Gastric cancer (GC) is a major global health burden, ranking fifth in incidence and fourth in cancer-related mortality worldwide (Sung et al., 2021). Early gastric cancer (EGC) is defined as adenocarcinoma that infiltrates the mucosa or submucosa of the stomach with or without lymph node metastases (T1, any N), which is associated with a favorable prognosis and a five-year survival rate of approximately 95% (Öhman et al., 1980; GASTRIC (Global Advanced/Adjuvant Stomach Tumor Research International Collaboration) Group et al., 2013; Katai et al., 2018; Yang et al., 2021). Consequently, early detection of EGC is critical for improving patient clinical outcomes.

Upper gastrointestinal endoscopy has been established as the gold standard for the diagnosis of EGC (Machlowska et al., 2020). Among its various imaging modalities, white-light endoscopy remains the preferred technique in routine clinical practice due to its widespread availability and ease of use (Nagula et al., 2024). Evidence from South Korea has demonstrated that screening upper gastrointestinal endoscopy has significantly increased the detection of EGC and reduced mortality by approximately 50% (OR = 0.53, 95% CI: 0.51–0.56) (Jun et al., 2017; Arnold et al., 2020). However, EGC lesions often present with subtle mucosal changes, such as microsurface architectural disruption and color irregularities, making their detection challenging under standard white-light endoscopy during routine screening (Zhang et al., 2011; Liu et al., 2023). As a result, the accuracy of EGC detection is highly dependent on endoscopist expertise, resulting in variability in diagnostic performance. Indeed, previous studies have shown that senior endoscopists with more than 10 years of experience achieved significantly higher diagnostic sensitivity in detecting EGC compared to junior endoscopists with only 2–3 years of training (Tang et al., 2020; Yuan et al., 2022).

To address the aforementioned challenges, deep learning (DL)-based artificial intelligence (AI) has been increasingly applied to medical imaging, showing substantial promise in improving diagnostic sensitivity and specificity (Esteva et al., 2019; Gandhi et al., 2025c). Compared to traditional machine learning, DL algorithms possess several advantages. First, they possess the ability to perform feature self-learning from medical image datasets, eliminating the need for manual feature extraction and avoiding potential performance degradation caused by inaccurate or inconsistent segmentation. Second, they can be trained in an end-to-end manner, mapping raw images to diagnostic outputs while jointly optimizing all components of the network (Baldominos et al., 2019; Wang et al., 2019b; Zhou Z. et al., 2023). In recent years, DL algorithms have been widely investigated in the field of pathological image analysis. Numerous studies have consistently demonstrated high diagnostic accuracy in tumor detection across multiple cancer types, including breast, lung, and colorectal cancers, as well as glioma (Wang et al., 2019a; Im et al., 2021; Li et al., 2022, 2025; Thalakottor et al., 2023). In the diagnosis of EGC using endoscopic images, a previous meta-analysis found that conventional AI achieved a sensitivity of 86% and a specificity of 90%, demonstrating diagnostic accuracy comparable to that of experienced endoscopists (Chen P.-C. et al., 2022). However, the aforementioned meta-analysis included a limited number of studies and did not specifically evaluate the performance of deep learning algorithms in detecting EGC under white-light endoscopy.

Therefore, this systematic review synthesizes the latest developments and analyzes the diagnostic performance of DL algorithms on white-light endoscopy image datasets in EGC diagnosis. Meanwhile, our study further compared the diagnostic performance for EGC between DL algorithms and expert endoscopists. The findings will provide evidence-based support for the clinical translation of DL algorithms in upper gastrointestinal endoscopy for EGC.

This meta-analysis was conducted in full compliance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses of Diagnostic Test Accuracy (PRISMA-DTA) guidelines (Supplementary Table 1) (Mdf et al., 2018). Additionally, the study protocol has been registered in the PROSPERO database (CRD420251112418).

To the best of our knowledge, this is the first meta-analysis to comprehensively evaluate the performance of DL algorithms in diagnosing EGC using white light endoscopic images. The results indicate that DL algorithms exhibit excellent diagnostic performance in the internal validation set, with a sensitivity of 0.91, a specificity of 0.93, and an AUC of 0.97. In the external validation set, the diagnostic sensitivity, specificity, and AUC were 0.82, 0.83, and 0.89, respectively, which were lower than those in the internal validation set. Furthermore, no significant differences were observed between DL algorithms and expert endoscopists in terms of diagnostic sensitivity or specificity. Meta-regression analysis indicates that the sample size of the training dataset contributes to the high heterogeneity in sensitivity and specificity observed in the internal validation sets. In summary, these results suggest that DL algorithms demonstrate good diagnostic performance in detecting EGC using white-light endoscopic images, indicating their potential as a reliable auxiliary diagnostic tool.

Sensitivity and specificity are key metrics for evaluating diagnostic performance. In this study, the DL model demonstrated high sensitivity and specificity in the internal validation set. High sensitivity indicates a low risk of missed diagnosis, facilitating the detection of EGC with atypical morphology or indistinct borders. High specificity reflects a low false-positive rate, conducive to reducing unnecessary biopsy procedures and thereby preventing overdiagnosis and overtreatment. The strong performance observed in the internal validation may be attributed to consistent data preprocessing, standardized image acquisition protocols, and uniform endoscopic imaging conditions (Li et al., 2025). These factors help minimize technical variability, enabling the model to more accurately distinguish EGC from non-EGC findings. However, in the external validation set, both sensitivity and specificity were lower than those observed in the internal validation. This performance decline is likely due to real-world variations across institutions, such as differences in endoscopist expertise, types of endoscopic equipment, and image quality (Campanella et al., 2019). These heterogeneities introduce noise and complexity that the model may not have fully accounted for during training. These findings underscore the importance of standardized data pipelines and the use of diverse, multi-center datasets during model development to improve model generalizability and robustness.

Currently, due to limitations in technical skills and clinical experience, trainee endoscopists exhibit significantly lower sensitivity and specificity in diagnosing EGC compared to expert endoscopists (Ende et al., 2018; Tang et al., 2020; Yuan et al., 2022). This performance gap contributes to instability in clinical endoscopic practice and increases the risk of missed or incorrect diagnoses, especially in primary care hospitals. Previous studies revealed that, with AI assistance, trained novices can produce expert-level lung and cardiac ultrasound images that can be used to assess pathology after a short training session, thereby enhancing access to diagnosis in resource-constrained settings (Narang et al., 2021; Baloescu et al., 2025). In this study, our results demonstrate that DL algorithms achieve sensitivity and specificity comparable to those of expert endoscopists. Therefore, it is reasonable to hypothesize that AI may serve as an effective assistive tool to enhance the sensitivity and specificity of trainee endoscopists in the detection of EGC during white-light endoscopy screening, thereby minimizing the likelihood of missed or incorrect diagnoses and facilitating earlier detection and timely intervention.

In the internal validation of deep-learning algorithms, meta-regression analysis demonstrated that models trained on large-scale public datasets exhibited significantly superior diagnostic sensitivity and specificity compared to those trained on small-scale institutional datasets. This finding indicates that the size of the training dataset may be one of the key factors determining the diagnostic performance of the deep-learning algorithms. Previous studies revealed that merely expanding the size of the training dataset can improve the classification performance of the DL network (Kiryati and Landau, 2021; Pei et al., 2021). However, due to the challenges in acquiring and annotating medical imaging data, particularly in the three-dimensional context of endoscopic examinations, constructing large and high-quality training datasets was difficult (Tajbakhsh et al., 2016; Chen X. et al., 2022). In contrast, public datasets offered a viable pathway to overcome these difficulties. ImageNet is a large-scale hierarchical visual recognition database developed in the United States, comprising 14 million manually labeled images (Kang et al., 2021). Kvasir-SEG is a publicly accessible high-quality gastrointestinal endoscopy dataset originating from Norway, comprising 1,000 images annotated with pixel-level segmentation masks (Jha et al., 2019). Consequently, in this meta-analysis, deep-learning algorithms trained on ImageNet and Kvasir-SEG datasets achieve superior performance in EGC detection. Furthermore, although cross-validation is an important technique for evaluating model robustness, particularly in studies with small datasets, our analysis did not observe a significant influence of cross-validation on heterogeneity within the internal dataset (Aggarwal et al., 2022). Similarly, factors including the number of participating centers, image size, and study quality did not contribute significantly to internal heterogeneity. However, this heterogeneity may stem from other potential factors such as clinical staging of EGC, image quality, and variations in the definition of EGC.

To our knowledge, this is the first meta-analysis specifically evaluating the diagnostic performance of DL algorithms for EGC. In contrast, a prior meta-analysis of 12 studies reported that AI—encompassing both machine learning and DL algorithms—achieved a sensitivity of 0.86 and a specificity of 0.90 in the diagnosis of EGC, values notably lower than the 0.91 and 0.93 observed in this study (Chen P.-C. et al., 2022). This discrepancy may be attributed to differences in algorithmic model selection (DL versus a combination of machine learning and DL). At the algorithmic level, traditional machine learning methods rely on handcrafted feature engineering and exhibit limited generalizability, particularly when applied to complex and heterogeneous medical imaging data (Moawad et al., 2022). In contrast, the DL models evaluated in this study enable end-to-end learning by automatically extracting hierarchical feature representations directly from raw images, thereby achieving enhanced robustness and higher diagnostic accuracy in complex visual recognition tasks (Wang et al., 2019b).

With a pre-test probability of 36%, the Fagan nomogram demonstrated a positive post-test probability of 73% and a negative post-test probability of 11%. This provides a practical tool for clinicians: for a patient with a pre-test suspicion of 36%, a positive result from the DL model would increase the probability of EGC to 73%, warranting a confirmatory biopsy. Conversely, a negative result would lower the probability to 11%, potentially supporting a decision for surveillance rather than immediate intervention, depending on the clinical context. From a clinical implementation perspective, these findings support the role of DL-based systems as decision-support tools rather than standalone diagnostic solutions. Practical deployment would require targeted training for endoscopists on AI-assisted interpretation within endoscopy suites, alongside clearly defined safety workflows to ensure clinician oversight (Olawuyi and Viriri, 2025). Moreover, regulatory approval is a prerequisite for clinical adoption. Similar to AI-based electrocardiogram detection systems, AI models for early gastric cancer detection require formal evaluation and regulatory clearance from authorities such as the FDA or CE bodies (Singla et al., 2025). Such approval usually depends on robust external and prospective validation, which remains limited in current studies. From a methodological perspective, future improvements in DL-based EGC detection may benefit from incorporating Transformer-based architectures (e.g., Vision Transformer and Swin Transformer), which have shown strong performance in medical image analysis by capturing long-range spatial dependencies (Gandhi et al., 2025a). In addition, generative data augmentation techniques could help mitigate data imbalance and enhance model robustness (Gandhi et al., 2025b). The integration of multimodal learning frameworks, combining endoscopic video data with relevant clinical information, may further improve diagnostic accuracy and clinical relevance (Qin et al., 2025). To address data privacy and enhance generalizability, federated learning offers a promising strategy for leveraging multicenter data without direct data sharing (Assaf et al., 2025). Finally, adoption of standardized reporting guidelines, such as CONSORT-AI, DECIDE-AI, and STARD-AI, is essential to improve transparency, reproducibility, and clinical interpretability of future studies (Goh et al., 2025).

In addition, it is important to note that the generalizability of these performance estimates may be further challenged by the lack of temporal validation in most included studies. Robust clinical prediction systems require testing on data from future time periods to ensure stability against shifts in clinical practice, equipment, or patient demographics. In our review, only one study employed a prospective external validation set (Sakai et al., 2018). Future studies should prioritize this design to provide a more rigorous and clinically realistic assessment of model performance over time. Furthermore, the presence of publication bias in the external validation dataset likely stems from the limited number of available studies and potential selective reporting of higher-performing models in externally validated literature. The observed bias suggests that the overall diagnostic performance of DL models in external validation settings may be overestimated in the current literature. Therefore, the establishment of multi-center, large-scale external validation cohorts is essential for a comprehensive evaluation of DL model performance.

Several limitations of this meta-analysis should be acknowledged when interpreting the findings.

First, a fundamental limitation of this analysis is that all included studies utilized retrospective datasets for both model development and validation. This retrospective design inherently carries risks of selection bias and spectrum bias, where the case mix may not fully represent the broader clinical population encountered in practice. Therefore, while our meta-analysis suggests promising diagnostic potential, the reported high accuracy likely represents a “best-case” scenario. Forthcoming multi-center, prospective trials are crucial to rigorously evaluate model performance in unselected, consecutive patients under real-world conditions (Tong et al., 2023). Second, there was variability in the definition of EGC across the included studies, and the inclusion criteria for control groups were inconsistent. This heterogeneity in the control population—ranging from purely normal mucosa to a mix of benign lesions (e.g., gastric ulcers, low-grade epithelial neoplasia, gastric polyps)—constitutes a potential source of classification bias. Such inconsistency may lead to systematic differences in model training and evaluation, as models trained against purely normal mucosa might achieve higher specificity in distinguishing cancer from normal tissue but potentially lower sensitivity for discriminating early cancer from challenging benign or precancerous conditions. Third, the current analysis was restricted to image-level evaluation of DL models due to incomplete patient-level data in the included studies. However, patient-level assessment better aligns with clinical practice. Image-based training risks overfitting to specific features within individual patients, which may limit the model’s applicability to external datasets (Lengerich et al., 2018). Fourth, all included studies focused on detecting gastric lesions from static, high-quality white-light endoscopic images, which inherently cannot reproduce the complexity of real-time endoscopy. In actual clinical practice, endoscopic observation is dynamic and often affected by motion blur caused by scope movement, variations in illumination, changes in viewing angle, and transient interference from mucus, blood, bubbles, or food residue. These in situ factors substantially increase diagnostic difficulty but were largely excluded from the training and validation datasets of the included studies. Consequently, the reported diagnostic performance of AI models derived from idealized image datasets may overestimate their effectiveness in real-world, real-time clinical settings. Future studies should therefore prioritize validation using video-based or real-time endoscopic data that better reflect routine clinical conditions.

In conclusion, our meta-analysis provides robust evidence that DL algorithms exhibit high diagnostic efficacy in detecting EGC from white-light endoscopic images. Moreover, the sensitivity and specificity of these algorithms are comparable to those of expert endoscopists. These findings highlight the potential for DL algorithms to serve as a clinical decision-support tool in routine practice.