Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice J Clin Med . 2026 Mar 29;15(7):2609. doi: 10.3390/jcm15072609 Search in PMC Search in PubMed View in NLM Catalog Add to search Real-World Evaluation of an AI-Assisted Diagnostic Support System for Early Gastric Cancer: Diagnostic Performance, Confidence Stratification, and Determinants of False-Positive Diagnosis Satoshi Osawa Satoshi Osawa 1 Department of Advanced Medical Science for Regional Collaboration, Hamamatsu University School of Medicine, Hamamatsu 431-3192, Japan 2 Department of Endoscopic and Photodynamic Medicine, Hamamatsu University School of Medicine, Hamamatsu 431-3192, Japan; [email protected] (T.Y.); Conceptualization, Methodology, Investigation, Writing – original draft, Writing – review & editing, Formal analysis Find articles by Satoshi Osawa 1, 2, *, † , Takanori Yamada Takanori Yamada 2 Department of Endoscopic and Photodynamic Medicine, Hamamatsu University School of Medicine, Hamamatsu 431-3192, Japan; [email protected] (T.Y.); Conceptualization, Methodology, Formal analysis, Investigation, Writing – review & editing Find articles by Takanori Yamada 2, † , Wataru Inui Wataru Inui 1 Department of Advanced Medical Science for Regional Collaboration, Hamamatsu University School of Medicine, Hamamatsu 431-3192, Japan Investigation Find articles by Wataru Inui 1 , Tomoyuki Niwa Tomoyuki Niwa 2 Department of Endoscopic and Photodynamic Medicine, Hamamatsu University School of Medicine, Hamamatsu 431-3192, Japan; [email protected] (T.Y.); Investigation Find articles by Tomoyuki Niwa 2 , Kenichi Takahashi Kenichi Takahashi 3 First Department of Medicine, Hamamatsu University School of Medicine, Hamamatsu 431-3192, Japan Find articles by Kenichi Takahashi 3 , Takatoshi Egami Takatoshi Egami 3 First Department of Medicine, Hamamatsu University School of Medicine, Hamamatsu 431-3192, Japan Investigation Find articles by Takatoshi Egami 3 , Keisuke Inagaki Keisuke Inagaki 3 First Department of Medicine, Hamamatsu University School of Medicine, Hamamatsu 431-3192, Japan Investigation Find articles by Keisuke Inagaki 3 , Tomohiro Takebe Tomohiro Takebe 3 First Department of Medicine, Hamamatsu University School of Medicine, Hamamatsu 431-3192, Japan Investigation Find articles by Tomohiro Takebe 3 , Tatsuhiro Ito Tatsuhiro Ito 3 First Department of Medicine, Hamamatsu University School of Medicine, Hamamatsu 431-3192, Japan Investigation Find articles by Tatsuhiro Ito 3 , Satoru Takahashi Satoru Takahashi 3 First Department of Medicine, Hamamatsu University School of Medicine, Hamamatsu 431-3192, Japan Investigation Find articles by Satoru Takahashi 3 , Shunya Onoue Shunya Onoue 3 First Department of Medicine, Hamamatsu University School of Medicine, Hamamatsu 431-3192, Japan Investigation Find articles by Shunya Onoue 3 , Yusuke Asai Yusuke Asai 3 First Department of Medicine, Hamamatsu University School of Medicine, Hamamatsu 431-3192, Japan Investigation Find articles by Yusuke Asai 3 , Kiichi Sugiura Kiichi Sugiura 3 First Department of Medicine, Hamamatsu University School of Medicine, Hamamatsu 431-3192, Japan Investigation Find articles by Kiichi Sugiura 3 , Tomoharu Matsuura Tomoharu Matsuura 4 Department of Laboratory Medicine, Hamamatsu University School of Medicine, Hamamatsu 431-3192, Japan Investigation Find articles by Tomoharu Matsuura 4 , Natsuki Ishida Natsuki Ishida 3 First Department of Medicine, Hamamatsu University School of Medicine, Hamamatsu 431-3192, Japan Investigation, Formal analysis, Writing – review & editing Find articles by Natsuki Ishida 3 , Mihoko Yamade Mihoko Yamade 3 First Department of Medicine, Hamamatsu University School of Medicine, Hamamatsu 431-3192, Japan Investigation, Writing – review & editing, Formal analysis Find articles by Mihoko Yamade 3 , Moriya Iwaizumi Moriya Iwaizumi 4 Department of Laboratory Medicine, Hamamatsu University School of Medicine, Hamamatsu 431-3192, Japan Writing – review & editing, Supervision Find articles by Moriya Iwaizumi 4 , Yasushi Hamaya Yasushi Hamaya 3 First Department of Medicine, Hamamatsu University School of Medicine, Hamamatsu 431-3192, Japan Writing – review & editing Find articles by Yasushi Hamaya 3 , Ken Sugimoto Ken Sugimoto 3 First Department of Medicine, Hamamatsu University School of Medicine, Hamamatsu 431-3192, Japan Supervision Find articles by Ken Sugimoto 3 Editors: Mattia Brigida , Gianluca Andrisani Author information Article notes Copyright and License information 1 Department of Advanced Medical Science for Regional Collaboration, Hamamatsu University School of Medicine, Hamamatsu 431-3192, Japan 2 Department of Endoscopic and Photodynamic Medicine, Hamamatsu University School of Medicine, Hamamatsu 431-3192, Japan; [email protected] (T.Y.); 3 First Department of Medicine, Hamamatsu University School of Medicine, Hamamatsu 431-3192, Japan 4 Department of Laboratory Medicine, Hamamatsu University School of Medicine, Hamamatsu 431-3192, Japan * Correspondence: [email protected] ; Tel.: +81-53-435-2261; Fax: +81-53-434-9447 † These authors contributed equally to this work. Roles Satoshi Osawa : Conceptualization, Methodology, Investigation, Writing – original draft, Writing – review & editing, Formal analysis Takanori Yamada : Conceptualization, Methodology, Formal analysis, Investigation, Writing – review & editing Wataru Inui : Investigation Tomoyuki Niwa : Investigation Takatoshi Egami : Investigation Keisuke Inagaki : Investigation Tomohiro Takebe : Investigation Tatsuhiro Ito : Investigation Satoru Takahashi : Investigation Shunya Onoue : Investigation Yusuke Asai : Investigation Kiichi Sugiura : Investigation Tomoharu Matsuura : Investigation Natsuki Ishida : Investigation, Formal analysis, Writing – review & editing Mihoko Yamade : Investigation, Writing – review & editing, Formal analysis Moriya Iwaizumi : Writing – review & editing, Supervision Yasushi Hamaya : Writing – review & editing Ken Sugimoto : Supervision Mattia Brigida : Academic Editor Gianluca Andrisani : Academic Editor Received 2026 Feb 25; Revised 2026 Mar 20; Accepted 2026 Mar 25; Collection date 2026 Apr. © 2026 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license . PMC Copyright notice PMCID: PMC13073999 PMID: 41976910 Abstract Background/Objectives : Artificial intelligence (AI)-assisted endoscopy has shown high sensitivity for early gastric cancer detection; however, false-positive diagnoses remain a clinical challenge. This study aimed to evaluate the real-world diagnostic performance of a commercially available AI system and to identify factors associated with false-positive diagnoses, focusing on repeated AI evaluations and confidence stratification. Methods : This single-center retrospective study included 47 patients with 89 localized gastric lesions evaluated between March 2024 and March 2025. Endoscopic examinations were performed under white-light, non-magnified observation with repeated AI assessments of each lesion. The rates of “Consider biopsy” (B) judgments were calculated. Lesions with a B judgment rate of ≥50% were defined as AI-positive and classified into four AI confidence categories. Diagnostic performance was assessed using sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). Factors associated with false-positive diagnoses were analyzed using penalized logistic regression. Results : The AI system demonstrated a sensitivity of 97.6% and an NPV of 95.7%, with a specificity of 45.8%. Pathology-positive rates decreased stepwise across the four AI confidence categories ( p < 0.001). Among AI-positive lesions, low regional reproducibility, lesion size ≥ 30 mm, scar, and erosion were independently associated with false-positive diagnoses. In analyses restricted to non-neoplastic lesions, lesion size ≥ 30 mm remained significantly associated with false-positive diagnosis. Conclusions : In real-world clinical practice, a commercially available AI system provides high sensitivity for early gastric cancer detection. Incorporating confidence stratification and regional reproducibility into clinical decision-making may enhance the effective use of AI-assisted endoscopic diagnosis beyond binary interpretations. Keywords: artificial intelligence, gastric cancer, upper gastrointestinal endoscopy, false positive, diagnostic confidence, real-world study 1. Introduction Upper gastrointestinal endoscopy involving image-enhanced endoscopy and magnified endoscopy plays a central role in the early detection and diagnosis of gastric cancer [ 1 , 2 ]. However, its diagnostic performance remains suboptimal. Previous studies have reported that the false-negative rate for gastric cancer ranges from 4.6% to 25.8%, mainly due to factors such as operator dependency, the small size or subtle morphology of lesions, and variability in observation quality [ 3 , 4 ]. These limitations highlight the need for diagnostic support tools that can improve detection accuracy and reduce interobserver variability. To address these challenges, artificial intelligence (AI)-assisted diagnostic technologies have been increasingly developed in the field of endoscopic imaging [ 5 , 6 , 7 ]. The first AI system designed to detect early gastric cancer was reported by Hirasawa et al. [ 8 ]. Since then, numerous deep learning–based AI models have been proposed. A recent meta-analysis of 21 studies showed that AI-assisted endoscopy significantly improved the sensitivity of early gastric cancer detection, particularly among non-expert endoscopists [ 9 ]. These findings suggest that AI may serve as a valuable adjunct to conventional endoscopy by enhancing diagnostic accuracy. A newly developed diagnostic support system, gastroAI™ model-G, aims to reduce missed diagnoses and provide high-accuracy endoscopic assessment for early gastric cancer. In a pre-market retrospective performance evaluation, gastroAI™ model-G achieved a sensitivity of 84.7% and a specificity of 58.2%, indicating promising diagnostic capability but also highlighting the issue of a relatively high false-positive rate [ 10 ]. In real-world clinical practice, however, several important questions remain unanswered regarding the optimal interpretation and use of AI outputs [ 11 , 12 ]. For example, it should be clarified how to handle situations where the same lesion is repeatedly evaluated and the results are inconsistent, how AI actually works for non-adenocarcinomatous tumors and various benign gastric lesions, and what factors related to the lesion or system contribute to false-positive diagnoses. The present study aimed to evaluate the diagnostic performance of gastroAI™ model-G (AI Medical Service Inc., Tokyo, Japan) across a broad spectrum of gastric lesions in real-world clinical practice, with particular emphasis on the clinical implications of repeated AI evaluations. In addition, we sought to identify factors associated with false-positive reactions and to clarify whether AI confidence stratification and regional reproducibility could serve as practical indicators for interpreting AI outputs beyond a simple binary classification. 2. Materials and Methods 2.1. Study Design and Patients This was a single-center, retrospective study conducted at Hamamatsu University School of Medicine. Consecutive patients who underwent endoscopic examination and/or treatment for suspected early gastric cancer between March 2024 and March 2025 were reviewed. Patients were considered to have “suspected early gastric cancer” if they were referred to our institution for further evaluation and potential endoscopic treatment following screening endoscopy at other institutions, or if suspicious lesions were identified during screening or surveillance endoscopy at our institution and required further diagnostic assessment. Endoscopic suspicion was based on findings such as irregular mucosal patterns, color changes, or abnormal morphology suggestive of neoplasia, at the discretion of the endoscopist. This approach reflects real-world clinical practice. In addition, inclusion was partly influenced by the availability of the AI system, which could be used only with a specific endoscopic platform dedicated to detailed examination. Among these, 47 patients with a total of 89 localized gastric lesions were included in the final analysis. All lesions were evaluated endoscopically as part of diagnostic assessment or therapeutic procedures for early gastric cancer. The final pathological diagnosis obtained from biopsy or endoscopic resection specimens was used as the reference standard for lesion classification. 2.2. Ethics Statement This study conformed to the principles of the Declaration of Helsinki. In accordance with the Ethical Guidelines for Medical and Health Research Involving Human Subjects issued by the Ministry of Education, Culture, Sports, Science and Technology and the Ministry of Health, Labour and Welfare of Japan, study information, including the objectives, was disclosed on our hospital website with an opt-out approach. The study protocol was reviewed and approved by the Ethics Committee of Hamamatsu University School of Medicine, Japan (institutional review board approval number: 24-071). Written informed consent was obtained from all participants. 2.3. Endoscopic Equipment and Examination Procedure All endoscopic examinations were performed using the EVIS X1 system (Olympus Medical Systems, Tokyo, Japan) with a GIF-XZ1200 gastroscope. Endoscopic examinations were performed under white-light, non-magnified observation, during which repeated AI assessments were conducted for each target lesion in real time. Video recordings were obtained simultaneously and subsequently used to analyze repeated AI assessments. 2.4. AI Assessment and Confidence Categorization The AI system used in this study was gastroAI™ model-G (AI Medical Service Inc., Tokyo, Japan), a deep learning–based diagnostic support system developed for gastric endoscopic image analysis. All AI evaluations are performed in real time during routine endoscopic examinations. When an endoscopist finds a lesion, they display it on the screen and press the freeze button, which activates the AI system and displays the AI evaluation. All video recordings were reviewed retrospectively to extract AI judgment data and assess regional reproducibility. This approach enabled a detailed evaluation of both diagnostic performance and the stability of AI outputs under real-world clinical conditions. For each lesion, multiple AI evaluations were performed, and the corresponding AI outputs were documented and recorded. The median number of AI assessments per lesion was five, allowing for analysis of repeated AI outputs and confidence variability. For each evaluation, the AI system classified the lesion as either “Consider biopsy (B)” or “Low confidence (LC)” ( Figure 1 ). The B judgment rate for each lesion was calculated as the proportion of evaluations classified as B among all AI assessments performed for that lesion. Lesions with a B judgment rate of ≥50% were defined as AI-positive, whereas those with a B judgment rate of <50% were defined as AI-negative. Figure 1. Open in a new tab AI assessment by gastroAI™ model-G. The system provides lesion-level outputs indicating the likelihood of neoplastic changes and categorizes findings using “Consider biopsy (B)” and “Low confidence (LC)” judgments. ( a ) The well-differentiated adenocarcinoma in the greater curvature of the antrum was assessed as “Consider biopsy” indicated in a square. ( b ) The benign gastric ulcer scar in the lesser curvature of the middle gastric body was assessed as “Low confidence (LC)”. Based on the distribution of B and LC judgments across repeated AI assessments, lesions were further classified into four AI confidence categories: B (100% B judgments), B/LC (B judgment rate 50–99%), LC/B (B judgment rate 1–49%), and LC (0% B judgments). These categories were used to evaluate the stepwise diagnostic characteristics and confidence structure of the AI system. 2.5. Regional Reproducibility Based on the recorded video data, regional reproducibility was assessed to determine whether the AI system consistently focused on the same lesion area across repeated evaluations ( Figure 2 ). Reproducibility was categorized using a predefined scoring system: score 1 (strong reproducibility, ≥70% overlap), score 2 (moderate reproducibility, 50–69% overlap), and score 3 (weak or absent reproducibility, <50% overlap). Figure 2. Open in a new tab Example of regional reproducibility assessed by AI. Regional reproducibility was assessed to evaluate whether the AI system consistently focused on the same lesion area across repeated evaluations. ( a ) Example of high regional reproducibility (score 1). Repeated AI assessments of gastric adenocarcinoma identified the same area in square, demonstrating regional reproducibility. ( b ) Example without regional reproducibility (score 3). Repeated AI assessment of benign gastric ulcer scars identified discordant areas in square and did not demonstrate regional reproducibility. This metric was used as an indicator of confidence for AI-positive outputs and was analyzed exclusively among AI-positive lesions, as true-negative lesions are inherently classified as having no reproducible lesion area by definition. 2.6. Statistical Analysis Categorical variables are presented as counts and percentages, and continuous variables are expressed as mean ± standard deviation or median with interquartile range, as appropriate. Differences between groups were assessed using Fisher’s exact test for categorical variables and the Mann–Whitney U test for continuous variables. Diagnostic performance of the AI system was evaluated using sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV), with the final pathological diagnosis as the reference standard. For analysis of AI confidence, lesions were classified into four categories (B, B/LC, LC/B, and LC) according to the distribution of B and LC judgments. Pathology-positive rates and regional reproducibility scores were compared across the four categories using the χ 2 test and the Kruskal–Wallis test, respectively. Monotonic trends across increasing AI confidence levels were evaluated using a trend test. Factors associated with false-positive AI diagnosis were examined in two predefined analyses. First, among AI-positive lesions (biopsy judgment rate ≥ 50%), true-positive and false-positive lesions were compared to identify factors associated with false-positive diagnosis. In this analysis, regional reproducibility was included as a candidate variable because it represents the consistency of AI attention among AI-positive outputs. Second, false-positive lesions were compared with true-negative lesions to identify factors associated with false-positive diagnosis among non-neoplastic lesions; regional reproducibility was excluded from this analysis because true-negative lesions are structurally classified as having no regional reproducibility by definition. Variables showing potential associations in univariate analyses or considered clinically relevant were entered into multivariable models. Multivariable analyses were performed using penalized logistic regression to account for small sample size and quasi-complete separation. Adjusted odds ratios (ORs) with 95% confidence intervals (CIs) were calculated. All statistical tests were two-sided, and a p -value < 0.05 was considered statistically significant. Statistical analyses were performed using appropriate statistical software. 3. Results 3.1. Overall Diagnostic Performance of the AI System A total of 89 gastric lesions were included in the analysis. Baseline characteristics and pathological diagnosis of gastric lesions are presented in Table 1 . Based on the predefined threshold for AI positivity (B judgment rate ≥ 50%), the AI system classified 66 lesions as positive and 23 as negative. According to the final pathological diagnosis, there were 40 true-positive, 26 false-positive, 22 true-negative, and 1 false-negative lesions. Table 1. Baseline characteristics and pathological diagnosis of gastric lesions (n = 89). Characteristics Value Patients, n 47 Lesions, n 89 Age, years (mean ± SD) 71.9 ± 9.5 Male sex, n (%) 74 (83.1%) Lesion size, mm (median, IQR) 10.0 (5.0–24.0) Lesion size ≥ 30 mm, n (%) 17 (19.1%) Lesion location Upper third 21 (23.6%) Middle third 31 (34.8%) Lower third 35 (39.3%) Esophagogastric junction 2 (2.2%) Helicobacter pylori status Never infected 18 (20.2%) Currently infected 15 (16.9%) Previously infected 53 (59.6%) Unknown 3 (3.4%) Pathological diagnosis Adenocarcinoma 41 (46.1%) Malignant lymphoma * 2 (2.2%) Non-neoplastic lesions 46 (51.7%) Fundic gland polyp 11 Hyperplastic polyp 8 Ulcer scar 8 Erosion 8 Others ‡ 11 Median number of evaluations per lesion 5 Total number of AI judgements 474 Number of considered biopsies (%) 373 (78.9%) Number of positivity in AI diagnosis (B judgment rate ≥ 50%) 66 (74.2%) Open in a new tab * Malignant lymphoma includes mucosa-associated lymphoid tissue (MALT) lymphoma and diffuse large B-cell lymphoma. ‡ Others include submucosal tumors and other non-epithelial lesions. The overall diagnostic performance of the AI system showed a high sensitivity of 97.6% and a negative predictive value of 95.7%, whereas specificity and positive predictive value were 45.8% and 60.6%, respectively ( Table 2 ). These results indicate that the AI system rarely missed early gastric cancer but generated a substantial number of false-positive judgments. Table 2. Diagnostic performance of the AI system for early gastric cancer (n = 89). AI Positive (n = 66) AI Negative (n = 23) Diagnostic Index Pathology positive (n = 41) 40 1 Sensitivity: 97.6% Pathology negative (n = 48) 26 22 Specificity: 45.8% PPV: 60.6% NPV: 95.7% Overall accuracy: 69.7% Open in a new tab AI positivity was defined as a “Consider biopsy” (B) judgment rate ≥ 50%. Sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV) were calculated using the final pathological diagnosis as the reference standard. 3.2. Stepwise Diagnostic Characteristics Across Four AI Confidence Categories Lesions were stratified into four AI confidence categories according to the distribution of B and LC judgments: B, B/LC, LC/B, and LC. As shown in Table 3 , the pathological positivity rate decreased stepwise across these categories, with the highest rate observed in the B group (75.6%), followed by the B/LC group (36.0%), the LC/B group (11.1%), and no cancer detected in the LC group (0%). The difference in pathological positivity rates among the four groups was statistically significant (χ 2 test, p < 0.001), and a strong monotonic trend was observed across increasing AI confidence levels (Spearman trend test, p < 0.001). Table 3. Stepwise diagnostic characteristics across four AI confidence categories. AI Confidence Category B Judgment Rate (%) N True Positive False Positive True Negative False Negative Pathologically Cancer Positive Rate (%) B 100 41 31 10 0 0 75.6 B/LC 70–99 25 9 16 0 0 36.0 LC/B 50–69 9 0 0 8 1 11.1 LC 1–49 14 0 0 14 0 0.0 Open in a new tab AI positivity was defined as a “Consider biopsy” (B) judgment rate ≥ 50%. Pathology-positive rate differed significantly across the four AI confidence categories (χ 2 test, p < 0.001). The pathology-positive rate in the LC/B group reflects one false-negative lesion. A significant monotonic trend was observed, with decreasing pathology-positive rates from B → B/LC → LC/B → LC (trend test, p < 0.001). Regional reproducibility scores also demonstrated a significant stepwise deterioration across the four AI confidence categories. It should be noted that lower reproducibility scores indicate stronger regional reproducibility, as score 1 represents high overlap of AI-detected regions. Median reproducibility scores were lowest (indicating the strongest reproducibility) in the B group and progressively increased in the B/LC, LC/B, and LC groups (Kruskal–Wallis test, p < 0.001). These findings indicate that the four-tier AI classification reflects a graded confidence structure rather than a simple binary decision. 3.3. Factors Associated with False-Positive Diagnosis Among AI-Positive Lesions (TP vs. FP) To evaluate factors associated with false-positive diagnosis among AI-positive lesions, analyses were restricted to lesions classified as AI-positive (B judgment rate ≥ 50%), consisting of 40 true-positive and 26 false-positive lesions. In univariate analyses, lower regional reproducibility, larger lesion size (≥30 mm), and the presence of scar or erosion were associated with false-positive diagnosis, whereas age, sex, Helicobacter pylori infection status, and color change were not significant independent predictors ( Figure 3 a). Multivariable penalized logistic regression demonstrated that low regional reproducibility, lesion size ≥ 30 mm, scar, and erosion were independently associated with false-positive diagnosis among AI-positive lesions ( Table 4 ). Figure 3. Open in a new tab Factors associated with false-positive AI diagnosis. ( a ) Forest plot of univariate odds ratios (ORs) with 95% confidence intervals (CIs) for false-positive AI diagnosis among AI-positive lesions (TP vs. FP). ( b ) Forest plot of univariate odds ratios (ORs) with 95% confidence intervals (CIs) for false-positive AI diagnosis among non-neoplastic lesions (FP vs. TN). Regional reproducibility was excluded from the FP vs. TN analysis due to its structural definition. Table 4. Multivariable analysis of factors associated with false-positive diagnosis among AI-positive lesions (TP vs. FP). Variable Adjusted OR 95% CI p Value Low regional reproducibility (score = 3) 2.70 1.22–6.67 0.015 Lesion size ≥ 30 mm 2.41 1.08–5.38 0.032 Scar 3.68 1.02–13.2 0.047 Erosion 5.43 1.11–26.6 0.037 Open in a new tab Multivariable analysis was performed using a penalized logistic regression model to account for quasi-complete separation. Odds ratios (ORs) and 95% confidence intervals (CIs) are shown. Analysis was restricted to AI-positive lesions (biopsy judgment rate ≥ 50%). These results indicate that, among AI-positive lesions, regional reproducibility functions as a confidence marker that discriminates true-positive from false-positive diagnoses. 3.4. Factors Associated with False-Positive Diagnosis Among Non-Neoplastic Lesions (FP vs. TN) To assess factors associated with false-positive diagnoses among non-neoplastic lesions, false-positive lesions were compared with true-negative lesions. Regional reproducibility was excluded from this analysis because, by definition, true-negative lesions are structurally classified as having no regional reproducibility. In this comparison, lesion size ≥ 30 mm was significantly associated with false-positive diagnosis, whereas patient age, sex, Helicobacter pylori infection status, and color change were not significantly associated ( Figure 3 b). Certain lesion types, such as fundic gland polyps, tended to be less frequently misclassified as false positives, although this association did not consistently reach statistical significance in multivariable analysis ( Table 5 ). Table 5. Multivariable analysis of factors associated with false-positive diagnosis among non-neoplastic lesions (FP vs. TN). Variable Adjusted OR 95% CI p Value Lesion size ≥ 30 mm 2.66 1.01–7.03 0.048 Fundic gland polyp 0.59 0.18–1.94 0.39 Open in a new tab Multivariable analysis was conducted using a penalized logistic regression model. 4. Discussion In this real-world, single-center retrospective study, we evaluated the diagnostic performance of an AI-based diagnostic support system for early gastric cancer and investigated factors associated with false-positive AI diagnoses. The principal findings of this study can be summarized as follows: (1) the AI system demonstrated high sensitivity and negative predictive value for early gastric cancer detection in routine clinical practice; (2) repeated AI evaluations revealed a stepwise confidence structure rather than a simple binary output; (3) among AI-positive lesions, false-positive diagnoses were systematically associated with low regional reproducibility and specific lesion characteristics; and (4) among non-neoplastic lesions, large lesion size was a major determinant of false-positive AI reactions. From a methodological perspective, this study provides a real-world framework for evaluating AI-assisted diagnosis using repeated assessments rather than single-image analysis. This approach may be particularly important for clinical implementation, as it reflects the dynamic interaction between endoscopists and AI systems during routine practice. First, the overall diagnostic performance observed in this study was characterized by high sensitivity with a very low false-negative rate. This finding is consistent with previous reports of AI-assisted endoscopic diagnosis and supports the role of AI as a safety net to reduce missed gastric cancers during routine upper gastrointestinal endoscopy [ 8 , 9 , 13 ]. Given that missed lesions remain a clinically significant problem even among experienced endoscopists, high sensitivity is a critical prerequisite for AI-assisted screening tools. However, this benefit was accompanied by relatively modest specificity, emphasizing that the clinical value of AI systems cannot be judged by sensitivity alone and that false-positive reactions must be appropriately interpreted and managed. Second, an important contribution of this study is the demonstration that AI outputs form a graded confidence hierarchy when repeated evaluations are considered. By stratifying lesions into four confidence categories based on the distribution of B and LC judgments, we observed a clear stepwise decrease in pathological positivity rates from the highest- to the lowest-confidence groups. These results indicate that AI-assisted diagnosis should not be interpreted as a binary decision but rather as a probabilistic signal conveying different levels of diagnostic confidence. Such confidence stratification may provide endoscopists with clinically meaningful context when deciding whether to perform a biopsy or continue observation, particularly in borderline or ambiguous lesions [ 14 ]. Third, our analysis revealed that false-positive AI diagnoses among AI-positive lesions were not random events but were associated with identifiable factors. Low regional reproducibility emerged as a key discriminator between true-positive and false-positive diagnoses. Regional reproducibility reflects the consistency with which the AI system focuses on the same lesion area across repeated evaluations, and our findings suggest that high reproducibility may serve as a practical confidence marker for AI-positive outputs. In addition, lesion size ≥ 30 mm and the presence of scar or erosion were independently associated with false-positive diagnoses, indicating that extensive mucosal changes and non-neoplastic structural alterations may trigger persistent AI activation despite the absence of malignancy. Fourth, when false-positive lesions were compared with true-negative lesions among non-neoplastic cases, lesion size ≥ 30 mm remained significantly associated with false-positive diagnosis. This finding suggests that large benign lesions may exhibit morphological features that resemble early gastric cancer and are therefore more likely to be misclassified by AI systems. Conversely, fundic gland polyps tended to be less frequently misclassified as false positives, implying that certain benign lesion types may be inherently less likely to activate cancer-oriented AI algorithms. These findings should be interpreted in the broader context of current research and clinical implementation of AI-assisted upper gastrointestinal endoscopy. Although numerous studies have reported high diagnostic performance of AI systems for early gastric cancer using image-enhanced or magnifying endoscopy, most of these investigations were conducted under highly controlled conditions, such as the use of static images, preselected datasets, or expert-centered environments [ 5 , 15 , 16 , 17 ]. In contrast, AI systems that have obtained regulatory approval and are currently available for routine clinical use remain limited, particularly for gastric cancer detection. This discrepancy highlights a critical gap between experimental AI performance and real-world clinical implementation [ 12 ]. In this study, one false-negative lesion was identified, which was a small raspberry-like tumor (foveolar-type gastric adenocarcinoma) ( Figure 4 ). This lesion presented as a subtle, reddish, protruded polyp and is known to represent a distinct morphological subtype that may be difficult to differentiate from hyperplastic polyps [ 18 ]. Notably, this lesion type is not included in the training dataset of the current AI system, as described in the product specifications. Therefore, the false-negative result in this case reflects a predefined limitation of the system rather than a random diagnostic error. This finding highlights an important boundary of AI-assisted diagnosis and underscores the need for continuous model updates to incorporate rare or newly recognized lesion subtypes. Figure 4. Open in a new tab Endoscopic findings of a false-negative case. A small, raspberry-like tumor less than 5 mm in size was observed in the greater curvature of the upper gastric body. White-light imaging revealed a small, reddish, protruding lesion. Chromoendoscopy clearly visualized the lesion surface without irregularity. Narrow-band imaging with magnification revealed irregular microvessels. The lesion was pathologically diagnosed as foveolar-type gastric adenocarcinoma, but the AI system showed a negative result, with a B rate of 33%. Recent editorials and review articles have emphasized that the clinical acceptability of AI-assisted endoscopy depends not only on sensitivity but also on the interpretability of AI outputs and the management of false-positive detections [ 15 ]. In this context, our findings provide practical insight into how AI outputs can be interpreted in routine practice. The observation that false-positive reactions are associated with specific lesion characteristics and AI confidence markers suggests that false positives should not be regarded simply as system errors but rather as signals requiring contextual interpretation by the endoscopist. Furthermore, recent pre-market studies of updated AI systems (gastroAI™ model-G2) using curated static image datasets have demonstrated the technical evolution of AI platforms under controlled conditions [ 19 ]. Although such studies differ fundamentally from real-world, video-based evaluations in terms of target population, dataset composition, and evaluation methodology, they underscore the importance of complementary validation strategies. Together, controlled image-based studies and real-world clinical assessments may provide a more comprehensive understanding of AI performance across different clinical contexts. From a clinical perspective, the results of this study suggest that future development of AI-assisted upper gastrointestinal endoscopy should move beyond binary classification toward confidence-aware and context-sensitive decision support systems. Incorporating confidence stratification and regional reproducibility into AI outputs may facilitate more appropriate endoscopist–AI interaction, optimize biopsy strategies, and ultimately promote wider clinical adoption of AI technologies in gastric cancer screening and diagnosis. Several limitations of this study should be acknowledged. First, this was a single-center, retrospective study with a relatively limited sample size, which may restrict the generalizability of the findings. Second, the selection process for “suspected early gastric cancer” may introduce a degree of referral and device-related selection bias, which should be considered when interpreting the generalizability of the results. Third, the number of repeated AI assessments per lesion was determined at the discretion of the examining endoscopist, which may have introduced variability and potential bias in the calculation of AI judgment rates and confidence stratification. Fourth, malignant lymphoma, including mucosa-associated lymphoid tissue lymphoma and diffuse large B-cell lymphoma, was classified as non-adenocarcinoma and analyzed as non-cancer. While this approach allowed a clear focus on early gastric adenocarcinoma detection, it may have influenced the interpretation of false-positive AI reactions for non-epithelial malignant lesions. Fifth, regional reproducibility was assessed using a predefined categorical scoring system based on spatial overlap, and alternative quantitative methods may yield different results. Finally, all examinations were performed using a single endoscopic system at a single institution, and operator- and equipment-specific effects cannot be completely excluded. In conclusion, this real-world study demonstrates that an AI-based diagnostic support system provides high sensitivity for early gastric cancer detection, while false-positive diagnoses are systematically associated with specific lesion characteristics and AI confidence markers. Integrating confidence stratification and regional reproducibility into clinical decision-making may enable more nuanced interpretation of AI outputs and support the safe and effective implementation of AI-assisted upper gastrointestinal endoscopy beyond binary diagnostic frameworks. Acknowledgments We would like to express our gratitude to AI Medical Service Inc. for providing technical support related to the gastroAI™ model-G system. Abbreviations The following abbreviations are used in this manuscript: AI Artificial intelligence B Consider biopsy LC Low confidence NPV Negative predictive value PPV Positive predictive value OR Odds ratio CI Confidence interval Open in a new tab Author Contributions Conceptualization, S.O. (Satoshi Osawa) and T.Y.; methodology, S.O. (Satoshi Osawa) and T.Y.; investigation, S.O. (Satoshi Osawa), T.Y., W.I., T.N., K.T., T.E., K.I., T.T., T.I., S.T., S.O. (Shunya Onoue), Y.A., K.S. (Kiichi Sugiura), T.M. and N.I.; formal analysis, S.O. (Satoshi Osawa), T.Y. and M.I.; data curation, S.O. (Satoshi Osawa) and T.Y.; validation, S.O. (Satoshi Osawa) and T.Y.; visualization, S.O. (Satoshi Osawa); supervision, K.S. (Ken Sugimoto); project administration, S.O. (Satoshi Osawa) and T.Y.; writing—original draft preparation, S.O. (Satoshi Osawa); writing—review and editing, T.Y., M.Y. and Y.H. All authors have read and agreed to the published version of the manuscript. Institutional Review Board Statement The Ethics Committee of Hamamatsu University School of Medicine in Japan reviewed and approved the study protocol (institutional review board approval number: 24-071, date of approval: 5 June 2024). Informed Consent Statement Written informed consent was obtained from all participants. Data Availability Statement The data presented in this study are available on reasonable request from the corresponding author. The data are not publicly available due to ethical and privacy restrictions related to patient information. Conflicts of Interest The authors declare that they have no conflicts of interest. Funding Statement This research received no external funding. Footnotes Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content. References 1. Muto M., Yao K., Kaise M., Kato M., Uedo N., Yagi K., Tajiri H. Magnifying endoscopy simple diagnostic algorithm for early gastric cancer (MESDA-G) Dig. Endosc. 2016;28:379–393. doi: 10.1111/den.12638. [ DOI ] [ PubMed ] [ Google Scholar ] 2. Yao K., Uedo N., Kamada T., Hirasawa T., Nagahama T., Yoshinaga S., Oka M., Inoue K., Mabe K., Yao T., et al. Guidelines for endoscopic diagnosis of early gastric cancer. Dig. Endosc. 2020;32:663–698. doi: 10.1111/den.13684. [ DOI ] [ PubMed ] [ Google Scholar ] 3. 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 ] [ Google Scholar ] 4. Pimenta-Melo A.R., Monteiro-Soares M., Libânio D., Dinis-Ribeiro M. Missing rate for gastric cancer during upper gastrointestinal endoscopy: A systematic review and meta-analysis. Eur. J. Gastroenterol. Hepatol. 2016;28:1041–1049. doi: 10.1097/MEG.0000000000000657. [ DOI ] [ PubMed ] [ Google Scholar ] 5. Hu H., Gong L., Dong D., Zhu L., Wang M., He J., Shu L., Cai Y., Cai S., Su W., et al. Identifying early gastric cancer under magnifying narrow-band images with deep learning: A multicenter study. Gastrointest. Endosc. 2021;93:1333–1341.e1333. doi: 10.1016/j.gie.2020.11.014. [ DOI ] [ PubMed ] [ Google Scholar ] 6. Wu L., Zhou W., Wan X., Zhang J., Shen L., Hu S., Ding Q., Mu G., Yin A., Huang X., et al. A deep neural network improves endoscopic detection of early gastric cancer without blind spots. Endoscopy. 2019;51:522–531. doi: 10.1055/a-0855-3532. [ DOI ] [ PubMed ] [ Google Scholar ] 7. Wu L., Shang R., Sharma P., Zhou W., Liu J., Yao L., Dong Z., Yuan J., Zeng Z., Yu Y., et al. 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. doi: 10.1016/S2468-1253(21)00216-8. [ DOI ] [ PubMed ] [ Google Scholar ] 8. Hirasawa T., Aoyama K., Tanimoto T., Ishihara S., Shichijo S., Ozawa T., Ohnishi T., Fujishiro M., Matsuo K., Fujisaki J., et al. Application of artificial intelligence using a convolutional neural network for detecting gastric cancer in endoscopic images. Gastric Cancer. 2018;21:653–660. doi: 10.1007/s10120-018-0793-2. [ DOI ] [ PubMed ] [ Google Scholar ] 9. Shi Y., Fan H., Li L., Hou Y., Qian F., Zhuang M., Miao B., Fei S. The value of machine learning approaches in the diagnosis of early gastric cancer: A systematic review and meta-analysis. World J. Surg. Oncol. 2024;22:40. doi: 10.1186/s12957-024-03321-9. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 10. Hirasawa T., Ikenoyama Y., Ishioka M., Namikawa K., Horiuchi Y., Nakashima H., Fujisaki J. Current status and future perspective of artificial intelligence applications in endoscopic diagnosis and management of gastric cancer. Dig. Endosc. 2021;33:263–272. doi: 10.1111/den.13890. [ DOI ] [ PubMed ] [ Google Scholar ] 11. Mori Y., Ishihara R., Ogata H., Kutsumi H., Saito Y., Sumiyama K., Sekiguchi M., Tajiri H., Fujishiro M., Matsuda K., et al. Artificial Intelligence in Gastrointestinal Endoscopy: The Japan Gastroenterological Endoscopy Society Position Statements. Dig. Endosc. 2025;37:1116–1122. doi: 10.1111/den.15075. [ DOI ] [ PubMed ] [ Google Scholar ] 12. Ebigbo A., Messmann H., Lee S.H. Artificial Intelligence Applications in Image-Based Diagnosis of Early Esophageal and Gastric Neoplasms. Gastroenterology. 2025;169:396–415.e392. doi: 10.1053/j.gastro.2025.01.253. [ DOI ] [ PubMed ] [ Google Scholar ] 13. Ikenoyama Y., Hirasawa T., Ishioka M., Namikawa K., Yoshimizu S., Horiuchi Y., Ishiyama A., Yoshio T., Tsuchida T., Takeuchi Y., et al. Detecting early gastric cancer: Comparison between the diagnostic ability of convolutional neural networks and endoscopists. Dig. Endosc. 2021;33:141–150. doi: 10.1111/den.13688. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 14. Topol E.J. High-performance medicine: The convergence of human and artificial intelligence. Nat. Med. 2019;25:44–56. doi: 10.1038/s41591-018-0300-7. [ DOI ] [ PubMed ] [ Google Scholar ] 15. Lei C., Sun W., Wang K., Weng R., Kan X., Li R. Artificial intelligence-assisted diagnosis of early gastric cancer: Present practice and future prospects. Ann. Med. 2025;57:2461679. doi: 10.1080/07853890.2025.2461679. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 16. Kanesaka T. Artificial Intelligence for the Detection of Neoplastic Lesions During Upper Gastrointestinal Endoscopy: Diagnostic Performance and Future Directions. Dig. Endosc. 2026;38:e70090. doi: 10.1111/den.70090. [ DOI ] [ PubMed ] [ Google Scholar ] 17. Ueyama H., Kato Y., Akazawa Y., Yatagai N., Komori H., Takeda T., Matsumoto K., Ueda K., Matsumoto K., Hojo M., et al. Application of artificial intelligence using a convolutional neural network for diagnosis of early gastric cancer based on magnifying endoscopy with narrow-band imaging. J. Gastroenterol. Hepatol. 2021;36:482–489. doi: 10.1111/jgh.15190. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 18. Yatagai N., Ueyama H., Ikemura M., Uchida R., Utsunomiya H., Abe D., Oki S., Suzuki N., Ikeda A., Akazawa Y., et al. Clinicopathological and Endoscopic Features of Raspberry-Shaped Gastric Cancer in Helicobacter pylori-Uninfected Patients. Digestion. 2020;15:41–48. doi: 10.1159/000511907. [ DOI ] [ PubMed ] [ Google Scholar ] 19. Mizutani H., Tsuji Y., Kubota D., Hisada H., Miura Y., Ohki D., Takeuchi C., Kakushima N., Yamamichi N., Kikuchi R., et al. Impact of interaction between an artificial intelligence endoscopic support system and endoscopists on diagnosis of gastric neoplastic lesions. Endosc. Int. Open. 2025;13:a26950556. doi: 10.1055/a-2695-0556. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement The data presented in this study are available on reasonable request from the corresponding author. The data are not publicly available due to ethical and privacy restrictions related to patient information. Articles from Journal of Clinical Medicine are provided here courtesy of Multidisciplinary Digital Publishing Institute (MDPI) ACTIONS View on publisher site PDF (1.8 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top