Skip Navigation
Skip to contents

Clin Endosc : Clinical Endoscopy

Search
Advanced Search
OPEN ACCESS
mobile search button mobile menu button

Articles

Page Path
HOME > Clin Endosc > Volume 59(3); 2026 > Article
Original Article Impact of real-time artificial intelligence integration on detection of gastric lesions: an exploratory single-center before-and-after study using low-definition routine endoscopy
Hwijun Lee1,*orcid, Huynh Cong Bang2,*orcid, Seokho Cho1orcid, Jungmin Ha1orcid, Tran Thien Khiem2orcid, Le Viet Tung2orcid, La Vinh Phuc3orcid, Duong Trong Si4orcid, Tran The Du4orcid, Diem Thi-Ngoc Vo5orcid, Vo Nguyen Trung2,5orcid
Clinical Endoscopy 2026;59(3):408-416.
DOI: https://doi.org/10.5946/ce.2025.272
Published online: April 22, 2026

1MedInTech Inc., Seoul, Korea

2University Medical Center, Ho Chi Minh City, Ho Chi Minh City, Vietnam

3Can Tho University of Medicine and Pharmacy, Can Tho, Vietnam

4Can Tho Oncology Hospital, Can Tho, Vietnam

5University of Medicine and Pharmacy at Ho Chi Minh City, Ho Chi Minh City, Vietnam

Correspondence: Vo Nguyen Trung University of Medicine and Pharmacy at Ho Chi Minh City, 215 Hong Bang Street, Cho Lon Ward, Ho Chi Minh City 700000, Vietnam E-mail: trung.vn@umc.edu.vn
*Hwijun Lee and Huynh Cong Bang contributed equally to this work.
• Received: August 9, 2025   • Revised: January 1, 2026   • Accepted: January 2, 2026

© 2026 Korean Society of Gastrointestinal Endoscopy

This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (https://creativecommons.org/licenses/by-nc/4.0/) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.

  • 5,685 Views
  • 121 Download
  • 1 Web of Science
  • 1 Crossref
  • 1 Scopus
prev next
Full Article

Download PDF Download PDF

See letter "Expanding role of artificial intelligence in gastric neoplasms: moving beyond resource limitations" in Volume 59 on page 400.
  • Background/Aims
    Early detection of gastric neoplasia, particularly subcentimeter lesions, using upper gastrointestinal (GI) endoscopy remains challenging. This study evaluated the impact of a real-time artificial intelligence (AI) detection system on the lesion detection rate (LDR) during routine upper GI endoscopy performed using a low-definition platform commonly used in resource-limited settings, with a focus on lesions ≤0.5 cm.
  • Methods
    Diagnostic upper GI endoscopies performed between September 2024 and May 2025 were analyzed. LDRs were compared between the pre- and post-AI periods, including subgroup analyses by lesion size and type.
  • Results
    A total of 2,329 patients were included (1,491 pre-AI, 838 post-AI). After AI implementation, overall LDR per person increased from 1.15±0.45 to 1.20±0.57 (p<0.05). Detection of lesions ≤0.5 cm increased from 18.0% to 19.8% (p<0.05), while detection of larger lesions remained unchanged. The biopsy rate decreased from 13.8% to 8.5% (p<0.05).
  • Conclusions
    Real-time AI modestly improved the detection of diminutive gastric lesions while reducing unnecessary biopsies without compromising malignancy detection, thereby supporting its utility in routine endoscopy under resource-limited conditions.
  • Keywords: Diagnosis, computer-assisted; Endoscopy, digestive system; Gastrointestinal; Stomach neoplasms
Small lesions in the upper gastrointestinal (GI) tract present a critical diagnostic challenge during endoscopic screening and are frequently associated with early-stage gastric neoplasia or precancerous changes. Despite advances in diagnostic imaging and therapeutic endoscopy, the prognosis of gastric cancer remains closely related to the stage of diagnosis.1,2 When identified at an early stage, gastric neoplasms can often be treated with minimally invasive procedures, such as endoscopic submucosal dissection (ESD),3-9 which provides curative results with organ preservation. However, once the disease progresses beyond the mucosal layer, treatment becomes more invasive and the outcomes worsen significantly. Thus, improving early detection of gastric cancer is critical for optimizing patient survival and quality of life.
Endoscopic screening is an effective tool for detecting early gastric neoplasia, particularly in regions with a high incidence.10,11 However, it is inherently subject to human limitations such as fatigue, inattentional blindness, and variability in expertise.12 Visual recognition of early gastric cancer remains challenging, especially when the lesions are small, flat, pale, or partially obscured by mucus or gastric folds.13,14 These subtle mucosal changes can easily be overlooked even by experienced endoscopists, leading to delayed diagnosis or missed opportunities for curative intervention.15,16 In particular, subcentimeter lesions (≤0.5 cm in diameter) often lack obvious morphological changes and may not raise clinical suspicion during routine examination.13,14 Failure to detect such lesions during primary screening can result in interval cancers, thus undermining the effectiveness of surveillance programs.15,16
Recent advances in computer vision and deep learning have led to the development of real-time artificial intelligence (AI)-based systems that assist endoscopists in automatically detecting and highlighting areas of potential abnormality during live procedures.17,18 AI systems, typically based on convolutional neural networks, are trained on large datasets of annotated endoscopic images to recognize visual patterns associated with neoplastic changes.17 When deployed in clinical practice, they are capable of providing frame-by-frame analysis and generating visual or auditory alerts to prompt further inspections of suspicious regions.18,19 Such systems aim to reduce variability among observers, minimize blind spots, and maintain consistent vigilance throughout the examination.20,21
Several clinical trials and pilot studies have reported that the integration of AI into colonoscopy can significantly increase the adenoma detection rate, which has been adopted as a benchmark for lower GI endoscopy.22,23 Although several real-time studies from Korea and China have reported encouraging results, these countries have a high prevalence of gastric cancer and widespread use of high-definition endoscopy systems.24-26 However, data from low- and middle-income countries (LMICs), where low-definition endoscopic systems remain predominant and may limit the diagnostic performance of AI, are lacking. This represents a critical gap in the understanding of the real-world effectiveness of AI-assisted endoscopy under routine resource-limited conditions.
Therefore, this study aimed to evaluate the effect of implementing a real-time AI-assisted detection system on the lesion detection rate (LDR) during upper GI endoscopy performed using a low-definition endoscopy system, commonly employed in LMICs in Vietnam. The secondary objectives included assessment of this system’s performance in detecting small lesions measuring ≤5 mm and comparing the malignant LDR between the pre-AI and post-AI periods.
This exploratory single-center before-and-after study was conducted at the University Medical Centre, Ho Chi Minh City (UMP HCMC), a high-volume tertiary referral hospital in Vietnam. This study aimed to evaluate the effects of integrating an AI-assisted endoscopic examination system with lesion detection outcomes during upper GI screening. All diagnostic upper endoscopies were performed by a single board-certified gastroenterologist with >10 years of clinical experience and 50,000 cases of upper endoscopic diagnosis, thereby ensuring procedural consistency throughout the study period.
We reviewed the electronic medical records and endoscopy reports of all patients who underwent diagnostic upper endoscopy at the UMP HCMC between September 1, 2024, and May 31, 2025. The study period was divided into two phases for comparison: a pre-AI phase (September 1, 2024 to March 3, 2025), during which endoscopies were performed without AI support; and a post-AI phase (March 4, 2025 to May 31, 2025), following the clinical deployment of a real-time AI detection system. March 4, 2025, marked the formal integration of the AI system into routine practice and served as the boundary between the two phases.
Only the first endoscopic examination of each patient was included in the analysis to ensure independent lesion detection for each individual. Patients were excluded if the procedure was incomplete or if the image quality was insufficient for a reliable assessment (e.g., due to poor insufflation or suboptimal gastric preparation).
For each included case, we extracted the following variables from the endoscopy records and pathology database: patient age, sex, date of the procedure, number of lesions, size of the largest lesion (cm), anatomical location, and histopathological diagnosis based on biopsy. The lesion size was defined as the maximum diameter recorded in the endoscopic or pathological reports. A lesion was defined as any visible abnormality (e.g., erosion, elevated nodule, or flat depression) that led to a targeted biopsy. To facilitate size-stratified analysis, the lesions were grouped into four categories by size. The morphology of gastric lesions was classified according to the Paris classification system.
The primary outcome was LDR, defined as the proportion of patients in whom at least one lesion was identified during the procedure. LDRs were calculated separately for the pre- and post-AI periods to assess the impact of AI implementation on detection performance.
Secondary outcomes included (1) the detection rate of small lesions (≤5 mm), defined as the proportion of patients in whom at least one lesion of ≤5 mm in diameter was identified; and (2) the malignant detection rate (MDR), defined as the proportion of patients with at least one lesion pathologically diagnosed as low-grade dysplasia, high-grade dysplasia, early gastric cancer, or advanced gastric cancer. Both secondary outcomes were analyzed separately for the pre- and post-AI phases.
In addition, we calculated the benign detection rate (BDR), defined as the proportion of patients with at least one benign gastric lesion in the corresponding size category. Benign lesions included intestinal metaplasia, chronic gastritis, fundic gland polyps, hyperplastic polyps, atrophic gastritis, and inflammatory polyps.
All training and internal‐test examinations were performed using GIF‐Q150 endoscopes (Olympus Medical Systems Corp.) in conjunction with an endoscopic video imaging system, Evis Exera CV‐180 (Olympus Medical Systems Corp.).
The AI system used in this study was the MD-GA-300, a commercially available real-time computer-aided detection (CADe) software developed by MedInTech Inc., to support upper GI endoscopy. This system was designed to assist endoscopists in automatically identifying and localizing suspicious mucosal lesions within a live endoscopic video stream. Representative examples of the detection images are shown in Figure 1. The MD-GA-300 was developed as a real-time object detection system based on a transformer architecture optimized for endoscopic image analysis. The system was operated on a workstation equipped with an Intel Core i5-12400 CPU (2.50 GHz), an NVIDIA RTX 4060 graphics processing unit (GPU), and 16 GB of RAM running Windows 10 (64-bit). The video input was processed using a capture board supporting 1080p resolution, and 1 GB of available disk space was required for software installation and temporary storage. The model was implemented using PyTorch with CUDA and TensorRT acceleration, achieving processing speeds exceeding 30 fps at full HD resolution. The AI system continuously analyzes live endoscopic video streams and highlights suspected lesions in real-time, providing visual feedback to endoscopists without perceptible latency.
The underlying model was trained using more than 10,000 annotated endoscopic images obtained from the Seoul National University Hospital, covering a broad spectrum of gastric pathologies, including early and advanced gastric cancer, dysplasia, benign ulcers, and inflammatory conditions. To improve robustness and reduce false positives, non-lesion frames (e.g., bubbles, glare, or mucus) were included, and data augmentation techniques were applied to enhance generalizability. Internal validation showed a high diagnostic performance, with both sensitivity and specificity exceeding 90% for the detection of abnormal lesions, including small and flat lesions.
For clinical integration, the MD-GA-300 system was operated using a local GPU workstation connected to a live endoscopy feed. Visual prompts, including bounding boxes and confidence scores, were overlaid in real time on the main endoscopic monitor with minimal latency, enabling the endoscopist to receive AI assistance without disrupting the examination. The system functions passively without storing patient data or interfering with manual control. All final decisions regarding further inspection or biopsy of highlighted areas were made at the discretion of the endoscopist.
Continuous variables were first assessed for normality using the Shapiro–Wilk test. Variables that followed a normal distribution were compared between groups using Welch’s t-test (unequal variance t-test), whereas non-normally distributed variables were compared using the Wilcoxon rank-sum test. Categorical variables were expressed as frequencies and percentages and compared using the chi-square test or Fisher’s exact test, as appropriate. Statistical significance was defined as a two-sided p-value of <0.05. All statistical analyses were performed using Python ver. 3.10 software.
Ethical statements
This study was approved by the Ethics Committee for Biomedical Research of the University of Medicine and Pharmacy at Ho Chi Minh City under Decision (No. 1605/DHYD-HDDD). All the procedures were performed in accordance with the ethical standards of the Declaration of Helsinki and its amendments. Written informed consent was obtained from all participants prior to the endoscopic procedure.
The primary outcome measures are presented in Table 1. A total of 2,329 upper endoscopic examinations were included in the analysis, comprising 1,491 procedures performed in the pre-AI period and 838 procedures in the post-AI period. Baseline characteristics were comparable between the groups. The mean age was 45.34±13.79 years in the pre-AI group and 46.31±13.62 years in the post-AI group (p=0.10). The sex distribution was similar in the pre-AI and post-AI groups (female, 802 vs. 474; male, 689 vs. 364; both p=0.22), indicating no significant demographic imbalance between the groups.
Following AI implementation, the overall LDR per person increased from 1.15±0.45 lesions before AI to 1.20±0.57 lesions after AI. This difference was statistically significant according to the Wilcoxon rank-sum test (p=0.035), indicating that the application of AI modestly improved the number of lesions detected per patient during endoscopy. The detection rate for lesions ≤0.5 cm increased from 0.18 in the pre-AI group to 0.20 in the post-AI group (p<0.05), representing a statistically significant improvement despite the modest absolute change. In contrast, detection rates for lesions >0.5–1.0 cm (0.21 vs. 0.18, p=0.11), >1.0–1.5 cm (0.006 vs. 0.007, p=0.73), and >1.5 cm (0.051 vs. 0.043, p=0.51) did not differ significantly between the groups.
The proportion of examinations in which a biopsy was performed decreased significantly after AI deployment from 16.2% to 9.7% (p<0.05). Analysis of the biopsy results demonstrated a significant decline in BDR from 12.42% in the pre-AI period to 5.8% in the post-AI period (p<0.05), whereas MDR remained unchanged (1.5% vs. 1.0%, p=0.32).
The distribution of lesion locations showed no significant differences between periods (p=0.16). Lesions were most frequently identified in the antrum (71.9% vs. 67.2%), followed by the corpus (17.4% vs. 19.3%) and fundus (10.7% vs. 13.4%).
The distribution of the Paris classification did not differ significantly between the pre- and post-AI groups (p=0.38, chi-squared test). Type 0-II lesions were the most common morphology, accounting for 39.7% (n=672) of lesions before AI and 41.5% (n=416) after AI. Type 0-I, 0-III mixed, and other types were less frequent (4.4% vs. 3.8%, 2.2% vs. 1.5%, and 0.5% vs. 0.3%, respectively).
Overall, the implementation of real-time AI assistance modestly improved the overall LDR, primarily driven by increased detection of small lesions (≤5 mm). MDR remained stable between the two periods, whereas the frequency of benign biopsies significantly decreased following AI integration, indicating a potential reduction in unnecessary biopsy procedures.
Numerous studies have evaluated the performance of AI-assisted endoscopy in the upper GI tract; however, most have been retrospective or image-based rather than applied in real-time clinical settings.17,18,24,25 However, real-world implementation studies are limited.24-26 To the best of our knowledge, our study is the first to evaluate the effectiveness of AI-assisted endoscopy for detecting gastric lesions in real-world clinical practice in Vietnam. The investigation was conducted using older generation endoscopy systems that were not specifically designed for early gastric cancer diagnosis, including the Olympus Medical Systems Corp. CV-180 processor and GIF-Q150 endoscope. These models are among the most commonly used in Vietnam and other LMICs. This feature distinguishes our study from previous real-world investigations conducted in China and South Korea, in which more advanced high-definition systems with enhanced diagnostic capabilities, such as the Olympus Medical Systems Corp. H260 and H290 series, were predominantly utilized.
In this retrospective study, we observed that real-time AI assistance during upper GI endoscopy enhanced the overall detection of gastric lesions, with a particularly notable improvement in identifying small mucosal lesions measuring ≤0.5 cm. This selective enhancement suggests that AI can effectively complement endoscopists’ visual limitations, particularly in the detection of subtle or inconspicuous lesions that are often missed during routine examinations. Although several prospective trials have demonstrated that AI systems improve the overall detection of early gastric neoplasia, few have provided size-stratified analyses or specifically evaluated the detection of subcentimeter lesions.24-26 By contrast, our study offers real-world evidence that AI-assisted endoscopy significantly improves lesion detection in this clinically challenging and diagnostically vulnerable subgroup of patients with early gastric neoplasia.
Notably, the observed improvement in lesion detection was achieved despite the use of an older, low-definition endoscopy platform (Olympus CV-180 processor with a GIF-Q150 scope). This finding is particularly remarkable, given that several studies have demonstrated a substantial decline in AI performance under non-ideal imaging conditions.27,28 Jaspers reported that endoscopic AI models trained on high-quality images from expert centers experienced a marked reduction in diagnostic accuracy when tested on lower-quality images affected by poor illumination, motion blurring, or image compression. In their robustness analysis, the performance of deep neural networks dropped by up to 10–30% in area under the curve when image quality was degraded, underscoring the vulnerability of AI systems to real-world visual variability.29 In contrast, our results showed that real-time AI-assistance retained measurable effectiveness, even on a low-definition endoscopic system, suggesting that the implemented AI algorithm demonstrated acceptable robustness against suboptimal image quality. This reinforces the potential clinical utility of AI in resource-limited environments where modern high-resolution endoscopy systems may not be widely available.
Our study also demonstrated the ability of AI-assisted endoscopy to enhance diagnostic performance even in regions with a relatively low incidence of gastric cancer. According to GLOBOCAN 2022, Vietnam’s age-standardized incidence rate (ASR) for gastric cancer is approximately 15.5–16.3 per 100,000 populations—higher than the global average (9.2 per 100,000) but considerably lower than the rates reported in East Asian countries such as Japan, Mongolia, and South Korea (ASR, 30–35 per 100,000).30 Within Vietnam, there is also a marked regional disparity in gastric cancer burden. Epidemiological data indicate that the incidence rate in Hanoi, the largest city in the north of Vietnam, was substantially higher than that in Ho Chi Minh City, the largest city in the south. Cancer registry data reported 377 new gastric cancer cases in Hanoi and 135 in Ho Chi Minh City in 2013. Projections for 2025 suggest a continued increase in Hanoi (up to approximately 410 cases), whereas the number in Ho Chi Minh City is expected to remain much lower, at approximately 117 cases.31 Therefore, our study, conducted in Ho Chi Minh City, provides one of the first real-world demonstrations that AI-assisted endoscopy can be effective in improving lesion detection, even in lower-incidence settings. This contrasts with most previous investigations conducted in China and South Korea, where gastric cancer incidence is much higher and screening programs are well established.24-26 Collectively, these findings highlight the potential clinical value of AI technology in enhancing diagnostic vigilance and early detection in regions with a lower disease prevalence and limited screening infrastructure.
Importantly, this improvement in subcentimeter lesion detection was accompanied by a marked reduction in biopsy frequency and a significant decrease in benign histologic yields while preserving the MDR. This finding is particularly relevant in resource-limited healthcare settings, such as in Vietnam, where reducing unnecessary biopsies can substantially lower the cost of gastric cancer screening. By decreasing the economic and procedural burdens associated with endoscopic screening, AI assistance may help improve physician adherence and patient participation in early detection programs.
In contrast, the detection rates for larger or more obvious lesions remained stable and even slightly decreased in some subgroups, implying that the incremental value of AI lies more in subtle, early-stage abnormalities than in advanced disease. These findings partially align with previous real-world data, as the effectiveness of AI in detecting gastric lesions remains unclear. In a two-center study from China, AI assistance produced divergent outcomes depending on the biopsy strategy: at a high-biopsy center, AI did not increase the overall neoplasm detection rate (1.39% vs. 1.36%, p=0.897) but significantly increased the proportion of early gastric cancers among detected neoplasms (25.4% vs. 43.0%; odds ratio [OR], 2.21).24 Conversely, at a low-biopsy center, the gastric neoplasm detection rate nearly doubled with AI (1.78% vs. 3.23%; OR, 1.84; p<0.001), mainly because of the increased detection of low-grade intraepithelial neoplasia, while the early cancer yield did not improve.24 More robust evidence from the Intelligent Quality-Control System randomized trial demonstrated a consistent increase in early upper GI neoplasm detection across both academic and non-academic centers (6.1% vs. 2.3%, p=0.0001), as well as for both junior and experienced endoscopists.25 In contrast, the Korean Clinical Decision Support System study showed only a non-significant trend towards improved lesion detection (2.0% vs. 1.3%, p=0.21), despite high accuracy in lesion classification and invasion depth prediction.26
Our study did not observe an increase in MDR following AI implementation. This may be attributable to the lack of follow-up data on posttreatment pathology, as our hospital has not yet implemented ESD. Patients with high-grade dysplasia or early gastric cancer are referred to tertiary centers where ESD is available, whereas those with low-grade dysplasia are typically followed up without immediate intervention. This strategy likely contributes to the absence of an observed AI-associated improvement in malignancy detection. Previous studies have consistently reported significant histopathological discrepancies between pre-procedural biopsy findings and post-ESD specimens, frequently resulting in pathological upgrading to higher-grade dysplasia or carcinoma.32,33 Concordance rates between biopsy and ESD pathology reportedly range from 41.9% to 66.1%, with upgrades observed in 26.4% to 40.7% of cases, particularly among lesions initially diagnosed as low- or high-grade intraepithelial neoplasia, up to 30% of which may reveal adenocarcinoma upon post-ESD evaluation.33-35 Consequently, incorporating post-intervention histology into future analyses could provide a more accurate assessment of the diagnostic performance of AI in identifying malignant or premalignant lesions that were initially underestimated.
Despite these encouraging findings, this study has some limitations. First, it employed a pre–post design, which inherently carries the risk of confounding due to temporal or procedural changes. However, the baseline characteristics of the two groups were comparable, suggesting that these effects were minimal. Second, this was a single-center pilot study involving only one expert endoscopist. Although this ensured procedural consistency and minimized inter-operator variability between the pre- and post-AI phases, it also limited the generalizability of our results. Based on these preliminary findings, future studies involving multiple centers, longer study periods, and broader participation of endoscopists are warranted to comprehensively evaluate the clinical impact of AI. Finally, a notable limitation of our study was the lack of follow-up data on posttreatment pathological outcomes. This was partly because ESD had not yet been implemented at our institution, precluding access to post-therapeutic histopathological confirmation. The absence of ESD is common across Vietnam, particularly in the southern regions, where the incidence of gastric cancer is relatively low. We plan to address this limitation in future studies by introducing ESD at our center, which will enable comprehensive follow-up and provide a more accurate evaluation of the impact of AI on the detection and management of early gastric cancer.
In this single-center before-and-after study using a low-definition endoscopy platform (CV-180/GIF-Q150), real-time AI assistance modestly but significantly improved overall LDR, driven chiefly by increased identification of subcentimeter (≤0.5 cm) gastric lesions. Simultaneously, biopsy frequency and benign histologic yields declined, whereas MDR remained stable, indicating more selective and efficient tissue sampling without compromising oncological vigilance. These results suggest that AI can help endoscopists overcome subtle visual blind spots in routine practice.
Given Vietnam’s heterogeneous gastric cancer burden and the widespread use of older endoscopy systems in resource-limited settings, our findings provide real-world evidence that AI can deliver clinically meaningful benefits even outside high-incidence regions and without high-end imaging hardware. Future prospective, multicenter studies with standardized follow-up and post-intervention pathology are warranted to validate the generalizability of our findings and to determine whether AI-assisted detection ultimately improves patient outcomes.

Supplementary Video 1.

Demonstration of AI‑assisted gastroscopy.
Supplementary materials related to this article can be found online at https://doi.org/ce.2025.272.
Fig. 1.
Example images of the detection results of MD-GA-300 (MedInTech Inc.). (A–F) Gastric lesions detected by endoscopy using AI software.
ce-2025-272f1.jpg
ce-2025-272f2.jpg
Table 1.
Comparison of baseline characteristics between the pre-AI and post-AI phases
Variable Pre-AI Post-AI p-value
Age (yr) 45.34±13.79 46.31±13.62 0.10j)
Sex
 Female 802 474 0.22k)
 Male 689 364 -
Lesion detected (by size)
 ≤0.5 cm 269 166
 >0.5 cm, ≤1.0 cm 316 153
 >1.0 cm, ≤1.5 cm 9 6
 >1.5 cm 76 36
 Overall LDRa) 1.15 1.20 0.035l)
LDR (by size)a)
 ≤0.5 cm 0.18 0.20 <0.05l)
 >0.5 cm, ≤1.0 cm 0.21 0.18 0.11l)
 >1.0 cm, ≤1.5 cm 0.006 0.007 0.73l)
 >1.5 cm 0.051 0.043 0.51l)
Biopsy performed (n, %) 242 (13.8) 81 (8.5) <0.05l)
Biopsy result
 Benign (BDR)b),c) 213 (0.124) 59 (0.058) <0.05l)
 Malignant (MDR)d),e) 26 (0.015) 10 (0.01) 0.32l)
Lesion location (n, %)
 Antrum 568 (72.1) 315 (67.2) 0.16k)
 Corpus 135 (17.1) 91 (19.4) -
 Fundus 85 (10.8) 63 (13.4) -
Paris classification (n, %)
 Type 0–If) 74 (4.4) 38 (3.8) 0.38k)
 Type 0–IIg) 672 (39.7) 416 (41.5) -
 Type 0–III mixedh) 38 (2.2) 15 (1.5) -
 Otheri) 8 (0.5) 3 (0.3) -

AI, artificial intelligence; LDR, lesion detection rate; BDR, benign detection rate; MDR, malignancy detection rate.

a)LDR defined as the proportion of patients with at least one lesion of the corresponding size.

b)Intestinal metaplasia, chronic gastritis, fundic gland polyp, hyperplastic polyp, atrophic gastritis, and inflammatory polyp.

c)BDR defined as the proportion of patients with at least one benign lesion of the corresponding size.

d)Includes early gastric cancer, advanced gastric cancer, high-grade dysplasia, and low-grade dysplasia.

e)MDR defined as the proportion of patients with at least one malignant lesion of the corresponding size.

f)Includes 0–I,0–Is, 0–Ip, 0–Isp.

g)Includes 0–IIa, 0–IIb, 0–IIc.

h)Includes 0–III, 0–IIa+0–III, 0–IIb+0–III, 0–IIc+0–III.

i)Includes I, II, II, IV.

j)Welch’s t-test.

k)Chi-square test.

l)Wilcoxon rank-sum test.

  • 1. Sano T, Coit DG, Kim HH, et al. Proposal of a new stage grouping of gastric cancer for TNM classification: International Gastric Cancer Association staging project. Gastric Cancer 2017;20:217–225.ArticlePubMedPDF
  • 2. Katai H, Ishikawa T, Akazawa K, et al. Five-year survival analysis of surgically resected gastric cancer cases in Japan: a retrospective analysis of more than 100,000 patients from the nationwide registry of the Japanese Gastric Cancer Association (2001-2007). Gastric Cancer 2018;21:144–154.ArticlePubMedPDF
  • 3. Tsubono Y, Hisamichi S. Screening for gastric cancer in Japan. Gastric Cancer 2000;3:9–18.ArticlePubMedPDF
  • 4. Yeoh KG. How do we improve outcomes for gastric cancer? J Gastroenterol Hepatol 2007;22:970–972.ArticlePubMed
  • 5. Jeon HK, Kim GH, Lee BE, et al. Long-term outcome of endoscopic submucosal dissection is comparable to that of surgery for early gastric cancer: a propensity-matched analysis. Gastric Cancer 2018;21:133–143.ArticlePubMedPDF
  • 6. Isomoto H, Shikuwa S, Yamaguchi N, et al. Endoscopic submucosal dissection for early gastric cancer: a large-scale feasibility study. Gut 2009;58:331–336.ArticlePubMed
  • 7. Choi MK, Kim GH, Park DY, et al. Long-term outcomes of endoscopic submucosal dissection for early gastric cancer: a single-center experience. Surg Endosc 2013;27:4250–4258.ArticlePubMedPDF
  • 8. Ahn JY, Jung HY. Long-term outcome of extended endoscopic submucosal dissection for early gastric cancer with differentiated histology. Clin Endosc 2013;46:463–466.ArticlePubMedPMC
  • 9. Gotoda T, Iwasaki M, Kusano C, et al. Endoscopic resection of early gastric cancer treated by guideline and expanded National Cancer Centre criteria. Br J Surg 2010;97:868–871.ArticlePubMed
  • 10. Zhang X, Li M, Chen S, et al. Endoscopic screening in asian countries is associated with reduced gastric cancer mortality: a meta-analysis and systematic review. Gastroenterology 2018;155:347–354.ArticlePubMed
  • 11. Hamashima C, Shibuya D, Yamazaki H, et al. The Japanese guidelines for gastric cancer screening. Jpn J Clin Oncol 2008;38:259–267.ArticlePubMed
  • 12. Aranda-Hernández J, Hwang J, Kandel G. Seeing better--Evidence based recommendations on optimizing colonoscopy adenoma detection rate. World J Gastroenterol 2016;22:1767–1778.ArticlePubMedPMC
  • 13. Yao K. The endoscopic diagnosis of early gastric cancer. Ann Gastroenterol 2013;26:11–22.PubMedPMC
  • 14. Sugano K, Sato K, Yao K. New diagnostic approaches for early detection of gastric cancer. Dig Dis 2004;22:327–333.ArticlePubMedPDF
  • 15. Pimenta-Melo AR, Monteiro-Soares M, Libânio D, et al. Missing rate for gastric cancer during upper gastrointestinal endoscopy: a systematic review and meta-analysis. Eur J Gastroenterol Hepatol 2016;28:1041–1049.ArticlePubMed
  • 16. Beck M, Bringeland EA, Qvigstad G, et al. Gastric cancers missed at upper endoscopy in central Norway 2007 to 2016: a population-based study. Cancers (Basel) 2021;13:5628.ArticlePubMedPMC
  • 17. Hirasawa T, Aoyama K, Tanimoto T, et al. Application of artificial intelligence using a convolutional neural network for detecting gastric cancer in endoscopic images. Gastric Cancer 2018;21:653–660.ArticlePubMedPDF
  • 18. Wu L, Zhou W, Wan X, et al. A deep neural network improves endoscopic detection of early gastric cancer without blind spots. Endoscopy 2019;51:522–531.ArticlePubMed
  • 19. Luo H, Xu G, Li C, et al. Real-time artificial intelligence for detection of upper gastrointestinal cancer by endoscopy: a multicentre, case-control, diagnostic study. Lancet Oncol 2019;20:1645–1654.ArticlePubMed
  • 20. Bou Jaoude J, Al Bacha R, Abboud B. Will artificial intelligence reach any limit in gastroenterology? Artif Intell Gastroenterol 2024;5:91336.
  • 21. Min JK, Kwak MS, Cha JM. Overview of Deep Learning in Gastrointestinal Endoscopy. Gut Liver 2019;13:388–393.ArticlePubMedPMC
  • 22. Barua I, Vinsard DG, Jodal HC, et al. Artificial intelligence for polyp detection during colonoscopy: a systematic review and meta-analysis. Endoscopy 2021;53:277–284.ArticlePubMed
  • 23. Repici A, Badalamenti M, Maselli R, et al. Efficacy of real-time computer-aided detection of colorectal neoplasia in a randomized trial. Gastroenterology 2020;159:512–520.ArticlePubMed
  • 24. Huang L, Xu M, Li Y, et al. Gastric neoplasm detection of computer-aided detection-assisted esophagogastroduodenoscopy changes with implement scenarios: a real-world study. J Gastroenterol Hepatol 2024;39:2787–2795.ArticlePubMedPDF
  • 25. Zhou R, Liu J, Zhang C, et al. Efficacy of a real-time Intelligent Quality-Control System for the detection of early upper gastrointestinal neoplasms: a multicentre, single-blinded, randomised controlled trial. EClinicalMedicine 2024;75:102803.ArticlePubMedPMC
  • 26. Gong EJ, Bang CS, Lee JJ, et al. Deep learning-based clinical decision support system for gastric neoplasms in real-time endoscopy: development and validation study. Endoscopy 2023;55:701–708.ArticlePubMed
  • 27. Zha B, Cai A, Wang G. Diagnostic accuracy of artificial intelligence in endoscopy: umbrella review. JMIR Med Inform 2024;12:e56361.ArticlePubMedPMC
  • 28. Jong MR, Jaspers TJM, Kusters CHJ, et al. Challenges in implementing endoscopic artificial intelligence: the impact of real-world imaging conditions on barrett's neoplasia detection. United European Gastroenterol J 2025;13:929–937.ArticlePubMedPMC
  • 29. Jaspers TJM, Boers TGW, Kusters CHJ, et al. Investigating the impact of image quality on endoscopic AI model performance. Lect Comput Sci 2024;32–41.Article
  • 30. Bray F, Laversanne M, Sung H, et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin 2024;74:229–263.ArticlePubMed
  • 31. Nguyen SM, Deppen S, Nguyen GH, et al. projecting cancer incidence for 2025 in the 2 largest populated cities in Vietnam. Cancer Control 2019;26:1073274819865274.ArticlePubMedPMCPDF
  • 32. Aliaga Ramos J, Pedrosa MS, Yoshida N, et al. Histopathologic diagnosis discrepancies between preoperative endoscopic forceps biopsies and specimens resected by endoscopic submucosal dissection in superficial gastric neoplasms. J Clin Gastroenterol 2023;57:74–81.ArticlePubMed
  • 33. Wang P, Zhao X, Wang R, et al. Risk factors for pathological upgrading and noncurative resection in patients with gastric mucosal lesions after endoscopic submucosal dissection. BMC Gastroenterol 2024;24:253.ArticlePubMedPMCPDF
  • 34. Noh CK, Jung MW, Shin SJ, et al. Analysis of endoscopic features for histologic discrepancies between biopsy and endoscopic submucosal dissection in gastric neoplasms: 10-year results. Dig Liver Dis 2019;51:79–85.ArticlePubMed
  • 35. Lu C, Lv X, Lin Y, et al. Retrospective study: The diagnostic accuracy of conventional forceps biopsy of gastric epithelial compared to endoscopic submucosal dissection (STROBE compliant). Medicine (Baltimore) 2016;95:e4353.ArticlePubMedPMC

Figure & Data

REFERENCES

    Citations

    Citations to this article as recorded by  
    • Expanding role of artificial intelligence in gastric neoplasms: moving beyond resource limitations
      Hannah Lee, Jun-Won Chung, Kyoung Oh Kim, Kwang An Kwon, Jung Ho Kim
      Clinical Endoscopy.2026; 59(3): 400.     CrossRef

    • Cite
      CITE
      export Copy Download
      Close
      Download Citation
      Download a citation file in RIS format that can be imported by all major citation management software, including EndNote, ProCite, RefWorks, and Reference Manager.
      Format:
      • RIS — For EndNote, ProCite, RefWorks, and most other reference management software
      • BibTeX — For JabRef, BibDesk, and other BibTeX-specific software
      Include:
      • Citation for the content below
      Impact of real-time artificial intelligence integration on detection of gastric lesions: an exploratory single-center before-and-after study using low-definition routine endoscopy
      Clin Endosc. 2026;59(3):408-416.   Published online April 22, 2026
      DOI: https://doi.org/10.5946/ce.2025.272
      Close
    • XML DownloadXML Download
    Figure
    • 0
    • 1
    Impact of real-time artificial intelligence integration on detection of gastric lesions: an exploratory single-center before-and-after study using low-definition routine endoscopy
    Image Image
    Fig. 1. Example images of the detection results of MD-GA-300 (MedInTech Inc.). (A–F) Gastric lesions detected by endoscopy using AI software.
    Graphical abstract

    Fig. 1.

    Graphical abstract

    Impact of real-time artificial intelligence integration on detection of gastric lesions: an exploratory single-center before-and-after study using low-definition routine endoscopy
    Variable Pre-AI Post-AI p-value
    Age (yr) 45.34±13.79 46.31±13.62 0.10j)
    Sex
     Female 802 474 0.22k)
     Male 689 364 -
    Lesion detected (by size)
     ≤0.5 cm 269 166
     >0.5 cm, ≤1.0 cm 316 153
     >1.0 cm, ≤1.5 cm 9 6
     >1.5 cm 76 36
     Overall LDRa) 1.15 1.20 0.035l)
    LDR (by size)a)
     ≤0.5 cm 0.18 0.20 <0.05l)
     >0.5 cm, ≤1.0 cm 0.21 0.18 0.11l)
     >1.0 cm, ≤1.5 cm 0.006 0.007 0.73l)
     >1.5 cm 0.051 0.043 0.51l)
    Biopsy performed (n, %) 242 (13.8) 81 (8.5) <0.05l)
    Biopsy result
     Benign (BDR)b),c) 213 (0.124) 59 (0.058) <0.05l)
     Malignant (MDR)d),e) 26 (0.015) 10 (0.01) 0.32l)
    Lesion location (n, %)
     Antrum 568 (72.1) 315 (67.2) 0.16k)
     Corpus 135 (17.1) 91 (19.4) -
     Fundus 85 (10.8) 63 (13.4) -
    Paris classification (n, %)
     Type 0–If) 74 (4.4) 38 (3.8) 0.38k)
     Type 0–IIg) 672 (39.7) 416 (41.5) -
     Type 0–III mixedh) 38 (2.2) 15 (1.5) -
     Otheri) 8 (0.5) 3 (0.3) -
    Table 1. Comparison of baseline characteristics between the pre-AI and post-AI phases

    AI, artificial intelligence; LDR, lesion detection rate; BDR, benign detection rate; MDR, malignancy detection rate.

    LDR defined as the proportion of patients with at least one lesion of the corresponding size.

    Intestinal metaplasia, chronic gastritis, fundic gland polyp, hyperplastic polyp, atrophic gastritis, and inflammatory polyp.

    BDR defined as the proportion of patients with at least one benign lesion of the corresponding size.

    Includes early gastric cancer, advanced gastric cancer, high-grade dysplasia, and low-grade dysplasia.

    MDR defined as the proportion of patients with at least one malignant lesion of the corresponding size.

    Includes 0–I,0–Is, 0–Ip, 0–Isp.

    Includes 0–IIa, 0–IIb, 0–IIc.

    Includes 0–III, 0–IIa+0–III, 0–IIb+0–III, 0–IIc+0–III.

    Includes I, II, II, IV.

    Welch’s t-test.

    Chi-square test.

    Wilcoxon rank-sum test.

    Table 1.

    Clin Endosc : Clinical Endoscopy Twitter Facebook
    © Korean Society of Gastrointestinal Endoscopy.
    Close layer
    TOP