INTRODUCTION
Gastric cancer (GC) is one of the leading causes of cancer-related mortality worldwide, with a particularly high incidence in Asia, including Korea. According to the Correa cascade, intestinal-type gastric adenocarcinoma develops through a stepwise progression from chronic gastritis to atrophic gastritis, intestinal metaplasia (IM), and dysplasia. Therefore, early detection and appropriate management of these premalignant lesions is essential for prevention of GC.
The use of artificial intelligence (AI) in upper gastrointestinal (UGI) endoscopy has been rapidly expanding, with growing evidence demonstrating its potential to improve diagnostic accuracy and reduce interobserver variability. Multiple narrative and systematic reviews on this topic have been published, including in this journal, and most have concluded that AI systems achieve high diagnostic accuracy across gastric premalignant conditions [1]. Rather than restating these pooled conclusions, the present review aims to provide a critical reappraisal of the evidence base with a focus on clinical implementation.
CHRONIC ATROPHIC GASTRITIS AND IM
Chronic atrophic gastritis (CAG) and gastric IM are well-established precursors of GC, and endoscopic surveillance is recommended for patients with extensive atrophy or IM [2]. However, the accuracy of conventional endoscopic diagnosis is limited, and the international standard for diagnosis and grading relies on histopathology-based systems such as the updated Sydney system and the Operative Link on Gastritis Assessment (OLGA)/Operative Link on Gastric Intestinal Metaplasia (OLGIM) staging systems [3-5]. These histologic assessments require multiple biopsies, which prolong procedure time, increase cost, and entail additional risk of complications. Consequently, an AI tool that can accurately recognize and grade these lesions on endoscopic images would have substantial clinical utility, and a number of investigations have addressed this need. Most of these studies have employed convolutional neural network (CNN)-based deep-learning models, with the overall diagnostic performance reported to approach or even exceed that of expert endoscopists. A meta-analysis of eight studies on CAG reported pooled AI sensitivity, specificity, and area under the curve (AUC) values of 94%, 96%, and 0.98, respectively, with AI achieving significantly higher specificity than endoscopists in head-to-head subgroup comparisons (96% vs. 82%, p<0.001) [6]. A separate meta-analysis of 12 studies on IM reported pooled AI sensitivity, specificity, and AUC values of 94%, 93%, and 0.97, respectively [7].
Use of AI for identification of CAG
Attempts to automate the endoscopic recognition of gastritis and its precancerous sequelae predate the deep-learning era. In a total of 104 prospectively enrolled patients with dyspepsia, Huang et al. [8] extracted 84 quantitative parameters from conventional endoscopic images and applied a refined feature selection with neural network approach to predict the histologic features of H. pylori density, acute and chronic inflammation, atrophy, IM, and lymphoid follicles. The prediction accuracy exceeded 80% for atrophy and IM at all three sampled sites, and the model was more accurate than all six comparator endoscopists for prediction of antral IM. The study also anticipated limitations that persist today: performance varied by anatomical site, with the sensitivity for IM reaching approximately 85% in the body and cardia but falling in the antrum, and the model was never tested outside its development cohort. Two decades later, CNNs addressed the same task with far larger datasets and without hand-crafted features.
Early CNN-based investigations focused on establishing the diagnostic baseline by using static images. Zhang et al. [9] developed a deep convolutional neural network (DCNN) that achieved an accuracy of 94.2% and sensitivity of 94.5% in diagnosing CAG from white-light endoscopy (WLI) images of the gastric antrum, outperforming human experts. Similarly, Guimarães et al. [10] utilized a VGG16-based model on real-world images from the proximal stomach (body and fundus), reporting an accuracy of 93% and an AUC of 0.98, which significantly surpassed human endoscopists in a head-to-head comparison. To address the limitations of WLI, Xu et al. [11] employed a multicenter dataset to train a DCNN for detecting gastric atrophy and IM by using image-enhanced endoscopy (narrow-band imaging [NBI]/blue-light imaging [BLI]); they achieved diagnostic accuracies of 0.869 for gastric atrophy and 0.888 for IM, which were comparable to those of expert endoscopists. In terms of risk stratification, Tao et al. [12] constructed a system based on the Kimura–Takemoto classification [13], which automatically stratified CAG severity (mild, moderate, severe) and achieved 94.12% accuracy in video tests. Recent studies have transitioned from static diagnoses to real-time clinical-decision support. Zhao and colleagues [14-16] developed a real-time video-monitoring system based on the U-Net architecture, which not only detected CAG with higher sensitivity than human endoscopists (89.3% vs. 67.6%), but also successfully automated OLGA staging during live procedures.
Beyond conventional endoscopy, AI has been also used in optical biopsies and cloud-based screening. Liu et al. [17] utilized deep learning with probe-based confocal laser endomicroscopy to diagnose the entire Correa cascade (inflammation, atrophy, IM, and neoplasia) with accuracies ranging from 91% to 97%, potentially eliminating the need for physical biopsies in some cases. In the field of digital pathology, Fang et al. [18] developed GasMIL, a semi-supervised multiple-instance learning model that graded gastric atrophy and IM on whole-slide images and achieved high diagnostic accuracy for IM (AUC=0.884) and gastric atrophy (AUC=0.877), consistently outperforming 10 pathologists. Finally, addressing healthcare disparities, Chiang et al. [19] implemented a “Rural-to-Center” AI model that allowed evaluation of endoscopic images from remote islands by cloud-based AI in a tertiary medical center, demonstrating high accuracy for H. pylori (91.3%) and CAG (79.9%) in routine practice and bridging gaps in access and expertise in resource-limited settings.
Use of AI for identification of IM
CNN-based systems have been used for the optical diagnosis of IM across a range of imaging modalities and clinical settings. Xu et al. [11] trained a system on 6250 image-enhanced endoscopy images from five Chinese hospitals and reported accuracies of 0.908, 0.859, and 0.898 for the identification of IM in the internal, multicenter external, and prospective video test sets, respectively; in a man–machine comparison, the performance of the system was comparable to that of experts (accuracy, 0.888 vs. 0.820, p=0.117) and superior to that of non-experts (accuracy, 0.736, p=0.028). In Korea, Gong et al. [20] developed the first model to classify the entire spectrum of gastric carcinogenesis (normal, atrophy, IM, dysplasia, early GC [EGC], and advanced GC) in a single architecture, and the model achieved internal-test lesion-classification accuracy of 91.2% and external-test accuracy of 82.3%, with a per-class AUC value of 91.3% for IM on external testing. Efforts to extend these results beyond East Asia and beyond still-image classification remain comparatively limited. Ligato et al. [21] reported a Western pilot study that used BLI of the gastric corpus, in which a patch-based ResNet-50 classified 200×200 pixel patches with accuracy of 74% and, after optimization of a learnable voting scheme, classified whole images with accuracy of 78% (recall, 100%; precision, 70%). Pornvoraphat et al. [22] went beyond image-level classification and trained a model to outline IM directly on the endoscopic image, at a speed sufficient for display during a live procedure without perceptible lag. The image-level performance of the model was high (sensitivity, 91%; specificity, 96%), but the average overlap between the region highlighted by the model and the region annotated by the endoscopist was only 55%, indicating that the system identifies where IM is likely to be present more reliably than it defines its borders. Niu et al. [23] used magnifying endoscopy with NBI not only to grade IM lesions (AUC: 0.872 in the internal validation set and 0.803 in the external validation set) but to assign a patient-level OLGIM stage, identifying high-risk stage III/IV disease with an accuracy of 84.0%.
Taken together, these studies indicate that AI models can recognize IM on endoscopic images with performance levels broadly comparable to those of experienced endoscopists and consistently superior to that of non-experts, and that this trend persisted across WLI, NBI, and BLI. The clinical value of this capability is not the diagnosis itself, but what it could replace. If IM can be reliably identified and graded optically, biopsies could be directed to the most representative mucosa rather than performed in accordance with a fixed protocol; patients could be assigned appropriate surveillance intervals; and the assessments could be extended to settings where expert endoscopy or histopathological assessment is not readily available. Several of the studies reviewed above were designed with exactly these goals. However, the ability of these models to retain such performance outside the datasets in which they were developed is the decisive question for clinical adoption.
GASTRIC DYSPLASIA
Detection and diagnosis of gastric neoplasia is one of the most actively investigated applications of AI in UGI endoscopy. Numerous studies have used WLI, NBI, and other image-enhanced modalities to demonstrate that the diagnostic performance of AI models for gastric dysplasia and EGC is comparable to that of expert endoscopists. However, most of these studies did not treat dysplasia as a separate endpoint but pooled low-grade intraepithelial neoplasia (LGIN), high-grade intraepithelial neoplasia (HGIN), and EGC under the single label “gastric neoplasm,” so their dysplasia-specific performance usually had to be recovered from subgroup data.
The strongest evidence for detection was obtained from randomized controlled trials (RCTs). A recent systematic review and meta-analysis pooled the findings from 11 RCTs with 57512 participants that compared AI-assisted esophagogastroduodenoscopy (AI-EGD) with conventional esophagogastroduodenoscopy (Co-EGD) [24]. Of these 11 trials, four evaluated esophageal neoplasms exclusively, four evaluated gastric neoplasms exclusively, and three evaluated neoplasms of the stomach, esophagus, and duodenum together; therefore, only some of the trials are presented in Table 1. AI-EGD increased the neoplasm-detection rate at both per-patient (risk ratio [RR], 1.57; 95% confidence interval [CI], 1.23–2.01) and per-lesion (RR, 1.55; 95% CI, 1.33–2.18) levels. Within the stomach, detection improved for LGIN (RR, 1.74; 95% CI, 1.09–2.78), HGIN (RR, 1.75; 95% CI, 1.44–2.11), and carcinoma (RR, 1.66; 95% CI, 1.37–2.02). The benefit was concentrated in small lesions: detection of lesions smaller than 10 mm more than doubled (RR, 2.24; 95% CI, 1.72–2.91), while no significant difference was observed for lesions larger than 10 mm (RR, 1.10; 95% CI, 0.84–1.43). However, these size-stratified estimates were derived from pooled data for the stomach, esophagus, and duodenum. In analyses based on anatomical region, the effect was the greatest in the cardia and fundus (RR, 2.93; 95% CI, 2.25–3.83), followed by the gastric body (RR, 2.20; 95% CI, 1.04–4.68) and the angulus (RR, 1.26; 95% CI, 1.03–1.53), whereas detection in the antrum and pylorus did not differ significantly between AI-EGD and Co-EGD. Blind spots were fewer with AI assistance, while inspection time, biopsy rate, and adverse events were comparable. These findings were consistent with the usual explanation for the effectiveness of AI-EGD: early dysplastic lesions are often flat or slightly depressed with indistinct margins, and they are most easily overlooked when they are small or located in regions that are difficult to inspect.
Nevertheless, these pooled figures should be interpreted in the context of the evidence from which they were derived. Ten of the 11 trials were conducted in China and one in Japan; all 11 trials showed at least some concerns regarding risk-of-bias assessment; and most of the trials employed systems belonging to the ENDOANGEL/WISENSE family. Therefore, the summary estimates approximated the performance of platforms of a single lineage in a single epidemiological setting rather than providing an average performance across AI systems. Because so much of the available evidence rests on this single lineage, this factor warrants specific mention. ENDOANGEL, which was developed by the group of Professor Honggang Yu at Wuhan University, was initially introduced as WISENSE and validated for monitoring gastric blind spots and reducing their occurrence in randomized trials. It has since evolved into an advanced system capable of detecting focal gastric lesions and predicting neoplasia in real time [11,25-33]. In a single-center tandem RCT, the miss rate for gastric neoplasms fell from 27.3% with routine WLI to 6.1% with ENDOANGEL-LD assistance (RR, 0.224; 95% CI, 0.068–0.744), which is the largest single effect reported for AI in the UGI tract to date [29]. However, these findings have not held up at scale. In a 24-center trial of 29514 patients led by the same group, ENDOANGEL-GN did not improve the gastric neoplasm-detection rate once the pathology was centrally reviewed (1.42% vs. 1.25%; RR, 1.13; 95% CI, 0.92–1.38). Notably, central pathology review downgraded 83.58% (804/962) of lesions originally diagnosed as LGIN and 14.00% (7/50) of those diagnosed as HGIN to benign lesions (i.e., nonneoplastic lesions such as IM, atrophy, inflammation, polyps, and xanthelasma), whereas only 0.72% (6/837) of the non-neoplastic lesions were upgraded to neoplasms. Any benefit was confined to endoscopists with less than three years of experience and to procedures performed during fatigue periods [34].
A clinically distinct question is whether AI can distinguish dysplasia from cancer or other benign conditions once a lesion has been found, since this determines whether a lesion is biopsied, resected, or staged. Gong et al. [20] already discussed this issue in the context of IM, and they addressed this by constructing a single model evaluating the entire spectrum of gastric carcinogenesis, so that dysplasia was classified alongside normal mucosa, atrophy, IM, EGC, and advanced cancer rather than in isolation. Within the same six-category classification, whose internal- and external-test accuracies are noted above, the per-class AUC for identification of dysplasia was 93.4%. Du et al. [32] fused white-light and weak-magnifying images at the feature level in ENDOANGEL-MM, which diagnosed gastric neoplasms with an accuracy of 93.6% in prospectively enrolled patients. A more clinically relevant finding is that ENDOANGEL-MM increased the accuracy of non-expert endoscopists from 70.8% to 85.8%, a level no longer statistically distinguishable from that of experts. In Korea, Chang et al. [35] validated a CADe/CADx system on several test settings. In still images from patients referred from community clinics who showed an initial biopsy finding of atypia, the system separated EGC, dysplasia, and benign lesions with accuracies of 82.5%, 88.3%, and 83.1%, respectively. In prospectively collected videos from patients undergoing endoscopic resection, this system identified EGC with a sensitivity of 91.5% against 41.5% for the initial forceps biopsy. In an external validation set of prospectively collected videos from two institutions (n=296), this system could distinguish EGC, dysplasia, and benign lesions with accuracies of 87.5%, 90.5%, and 88.9%, respectively.
The translation of these improved detection and characterization rates into clinical benefits depends on the endoscopic practices. In a two-center real-world study, the same computer-aided detection (CAD) device produced qualitatively different effects on the basis of the biopsy policy. In a hospital where the biopsy rate exceeded 95%, the overall gastric neoplasm-detection rate remained unchanged after the use of CAD devices (1.39% vs. 1.36%, p=0.897), but the proportion of detected neoplasms that were EGC nearly doubled (25.4% to 43.0%, p=0.006); in a hospital with biopsy rate below 20%, the overall gastric neoplasm-detection rate increased from 1.78% to 3.23% and LGIN detection nearly tripled (0.54% to 1.43%, p<0.001) [36]. Thus, AI compensated for the weakness of each strategy: where biopsies were already obtained from nearly everyone, it helped direct them to the lesions that mattered; where biopsies were rarely obtained, it prompted endoscopists to sample lesions they would otherwise have passed over. However, the actual benefits of the additional identification of LGIN in the latter setting are less obvious. LGIN may regress or persist rather than progress, and each detected lesion commits the patient to additional costs and surveillance.
Despite these advances, evidence for the clinical effectiveness of AI for gastric lesion detection from RCTs remains limited, with only a small number of studies available, all of which were conducted in China (Table 1) [28,29,34,37-39]. The Japan Gastroenterological Endoscopy Society (JGES) position statement on AI in gastrointestinal endoscopy similarly concluded that current evidence is insufficient to determine whether CADe or CADx for UGI endoscopy improve tumor-detection rates or the accuracy of optical diagnosis [40].
CONCLUSIONS AND FUTURE DIRECTIONS
The available evidence indicates that the use of AI in UGI endoscopy can yield diagnostic performance comparable to that of expert endoscopists and consistently superior to that of non-experts. Therefore, AI assistance is likely to serve as a valuable adjunct for most practicing endoscopists. However, the real-world clinical impact of AI cannot be fully captured by accuracy metrics alone; additional evaluations using clinically meaningful endpoints, including improvements in detection rates, reductions in miss rates, avoidance of unnecessary biopsy and surveillance, and changes in patient outcomes, are essential. Many of these outcomes will require validation through well-designed clinical trials as well as real-world studies, because the findings derived from highly controlled trial settings often do not fully translate into routine clinical practice [41].
For successful integration of AI into clinical practice, several key considerations have to be addressed. First, the human–AI interface is a critical determinant of clinical utility. Reliable real-time performance without latency or interruption, seamless alert delivery that does not disrupt the examination flow, and intuitive outputs that facilitate rapid clinical interpretation are all essential considerations beyond the diagnostic accuracy of the underlying algorithm. Second, the performance of AI is inherently context-dependent. Thus, variations in disease prevalence, endoscopic systems, image quality, operator behavior, and histopathological practices across institutions can substantially influence model performance, limiting the generalizability of the models. Even identical AI systems may produce different clinical outcomes across hospitals [34,36]. Overcoming this limitation will require multicenter RCTs, robust external validation, and large-scale real-world data, and potentially necessitate site-specific model adaptation. Third, the highly specialized nature of AI systems poses a challenge to clinical implementation. Real-world practice requires integrated, multitask platforms rather than a “one AI per disease” approach. Thus, the parallel use of multiple independent AI systems is neither practical nor scalable. Future efforts should focus on the development of unified architectures that also address the risk of performance trade-offs associated with multitask learning. Finally, the legal and ethical challenges associated with the clinical adoption of AI remain significant. These include accountability for false-positive and false-negative outputs as well as patient privacy and data security.
Nevertheless, the potential for AI to substantially reshape clinical practice in the coming years has become increasingly clear. As performance and user interfaces improve, and as evidence from multicenter RCTs and real-world data continues to accumulate, the impact of AI will become more apparent. Rather than relying uncritically on RCT findings, clinicians should cultivate an objective approach to evaluating the effectiveness of AI within their own institutional environment, as well as the ability to identify, among the rapidly expanding range of AI technologies, those that provide meaningful benefit in their specific practice settings. Simultaneously, clinicians should maintain their own endoscopic competence and clinical judgment, which will allow them to perform thorough examinations without AI assistance and to recognize the limitations of AI and critically appraise and correct erroneous outputs.
PDF Links
PubReader
ePub Link
XML Download
Full text via DOI
Download Citation
