
Li SW, Zhang LH, Cai Y, Zhou XB, Fu XY, Song YQ, Xu SW, Tang SP, Luo RQ, Huang Q, Yan LL, He SQ, Zhang Y, Wang J, Ge SQ, Gu BB, Peng JB, Wang Y, Fang LN, Wu WD, Ye WG, Zhu M, Luo DH, Jin XX, Yang HD, Zhou JJ, Wang ZZ, Wu JF, Qin QQ, Lu YD, Wang F, Chen YH, Chen X, Xu SJ, Tung TH, Luo CW, Ye LP, Yu HG, Mao XL. Deep learning assists detection of esophageal cancer and precursor lesions in a prospective, randomized controlled study.Sci Transl Med. 2024;16:eadk5395.
[RCA] [PubMed] [DOI] [Full Text][Cited by in Crossref: 34][Cited by in RCA: 28][Article Influence: 14.0][Reference Citation Analysis (1)]
