Introduction
Upper gastrointestinal (UGI) cancers, principally gastric and esophageal cancers, collectively impose a substantial global health burden.1,2 Despite overall improvements in treatment modalities, the 5-year survival rate for advanced-stage gastric and esophageal cancers remains below 30%, largely because the majority of patients are diagnosed at late stages when curative resection is no longer feasible.3 In contrast, when detected at an early stage, the 5-year survival rate can exceed 90%, underscoring the importance of early detection strategies.4
Both gastric and esophageal cancers develop through well-defined precancerous cascades that offer critical windows for intervention. Gastric carcinogenesis follows the Correa cascade: chronic Helicobacter pylori gastritis progresses to chronic atrophic gastritis (CAG), gastric intestinal metaplasia (GIM), dysplasia, and finally adenocarcinoma. Patients with CAG or GIM carry an annual gastric cancer risk of 0.1–0.25%, rising to > 6% for high-grade dysplasia.5,6 In the esophagus, Barrett’s esophagus, a form of intestinal metaplasia driven by chronic reflux, confers a 30–60-fold increased risk of esophageal adenocarcinoma.7 Esophageal squamous intraepithelial neoplasia (ESIN) represents the precursor spectrum for esophageal squamous cell carcinoma, with annual progression rates of approximately 1% for low- to moderate-grade dysplasia.8 Despite the potential for curative intervention, endoscopic recognition of these lesions remains suboptimal.
Esophagogastroduodenoscopy with biopsy is the standard approach to surveillance, yet conventional white-light endoscopy (WLE) has significant limitations. Detection of CAG and GIM relies on subtle mucosal changes (pallor, loss of rugal folds, irregular surface pattern) that are easily overlooked, with reported sensitivities of only 50–70% and substantial interobserver variability.9 The Prague C & M criteria provide a standardized endoscopic grading system for Barrett’s esophagus,10 but dysplastic lesions may still be subtle against a heterogeneous background. ESIN is even more challenging to detect because it frequently lacks visible abnormalities and requires Lugol’s chromoendoscopy, which is time-consuming, poorly tolerated, and limited in availability.11 Consequently, a substantial proportion of precancerous lesions are missed during routine esophagogastroduodenoscopy, which may contribute to delayed diagnosis and preventable cancer progression.
Artificial intelligence (AI), particularly deep learning using convolutional neural networks (CNNs) and vision transformers, can support real-time identification of subtle mucosal abnormalities and may improve consistency relative to unaided human interpretation.12 Retrospective studies report areas under the receiver operating characteristic curves (AUCs) > 0.90 for early gastric cancer, GIM, CAG, and Barrett’s neoplasia.13 Despite advances from CNNs to transformer architectures and real-time video integration,14,15 the comparative performance and generalizability of AI across UGI precancerous lesions remain incompletely characterized.
Prior meta-analyses have evaluated AI for individual lesion types (e.g., early gastric cancer or Barrett’s neoplasia), but evidence comparing diagnostic performance across different lesion categories (ESIN, Barrett’s dysplasia, CAG, GIM, and gastric dysplasia), AI model architectures (CNN versus transformer), and imaging modalities (WLE versus image-enhanced endoscopy) remains limited. To address these gaps, we conducted a systematic review and meta-analysis to evaluate pooled diagnostic accuracy, perform subgroup analyses by lesion type, architecture, and modality, and assess heterogeneity and publication bias. Our findings provide an evidence base to guide clinical integration of AI-assisted endoscopic surveillance.
Given that precancerous lesions are actionable targets of organized cancer-screening programs, such a synthesis is directly relevant to setting risk-stratified screening intervals and allocating resources for cancer prevention.
Materials and methods
This systematic review and meta-analysis was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines (the completed PRISMA checklist is provided in Supplementary File 1).16 The study protocol was prospectively registered with PROSPERO (registration number: CRD420261349387).
Search strategy
Two authors independently searched electronic databases (PubMed, Ovid MEDLINE(R) ALL, Embase, Cochrane Library, and IEEE Xplore) from inception to March 31, 2026. We used the following keywords and corresponding Medical Subject Headings (MeSH) terms: “upper gastrointestinal tract”, “precancerous conditions”, “Barrett’s esophagus”, “intestinal metaplasia”, “atrophic gastritis”, “dysplasia”, “artificial intelligence”, “deep learning”, “convolutional neural network”, “computer-aided diagnosis”, “endoscopy”, “gastroscopy”, “sensitivity”, “specificity”, and “accuracy”. The full search strategy is shown in Supplementary File 2. The search was limited to English-language publications, and the references of the included studies and related reviews were manually searched to identify additional eligible studies. Two authors independently screened all retrieved records in two sequential stages. In the first stage, titles and abstracts were screened against the predefined eligibility criteria to exclude obviously irrelevant studies. In the second stage, full texts of potentially eligible studies were retrieved and assessed in detail. Any discrepancies between the two authors at either stage were discussed and, if consensus could not be reached, resolved by a third author.
Inclusion and exclusion criteria
Inclusion criteria
Studies were eligible if they: (1) evaluated the diagnostic performance of an AI model for UGI precancerous lesions using endoscopic images or videos as input data; (2) targeted lesions of interest including esophageal squamous intraepithelial neoplasia (ESIN), Barrett’s esophagus with dysplasia, CAG, GIM, or gastric intraepithelial dysplasia; (3) reported sensitivity and specificity, or provided sufficient raw data (true positive (TP), false positive (FP), true negative (TN), false negative (FN)) to calculate these metrics; (4) used histopathological examination as the reference standard, except for one intestinal metaplasia dataset that used a validated endoscopic visual criterion with previously demonstrated concordance with histopathology; and (5) were published in peer-reviewed journals with full text available.
Exclusion criteria
Studies were excluded if they: (1) did not clearly describe the AI model type or failed to report diagnostic performance indicators; (2) focused exclusively on advanced-stage cancers or non-precancerous gastrointestinal conditions; (3) were non-original research, including case reports, narrative reviews, systematic reviews, editorials, or conference abstracts; or (4) used non-endoscopic imaging modalities (e.g., radiological or pathological whole-slide images) as primary input data.
Data extraction
Two reviewers independently extracted data using a prespecified standardized form; discrepancies were resolved through discussion or adjudication by a third reviewer. The following data were extracted: (1) study characteristics: first author, publication year, country, study design (prospective or retrospective), and total sample size; (2) lesion and imaging characteristics: lesion type (ESIN, Barrett’s dysplasia, CAG, GIM, or gastric dysplasia), endoscopic modality (WLE, narrow-band imaging (NBI), magnifying narrow-band imaging (M-NBI), blue laser imaging (BLI), linked color imaging (LCI), or combinations thereof), and whether an endoscopist comparator group was included; (3) AI model characteristics: model architecture (e.g., CNN, deep convolutional neural network, transformer, or hybrid model) and whether the study reported internal or external validation; and (4) diagnostic performance data: TP, FP, TN, FN, sensitivity, specificity, accuracy, and AUC. When data were missing or ambiguous, corresponding authors were contacted for clarification.
Quality assessment
The risk of bias and applicability of included studies were assessed using the Quality Assessment of Diagnostic Accuracy Studies-2 (QUADAS-2).17 Risk of bias was assessed across four domains. For patient selection, we assessed whether participants were enrolled consecutively and whether selection bias was present. For the index test, we assessed whether AI model implementation and validation were clearly described. For the reference standard, we assessed whether the reference standard was appropriate for classifying the target condition. For Gong E 2024, dysplasia was confirmed histopathologically, while intestinal metaplasia was assessed using a validated endoscopic visual grading criterion; because the developers of this criterion had previously demonstrated its concordance with histopathology in an independent validation study, this dataset was retained and its reference standard rated as low risk. For flow and timing, we assessed whether the interval between the index test and reference standard was appropriate. Each domain was rated as having a low, high, or unclear risk of bias.
Statistical analysis
When a single study reported diagnostic performance separately for multiple lesion types or imaging modalities, each subgroup was treated as an independent dataset for pooled analysis; accordingly, the 15 included studies contributed a total of 24 datasets.Diagnostic performance data (TP, FP, TN, FN) were pooled using a bivariate random-effects model to simultaneously estimate summary sensitivity and specificity with 95% confidence intervals (CIs). Summary receiver operating characteristic (SROC) curves were constructed and AUC was calculated to reflect overall diagnostic accuracy. All pooled analyses and figure generation were performed using Python (version 3.12) and Meta-DiSc 2.0.18,19 Pooled sensitivity and specificity were estimated using a DerSimonian–Laird random-effects model applied to logit-transformed proportions20; a continuity correction of 0.5 was applied to studies with zero cells.21 SROC curves and the bivariate random-effects model were implemented according to the method described by Reitsma et al.22 Leave-one-out sensitivity analyses were conducted at both the dataset and study levels following the approach of Viechtbauer and Cheung to assess the influence of individual datasets or studies on the pooled estimates.23 Datasets from the same study were also aggregated into study-level records and re-pooled as a sensitivity analysis to examine within-study non-independence. Publication bias was assessed using Egger’s regression test, Begg’s rank correlation test, Deeks’ funnel plot asymmetry test, and the Duval and Tweedie trim-and-fill method; a two-sided P < 0.05 was considered statistically significant.24 Subgroup analyses and meta-regression were performed to explore sources of heterogeneity. For subgroup visualization, Figure 1b–f were fitted using the Moses–Littenberg method; the summary points and 95% CIs are the subgroup-specific pooled estimates reported in Table 1. Bivariate meta-regression was performed only for target organ, imaging modality, and study design; analyses by AI architecture and target lesion category were descriptive.
| Subgroup | n | Sensitivity (95% CI) | Specificity (95% CI) | DOR (95% CI) | LR+ (95% CI) | LR− (95% CI) | FPR (95% CI) | Relative sensitivity (95% CI) | P (sensitivity) | Relative specificity (95% CI) | P (specificity) | Global test P |
|---|
| Target organ | | | | | | | | | | | | |
| Esophagus | 5 | 0.91 (0.745–0.972) | 0.978 (0.919–0.994) | 442.64 (55.17–3551.11) | 40.963 (10.462–160.386) | 0.093 (0.03–0.287) | 0.022 (0.006–0.081) | 0.954 (0.907–1.004) | 0.127 | 1.026 (0.912–1.155) | 0.636 | 0.195 |
| Stomach | 19 | 0.933 (0.886–0.962) | 0.933 (0.884–0.962) | 195.24 (73.13–521.23) | 13.936 (7.82–24.834) | 0.071 (0.04–0.126) | 0.067 (0.038–0.116) | Reference | — | Reference | — | 0.195 |
| Imaging modality | | | | | | | | | | | | |
| WLE | 12 | 0.875 (0.781–0.933) | 0.958 (0.913–0.98) | 160.28 (46.43–553.32) | 20.857 (9.499–45.799) | 0.13 (0.071–0.239) | 0.042 (0.02–0.087) | 0.911 (0.834–0.996) | 0.017 | 1.035 (0.969–1.105) | 0.285 | 0.005 |
| Others (IEE/NBI/ME) | 12 | 0.96 (0.926–0.979) | 0.926 (0.853–0.964) | 302.59 (88.66–1032.79) | 12.955 (6.28–26.727) | 0.043 (0.022–0.083) | 0.074 (0.036–0.147) | Reference | — | Reference | — | 0.005 |
| Study design | | | | | | | | | | | | |
| Prospective | 8 | 0.872 (0.738–0.943) | 0.941 (0.858–0.977) | 109.54 (24.35–492.81) | 14.873 (5.697–38.831) | 0.136 (0.062–0.299) | 0.059 (0.023–0.142) | Reference | — | Reference | — | 0.22 |
| Retrospective | 16 | 0.947 (0.907–0.971) | 0.946 (0.898–0.972) | 313.02 (109.61–893.89) | 17.443 (9.035–33.675) | 0.056 (0.031–0.1) | 0.054 (0.028–0.102) | 1.086 (0.966–1.221) | 0.09 | 1.005 (0.938–1.076) | 0.894 | 0.22 |
| AI architecture | | | | | | | | | | | | |
| CNN | 18 | 0.945 (0.923–0.962) | 0.945 (0.897–0.971) | 295.29 (104.078–837.8) | 17.062 (8.984–32.949) | 0.058 (0.039–0.086) | 0.055 (0.029–0.103) | Not performed | — | Not performed | — | Not performed |
| Transformer | 6 | 0.792 (0.408–0.954) | 0.902 (0.816–0.950) | 34.99 (3.058–400.268) | 8.074 (2.218–19.169) | 0.231 (0.048–0.725) | 0.098 (0.050–0.184) | Not performed | — | Not performed | — | Not performed |
| Target lesion | | | | | | | | | | | | |
| IM | 10 | 0.944 (0.864–0.978) | 0.911 (0.805–0.962) | 171.56 (32.30–911.19) | 10.633 (4.442–25.870) | 0.062 (0.023–0.169) | 0.089 (0.038–0.195) | Not performed | — | Not performed | — | Not performed |
| Dysplasia | 6 | 0.871 (0.796–0.921) | 0.981 (0.912–0.996) | 336.72 (81.23–1395.80) | 45.647 (9.029–236.298) | 0.132 (0.079–0.224) | 0.019 (0.004–0.088) | Not performed | — | Not performed | — | Not performed |
| GA | 6 | 0.910 (0.834–0.953) | 0.913 (0.766–0.971) | 103.21 (14.98–711.22) | 10.430 (3.568–32.764) | 0.098 (0.048–0.216) | 0.087 (0.029–0.234) | Not performed | — | Not performed | — | Not performed |
| Precancerous | 2 | | | | | | | Not performed | — | Not performed | — | Not performed |
Heterogeneity across studies was quantified using the I² statistic and the χ² test; I² > 50% was considered indicative of substantial heterogeneity. To explore potential sources of heterogeneity, prespecified subgroup analyses were conducted according to: (1) target organ (esophageal versus gastric precancerous lesions); (2) endoscopic imaging modality (WLE versus image-enhanced endoscopy [IEE], including NBI, M-NBI, BLI, and LCI); (3) study design (retrospective versus prospective); (4) AI model architecture (CNN versus transformer-based models); and (5) target lesion category (intestinal metaplasia, dysplasia, or gastric atrophy, versus precancerous change not further subtyped). Meta-regression was performed for each subgroup variable when the number of datasets per stratum was sufficient.
Results
General information and baseline characteristics of the included studies
The PRISMA 2020 flow diagram (Fig. 2) was generated using the PRISMA2020 R package.25 A total of 15 studies were included (Fig. 2).26-40 These studies comprised 24 datasets (as several studies reported results separately for multiple lesion types or imaging modalities) and collectively encompassed 19,578 diagnostic records (Fig. 3a), including image-, video-, and patient-level data, of which 8,336 (42.6%) were positive records and 11,242 (57.4%) were negative records. Publication years ranged from 2020 to 2026, three studies each were published in 2020 (Liu G, Liu X, de Groof AJ), 2021 (Mu G, Lin N, Xu M), 2022 (Wong PK, Gong E [esophageal], Siripoppohn V); four studies were published in 2024 (Tan JL, Li Z, Gong E [gastric], Jhang JY); one study was published in 2023 (Yang J), and one in 2026 (Chiang TH). Studies originated predominantly from China (n = 10), with the remainder from Korea (n = 2), Australia (n = 1), Thailand (n = 1), and the Netherlands (n = 1). Eleven studies addressed gastric precancerous lesions and four targeted esophageal precancerous lesions. Ten studies were retrospective, three were prospective, and two included both retrospective and prospective datasets. At the dataset level, which is the unit used in Table 1, 8 of 24 datasets were prospective and 16 were retrospective, reflecting that Xu M 2021 and Siripoppohn V 2022 each contributed both a prospective and a retrospective dataset. Endoscopic imaging modalities were predominantly WLE (n = 12), with the remainder employing image-enhanced endoscopy techniques. AI architectures ranged from CNN-based models (ResNet-50, TResNet, Two-stream CNN, BiSeNet) to transformer-based and hybrid architectures (BLS²-MSA, GSCNet). Among the 24 datasets, 18 used CNN-based architectures and 6 used transformer-based architectures (Table 1),26-40 reflecting the diversity of deep learning approaches applied in this field. Detailed baseline characteristics of all included studies are presented in Table 2.
| Study | Year | Country | Study design | Target organ | Target lesion | Imaging type | AI model | Reference standard | Test sample sized |
|---|
| Liu G26 | 2020 | China | Retrospective | Esophagus | Dysplasia | WLE | Two-stream CNN | Pathology | 129 images |
| Wong PK27 | 2022 | China | Retrospective | Stomach | IM | M-NBI | BLS²-MSA | Pathology | 146 images |
| Gong E28 | 2022 | Korea | Prospective | Esophagus | Dysplasia | WLE | Neuro-T | Pathology | 836 images |
| Mu G29 | 2021 | China | Retrospective | Stomach | Gastric atrophy | WLE | UNet++ + ResNet-50 | Pathology | 258 images / 80 videos |
| Liu X30 | 2020 | China | Retrospective | Stomach | Precancerous (LGN) | M-NBI | ResNet-50 | Pathology | 3,871 images |
| Jhang JY31 | 2024 | China | Retrospective | Stomach | Precancerous (CGI) | WLE | GSCNet | Pathology | 912 images (304 patients) |
| Tan JL32 | 2024 | Australia | Retrospective | Esophagus | Dysplasia | WLE / NBI | ViT / DeiT | Pathology | 146 images (total) |
| Siripoppohn V33a | 2022 | Thailand | Prospective/retrospective | Stomach | IM | WLE / NBI | BiSeNet | Pathology | 320 images (total) |
| Li Z34 | 2024 | China | Retrospective | Stomach | IM | IEE (NBI) | Deformable transformer | Pathology | 587 images |
| Chiang TH35 | 2026 | China | Prospective | Stomach | IM / GA | WLE | Vision transformer | Pathology | 134 patients |
| Gong E36b | 2024 | Korea | Prospective | Stomach | Dysplasia / IM | WLE | Neuro-X (CNN) / Edge AI | Pathology / validated visual criterion for IM | 718 images |
| Lin N37 | 2021 | China | Retrospective | Stomach | IM / GA | WLE | TResNet | Pathology | 548 images (273 patients) |
| Yang J38 | 2023 | China | Retrospective | Stomach | IM / GA | WLI + LCI | LAG | Pathology | 4,284 images |
| Xu M39c | 2021 | China | Prospective/retrospective | Stomach | IM / GA | ME-NBI/BLI | CNN | Pathology | 1,052 images ×2 (IM/GA) / 98 videos ×2 (IM/GA) |
| de Groof AJ40 | 2020 | Netherlands | Retrospective | Esophagus | Neoplasia / dysplasia | WLE + NBI | CNN | Pathology | 80 images |
QUADAS-2 quality assessment
Histopathological examination was used as the reference standard in the included studies, except that intestinal metaplasia in Gong E (2024) was assessed using a validated endoscopic visual criterion previously shown to be concordant with histopathology (see Table 2 footnote) (Table 3).26-40 The studies encompassed a variety of endoscopic imaging modalities (WLE, NBI, M-NBI, magnifying endoscopy with blue laser imaging, and LCI) and AI architectures (CNN-based, transformer-based, and hybrid models), reflecting the current diversity of deep learning approaches in UGI endoscopy research. Methodological quality was evaluated using the QUADAS-2 tool, which assesses both risk of bias and applicability concerns across four domains: patient selection, index test, reference standard, and flow and timing.
| Study | Year | Risk of bias | Applicability |
|---|
| Patient selection | Index test | Reference standard | Flow and timing | Patient selection | Index test | Reference standard |
|---|
| Liu G26 | 2020 | High | Low | Low | Low | High | Low | Low |
| Wong PK27 | 2022 | High | Low | Low | Low | High | Low | Low |
| Gong E28 | 2022 | Low | Low | Low | Low | Low | Low | Low |
| Mu G29 | 2021 | Unclear | Low | Low | Low | Unclear | Low | Low |
| Liu X30 | 2020 | High | Low | Low | Low | High | Low | Low |
| Jhang JY31 | 2024 | High | Low | Low | Low | High | Low | Low |
| Tan JL32 | 2024 | High | Low | Low | Low | High | Low | Low |
| Siripoppohn V33 | 2022 | Low | Low | Low | Low | Low | Low | Low |
| Li Z34 | 2024 | High | Low | Low | Low | High | Low | Low |
| Chiang TH35 | 2026 | Low | Low | Low | Low | Low | Low | Low |
| Gong E36 | 2024 | Low | Low | Low | Low | Low | Low | Low |
| Lin N37 | 2021 | High | Low | Low | Low | High | Low | Low |
| Yang J38 | 2023 | High | Low | Low | Low | High | Low | Low |
| Xu M39 | 2021 | Low | Low | Low | Low | Low | Low | Low |
| de Groof AJ40 | 2020 | High | Low | Low | Low | High | Low | Low |
Risk of bias
Patient selection: Of the 15 studies, 5 (33.3%) were classified as low risk, 9 (60.0%) as high risk, and 1 (6.7%) as unclear risk. High-risk ratings in this domain predominantly stemmed from non-consecutive or retrospective image sampling, which may introduce selection bias and potentially overestimate diagnostic accuracy. Index test: All 15 studies (100%) were rated as low risk, reflecting adequate description of AI model implementation and validation procedures. Reference standard: All 15 studies (100%) were rated as low risk; histopathology was used for the target condition in all datasets except the Gong E 2024 intestinal metaplasia dataset, which used the validated visual criterion described above. Flow and timing: All 15 studies (100%) were rated as low risk, indicating appropriate intervals between the index test and reference standard across all included studies.
Applicability concerns
Patient selection: Five studies (33.3%) demonstrated low applicability concerns, 9 (60.0%) had high applicability concerns, and 1 (6.7%) was rated unclear, mirroring the risk of bias pattern and reflecting the predominance of retrospective, single-center designs that may limit generalizability to routine clinical practice. Index test: All 15 studies (100%) had low applicability concerns for the index test, indicating that the AI models and imaging protocols were considered applicable to the review question. Reference standard: All 15 studies (100%) were at low applicability concern; histopathology was used in all datasets except the Gong E 2024 intestinal metaplasia dataset, for which the validated visual criterion described above was used.
Results of univariate model analysis
Overall diagnostic performance
Based on the univariate random-effects model, the pooled sensitivity of the AI model in the diagnosis of upper gastrointestinal precancerous lesions was 0.928 (95% CI: 0.88–0.96), the specificity was 0.94 (95% CI: 0.91–0.97), and the diagnostic odds ratio (DOR) was 220.47 (95% CI: 100.86–481.89) (Table 4). The positive likelihood ratio (LR+) was 16.75 (95% CI: 9.78–28.68), and the negative likelihood ratio was 0.076 (95% CI: 0.046–0.125), indicating high discriminatory ability of AI-assisted endoscopy in the detection of upper gastrointestinal precancerous lesions.
| Index | Estimate | 95% LCI | 95% UCI |
|---|
| Univariate random-effects model | | | |
| Sensitivity | 0.928 | 0.883 | 0.957 |
| Specificity | 0.945 | 0.906 | 0.968 |
| DOR | 220.47 | 100.86 | 481.89 |
| LR+ | 16.75 | 9.78 | 28.68 |
| LR− | 0.076 | 0.046 | 0.125 |
| FPR | 0.055 | 0.032 | 0.094 |
| Bivariate random-effects model | | | |
| Sensitivity | 0.930 | 0.886 | 0.958 |
| Specificity | 0.944 | 0.906 | 0.967 |
| DOR | 223.79 | 92.24 | 542.94 |
| LR+ | 16.66 | 9.69 | 28.65 |
| LR− | 0.074 | 0.045 | 0.124 |
| FPR | 0.056 | 0.033 | 0.094 |
Sensitivity and specificity analysis
The heterogeneity in sensitivity was 91.6%, indicating extremely high heterogeneity among the included studies (P < 0.01). Sensitivity across datasets ranged from 0.30 to 0.99. Notably, studies targeting intestinal metaplasia (IM) with WLE tended to report lower sensitivity estimates (e.g., Chiang TH 2026 IM subgroup, sensitivity 0.30, 95% CI: 0.15–0.49), which may reflect the inherently subtle endoscopic appearance of IM under standard imaging.
Heterogeneity in specificity was equally pronounced, with an I² of 96.7% and specificities ranging from 0.73 to 1.00 across individual datasets. In the specificity forest plot (Fig. 3c), the summary estimate was 0.94 (95% CI: 0.91–0.97). At the subgroup level, specificity varied by target lesion category: dysplasia datasets achieved the highest pooled specificity (0.981), whereas intestinal metaplasia and gastric atrophy datasets were lower and more variable (0.911 and 0.913, respectively; Table 1). Accordingly, in the meta-regression the significant association between imaging modality and performance was confined to sensitivity (0.960 vs. 0.875; P = 0.017), with no corresponding difference in specificity (relative specificity 1.035; P = 0.285). Figure 3d presents the overall SROC curve based on the bivariate model (the same curve shown in Fig. 1a), with a summary AUC of 0.9661, indicating excellent overall discriminative ability (Fig. 3d).
Results of bivariate model analysis
Based on the bivariate random-effects model, the summary sensitivity and specificity of the AI model in the diagnosis of upper gastrointestinal precancerous lesions were 0.93 (95% CI: 0.886–0.958) and 0.944 (95% CI: 0.906–0.967), respectively, and the DOR was 223.79 (95% CI: 92.24–542.94), indicating that AI-assisted endoscopy has a high discriminatory ability between precancerous and non-precancerous lesions. The LR+ was 16.66 (95% CI: 9.69–28.65), indicating that a positive AI result substantially increases the post-test probability of a precancerous lesion; the negative likelihood ratio was 0.074 (95% CI: 0.045–0.124), indicating that a negative AI result substantially lowers the post-test probability of precancerous lesions. The false-positive rate was 0.056 (95% CI: 0.033–0.094) (Table 4).
The bivariate I² was 0.73 (73%), indicating substantial heterogeneity across the included datasets (P < 0.01). The sensitivity forest plot demonstrated considerable variability across studies, with individual sensitivities ranging from 0.30 to 0.99 (Fig. 4a), while specificities ranged from 0.72 to 1.00 (Fig. 4b). Studies employing image-enhanced endoscopy modalities such as magnifying narrow-band imaging and magnifying endoscopy with blue laser imaging tended to show higher sensitivity, whereas studies focusing on intestinal metaplasia under white-light endoscopy showed comparatively lower sensitivity, likely reflecting the subtle endoscopic appearance of IM. The ROC plane scatter plot (Fig. 4c) demonstrated that the majority of studies clustered in the upper-left region, consistent with high overall diagnostic accuracy. The 95% prediction ellipse was broad, reflecting the expected range of diagnostic performance in future comparable studies.
ROC and model consistency
The SROC curve showed that the summary AUC was 0.9661, indicating excellent overall discriminative ability (Fig. 1a). The majority of individual study points clustered in the upper-left region of the ROC plane, indicating the coexistence of high sensitivity and high specificity across most included datasets. The 95% confidence ellipse was narrow and centered near the summary point, reflecting reliable estimation of the pooled diagnostic accuracy. However, the 95% prediction ellipse was substantially wider, extending toward the lower-right region of the ROC plane, suggesting considerable variability in the expected diagnostic performance of AI models in future comparable studies. Subgroup-stratified Moses–Littenberg SROC plots (Fig. 1b–f; summary points and confidence intervals from Table 1) further illustrated the sources of this heterogeneity. When stratified by target organ (Fig. 1b), esophageal studies (purple summary point) and gastric studies (orange summary point) showed distinct distributional patterns, with esophageal studies tending to cluster closer to the upper-left corner, potentially reflecting the more distinct endoscopic features of esophageal dysplasia compared with gastric precancerous lesions. Stratification by imaging modality (Fig. 1c) revealed that studies using WLE (navy circles) demonstrated a wider spread of diagnostic performance compared with those employing other modalities (teal squares), consistent with the known limitations of WLE in detecting subtle mucosal changes. Stratification by study design (Fig. 1d) showed that prospective studies (triangles) and retrospective studies (circles) were similarly distributed across the ROC plane, suggesting that study design alone did not fully account for the observed heterogeneity. Stratification by AI model architecture (Fig. 1e) showed that CNN-based models (blue circle) achieved a tightly clustered summary point close to the upper-left corner with a comparatively narrow prediction ellipse, whereas transformer-based models (red diamond) showed a summary point shifted toward lower sensitivity and a markedly wider prediction ellipse extending toward the lower-right region of the ROC plane, reflecting the smaller number and greater architectural/lesion-type heterogeneity of the transformer-based datasets. Stratification by target lesion category (Fig. 1f) showed that dysplasia (red circle) achieved the summary point closest to the upper-left corner, consistent with its highest specificity among the pooled lesion categories, while intestinal metaplasia (blue square) and gastric atrophy (green triangle) showed broadly overlapping distributions with wider, similarly shaped prediction ellipses, indicating that these two lesion types were not clearly separated in overall diagnostic performance.
Subgroup analysis results
Subgroup analyses were performed according to target organ, imaging modality, study design, AI model architecture, and target lesion category, with results summarized in Table 1 and illustrated by Moses–Littenberg SROC curves in Figure 1b–f.
By target organ, esophageal studies (n = 5 datasets) showed numerically higher specificity than gastric studies (0.978, 95% CI: 0.919–0.994 vs. 0.933, 95% CI: 0.884–0.962), alongside a substantially higher DOR (442.642, 95% CI: 55.175–3551.110 vs. 195.238, 95% CI: 73.130–521.231) and LR+ (40.963 vs. 13.936). However, meta-regression analysis revealed no statistically significant difference in sensitivity (relative sensitivity: 0.954, 95% CI: 0.907–1.004, P = 0.127) or specificity (relative specificity: 1.026, 95% CI: 0.912–1.155, P = 0.636) between gastric and esophageal studies, and the global test was non-significant (P = 0.195), indicating that target organ did not significantly contribute to the observed heterogeneity.
By imaging modality, studies employing image-enhanced endoscopy or other advanced modalities (IEE/NBI/ME, n = 12 datasets) achieved higher sensitivity than WLE studies (0.960, 95% CI: 0.926–0.979 vs. 0.875, 95% CI: 0.781–0.933). Meta-regression confirmed that WLE was associated with significantly lower sensitivity compared with other modalities (relative sensitivity: 0.911, 95% CI: 0.834–0.996, P = 0.017). Although the difference in specificity did not reach statistical significance (relative specificity: 1.035, 95% CI: 0.969–1.105, P = 0.285), the global test demonstrated a statistically significant overall difference between modality subgroups (P = 0.005), confirming that imaging modality was a significant source of heterogeneity in the dataset.
By study design, retrospective studies (n = 16 datasets) showed numerically higher sensitivity (0.947, 95% CI: 0.907–0.971 vs. 0.872, 95% CI: 0.738–0.943) and comparable specificity (0.946 vs. 0.941) relative to prospective studies (n = 8 datasets). However, meta-regression analysis showed no significant difference in sensitivity (relative sensitivity: 1.086, 95% CI: 0.966–1.221, P = 0.09) or specificity (relative specificity: 1.005, 95% CI: 0.938–1.076, P = 0.894) between retrospective and prospective studies, with a non-significant global test (P = 0.22), suggesting that study design did not independently explain the heterogeneity observed.
By AI model architecture, CNN-based models (n = 18 datasets) achieved a pooled sensitivity of 0.945 (95% CI: 0.923–0.962) and specificity of 0.945 (95% CI: 0.897–0.971), compared with 0.792 (95% CI: 0.408–0.954) and 0.902 (95% CI: 0.816–0.950) for transformer-based models (n = 6 datasets). The markedly wider confidence interval for the transformer subgroup reflects both the small number of transformer-based datasets and their heterogeneous composition (spanning Barrett’s dysplasia, gastric intestinal metaplasia, and gastric atrophy); this comparison should therefore be interpreted as preliminary rather than a definitive architecture-level ranking.
By target lesion category, sensitivity and specificity varied across lesion types: intestinal metaplasia (IM, n = 10 datasets) showed sensitivity of 0.944 (95% CI: 0.864–0.978) and specificity of 0.911 (95% CI: 0.805–0.962); dysplasia (n = 6 datasets) showed the highest specificity at 0.981 (95% CI: 0.912–0.996) but a comparatively lower sensitivity of 0.871 (95% CI: 0.796–0.921); and gastric atrophy (n = 6 datasets) showed intermediate performance (sensitivity 0.910, 95% CI: 0.834–0.953; specificity 0.913, 95% CI: 0.766–0.971). The remaining precancerous-lesion category (Liu X 2020 and Jhang JY 2024, n = 2 datasets) was considered too small to support a stable pooled estimate and is reported descriptively only (Table 1). These lesion-specific differences indicate that the overall pooled estimate should not be treated as uniformly applicable across lesion types, and that AI performance is better characterized at the level of individual lesion categories than by a single pooled figure.
In summary, among the subgroup variables formally evaluated by meta-regression, imaging modality was significantly associated with heterogeneity in AI diagnostic performance (global P = 0.005), with WLE demonstrating significantly lower sensitivity compared with image-enhanced modalities. Neither target organ nor study design reached statistical significance in meta-regression analyses, and meta-regression was not extended to the AI-architecture or target-lesion subgroups because of the small number of datasets in several strata (e.g., transformer, n = 6; precancerous, n = 2); these subgroup comparisons are therefore descriptive rather than confirmatory.
Publication bias assessment
Publication bias was assessed using Egger’s regression test, Begg’s rank correlation test, Deeks’ funnel plot asymmetry test, and the trim-and-fill method. Egger’s test showed no statistically significant small-study effect (intercept = −0.152, standard error = 0.714, P = 0.833). Begg’s rank correlation test similarly indicated no significant asymmetry (Kendall’s τ = 0.073, P = 0.641). Deeks’ funnel plot asymmetry test, which is specifically recommended for diagnostic accuracy meta-analyses, yielded a P value of 0.053, marginally above the threshold for statistical significance. Visual inspection of the funnel plot (Fig. 5) showed an approximately symmetric distribution of study points around the pooled DOR estimate. The trim-and-fill analysis (Duval & Tweedie’s L0 estimator, applied in Python to the same random-effects log(DOR) model used in the leave-one-out analysis above) identified 2 potentially missing studies, yielding an adjusted pooled DOR of 136.10 (95% CI: 55.64–332.86), compared with an unadjusted DOR of 186.98 (95% CI: 80.55–434.03) from the same univariate log(DOR) model reported as the leave-one-out baseline above; this value is numerically distinct from, and not a substitute for, the bivariate DOR of 223.79 reported as this study’s primary effect estimate (Table 4), with the adjusted estimate remaining statistically significant. These results indicate no substantial evidence of publication bias, and the overall conclusions of this meta-analysis are unlikely to be materially affected by selective publication.
Leave-one-out sensitivity analysis
Leave-one-out sensitivity analysis was performed to evaluate the robustness of the pooled estimates, using a univariate DerSimonian–Laird random-effects model implemented in Python and applied directly to each dataset’s sensitivity, specificity, and DOR. Because this implementation differs both from the bivariate model underlying the study’s primary results and from the Meta-DiSc univariate summary reported in Table 4, its baseline point estimates (sensitivity 0.923, specificity 0.938) are close to, but not identical to, those values (e.g., the Table 4 univariate sensitivity of 0.928); this reflects only the different pooling implementation, not any difference in the underlying data. Sequential exclusion of each of the 24 datasets yielded a pooled sensitivity ranging from 0.913 to 0.934 and specificity ranging from 0.931 to 0.943, with no single dataset altering the pooled sensitivity or specificity by more than 0.02 from the overall estimate (baseline: sensitivity = 0.923, 95% CI: 0.888–0.948; specificity = 0.938, 95% CI: 0.897–0.964; DOR = 186.98, 95% CI: 80.55–434.03), confirming that our conclusions are not driven by any individual study. Exclusion of the two Chiang 2026 datasets, which reported the lowest sensitivity among included studies (0.30–0.31), resulted in only a marginal increase in pooled sensitivity (+0.009 to +0.011), while exclusion of Yang J 2023 (IM), the largest dataset by sample size, produced the largest single-dataset effect (ΔSensitivity = −0.010, ΔSpecificity = −0.006) due to its high statistical weight. Full leave-one-out results for all 24 datasets are provided in Supplementary File 3.
Discussion
The present systematic review and meta-analysis, incorporating 15 studies and 24 datasets comprising 19,578 diagnostic records, indicates that AI-assisted endoscopy has high diagnostic accuracy for upper gastrointestinal precancerous lesions, with a pooled sensitivity of 0.930 (95% CI: 0.886–0.958), specificity of 0.944 (95% CI: 0.906–0.967), DOR of 223.79, and AUC of 0.9661. The positive likelihood ratio of 16.658 and negative likelihood ratio of 0.074 indicate that AI-assisted endoscopy has meaningful discriminatory ability, substantially increasing the post-test probability when positive and lowering it when negative. Among the variables formally evaluated by meta-regression, imaging modality was associated with heterogeneity (global P = 0.005), with image-enhanced endoscopy (IEE/NBI/ME) achieving markedly higher sensitivity than white-light endoscopy (0.960 vs. 0.875). No evidence of substantial publication bias was detected across four complementary assessments. These findings support and extend prior evidence.13,41 This review incorporates a broader spectrum of lesion types (ESIN, Barrett’s dysplasia, CAG, GIM, and gastric dysplasia), as well as literature published through March 2026, and provides a broad synthesis of AI diagnostic performance across the UGI precancerous lesion spectrum.
The pooled diagnostic estimates observed in the present analysis are broadly consistent with, yet extend beyond, those reported in prior meta-analyses of AI for individual UGI lesion categories. Dilaghi et al.13 reported a pooled diagnostic accuracy of 90% for AI-based detection of gastric precancerous lesions (and 80% for H. pylori infection), whereas the present analysis, which spans a wider lesion spectrum, reports sensitivity and specificity directly and incorporates more recent architectures. The higher sensitivity observed with image-enhanced endoscopy modalities than with WLE (sensitivity 0.960 vs. 0.875; P = 0.017) is consistent with findings from individual high-quality studies and suggests potential complementarity between AI and advanced imaging. Yuan et al.42 evaluated an AI-assisted system for detecting esophageal precancerous lesions under WLE and non-magnified NBI in a multicenter randomized controlled trial. Similarly, Xu et al.39 reported a multicenter computer-aided detection system for gastric atrophy and GIM using magnifying narrow-band imaging and blue laser imaging, achieving accuracy of 0.901 in internal and 0.864 in external validation. These data suggest that the diagnostic yield of AI may be enhanced when deployed in conjunction with advanced endoscopic imaging, and that imaging modality should be considered an important factor associated with AI system performance in both study design and clinical implementation.
The rapid advancement of AI endoscopy represents a timely opportunity to address the persistent gaps in UGI cancer surveillance, particularly given the growing global burden of gastric and esophageal cancers and the well-documented limitations of conventional WLE-based surveillance. As health systems increasingly pursue population-level endoscopic screening programs, particularly in high-incidence regions of East Asia, AI-assisted endoscopy may help standardize detection quality, reduce interobserver variability, and extend specialist-level diagnostic capability to lower-resource settings. However, translating the high accuracy observed in controlled research settings into routine clinical benefit will require several critical developments.43 Future studies should prioritize prospective, multicenter designs with consecutive patient enrollment to minimize spectrum bias and enhance generalizability. Standardized reporting of performance metrics stratified by image quality, lesion subtype, and endoscopist experience level would facilitate more meaningful cross-study comparisons. Furthermore, the development of interpretable AI architectures capable of providing lesion-level visual explanations alongside diagnostic outputs may help build clinician trust and support regulatory review. The integration of AI with real-time video endoscopy, rather than static image analysis, represents a natural next step, as evidenced by emerging systems providing frame-by-frame lesion delineation during live procedures.42 Addressing these priorities will be important for bridging the gap between experimental research and clinical-grade AI tools for UGI precancerous lesion surveillance.
An important caveat concerns the direction and likely magnitude of the patient-selection bias identified above. Because 9 of 15 studies (60%) drew their test images from retrospectively curated, non-consecutive samples, often selecting technically adequate, unambiguous images for algorithm testing, the sensitivity and specificity estimates reported here are best interpreted as an upper bound on achievable performance under optimal, curated conditions, rather than a reliable estimate of accuracy in unselected, real-world screening populations, where image quality is more variable and lesion presentations are less clear-cut. Prospective studies with consecutive enrollment, reflecting the full spectrum of lesion appearances encountered in routine practice, are needed to determine the extent to which current estimates will generalize to unselected screening populations.
The clinical value of AI-assisted diagnosis is also likely to vary substantially across intended-use scenarios, and a one-size-fits-all interpretation of these pooled estimates is unlikely to be appropriate. AI in endoscopy can serve several distinct roles, including flagging suspicious areas during screening, guiding targeted biopsy sampling, and informing treatment planning; its clinical utility is not equivalent across these roles. Because the included studies overwhelmingly evaluated AI as a static-image or video-clip classifier rather than in these more specific decision-support capacities, the available evidence most directly supports AI’s use in flagging lesions for endoscopist review during screening, a role in which randomized evidence has shown that AI assistance can improve endoscopists’ diagnostic performance for Barrett’s neoplasia.44 Its incremental value is likely to be considerably smaller in scenarios where lesion size or morphology already mandates endoscopic resection irrespective of the AI-assigned grade, since management would not change regardless of the algorithm’s output. Clarifying the specific clinical scenarios in which AI is most likely to add value, and those in which its contribution may be limited, will help clinicians and readers critically evaluate the applicability of these findings to their own practice setting.
A further limitation relates to the gap between offline, static-image (or preselected video-clip) validation, the paradigm used by nearly all included studies, and the demands of real-time video endoscopy in clinical practice. Diagnostic performance measured on curated images does not necessarily translate to frame-by-frame performance during continuously moving, live endoscopic examinations, where motion artifact, variable insufflation, mucus, and bubbles introduce sources of difficulty that static-image benchmarks do not capture. Findings regarding the clinical readiness of AI-assisted UGI endoscopy should therefore be extrapolated to routine practice with caution until performance has been confirmed in prospective, real-time video studies.
Several limitations of the present meta-analysis warrant consideration. First, substantial heterogeneity was observed across included studies (bivariate I² = 73%), which, although partially explained by imaging modality, could not be fully accounted for by the prespecified subgroup variables, suggesting the presence of additional unmeasured sources of variability such as lesion prevalence, AI model architecture, and training data composition. Notably, diagnostic accuracy also varied meaningfully by target lesion category (Table 1; Fig. 1f), with dysplasia showing the highest specificity but comparatively lower sensitivity than intestinal metaplasia or gastric atrophy; pooling across lesion types into a single overall estimate, as in the headline result, therefore risks obscuring clinically important differences in AI performance by lesion subtype, and lesion-specific estimates should be preferred when applying these results to a specific clinical question. Second, among the included studies, 9 of 15 studies (60%) were judged to be at high risk of bias in the patient selection domain of QUADAS-2, primarily owing to retrospective image curation and non-consecutive enrollment; the performance estimates derived from such studies are likely to overestimate real-world diagnostic accuracy. Third, the marked geographic concentration of included studies, with 10 of 15 (66.7%) originating from China, limits the generalizability of findings to populations with different disease epidemiology, H. pylori prevalence, and endoscopic practice patterns. The GRAIDS study, the largest multicenter AI endoscopy study to date, demonstrated that performance equivalent to expert endoscopists could only be achieved after training on images from multiple centers,41 underscoring that single-center models are unlikely to generalize without geographically diverse training data. Federated learning approaches, which enable cross-institutional model training without centralizing patient data, represent a promising methodological avenue to address this challenge.45 Fourth, multiple datasets derived from the same study and, in some cases, the same patient cohort were treated as independent observations in the primary analysis. To assess the impact of this non-independence, we performed leave-one-out sensitivity analyses at both the dataset and study level, which showed that excluding any single dataset or study changed the pooled sensitivity or specificity by no more than 2 percentage points, indicating that the overall results were not driven by any individual study. We additionally collapsed the 24 datasets into 15 independent study-level records (summing counts within each study) and re-pooled the estimates as a study-level aggregated sensitivity analysis: the point estimates were essentially unchanged (sensitivity 0.919 vs. 0.924; specificity 0.945 vs. 0.938), but the corresponding 95% confidence intervals widened by approximately 21–26% relative to the primary 24-dataset analysis. This confirms that treating non-independent datasets as independent produced somewhat anti-conservative confidence intervals, although it did not materially alter the point estimates or overall conclusions; subgroup and meta-regression P values in this study should nonetheless be interpreted with appropriate caution (Supplementary File 3). In addition, the included datasets used different units of analysis, including images, video clips, and patients; pooled estimates combining these analytic units should therefore be interpreted with caution. Notwithstanding these limitations, the consistency of findings across diverse study designs, lesion types, and geographic settings supports the robustness of the overall conclusions.
Conclusions
This systematic review and meta-analysis indicated that AI-assisted endoscopy has high diagnostic accuracy for upper gastrointestinal precancerous lesions, under the largely retrospective, curated conditions that predominate in the current evidence base, with pooled sensitivity of 0.930, specificity of 0.944, and AUC of 0.9661. Among the variables formally evaluated by meta-regression, imaging modality was associated with heterogeneity, while diagnostic performance also varied meaningfully by target lesion category, with image-enhanced endoscopy showing significantly higher sensitivity than white-light endoscopy. These findings most directly support a role for AI as a detection-assistance tool during routine screening endoscopy, flagging subtle or easily overlooked mucosal abnormalities for endoscopist review, rather than as a stand-alone diagnostic or treatment-planning tool, particularly given the predominance of retrospective, single-center studies with non-consecutive image selection and the current scarcity of real-time, multicenter prospective validation. Future research should prioritize multicenter prospective trials with consecutive patient enrollment, standardized outcome reporting encompassing real-time video evaluation and lesion-subtype-specific performance, and the development of interpretable AI systems to facilitate a more confident transition from experimental research to routine clinical practice in appropriate use scenarios.
Supporting information
Supplementary material for this article is available at https://doi.org/10.14218/CSP.2026.00010 .
Supplementary File 1
PRISMA 2020 Checklist.
(PDF)
Supplementary File 2
Literature Search Strategies.
(DOCX)
Supplementary File 3
Leave-One-Out Sensitivity Analysis and Study-Level Aggregated Sensitivity Analysis.
(XLSX)
Declarations
Funding
This work was supported by the Noncommunicable Chronic Diseases–National Science and Technology Major Project (Grant Nos. 2025ZD0545303, 2025ZD0545300), the CAMS Innovation Fund for Medical Sciences (CIFMS) (Grant No. 2025-I2M-XH-XX-047), and the National High-Level Hospital Clinical Research Funding Program (Grant No. 2025-LYZX-R-B04). The funders had no role in the study design; data collection, analysis, or interpretation; manuscript preparation; or the decision to submit the manuscript for publication.
Conflict of interest
The authors declare that they have no conflicts of interest.
Author contributions
Conceptualization (YB, TG, JL), methodology (YB, DH, TG), software (YB, DH), validation (JyL), formal analysis (YB, DH), investigation (YB, DH), data curation (YB, DH, JyL), writing – original draft (YB, DH, YL), writing – review and editing (JyL, YL, JZ, TG, JL), visualization (YL), supervision (JZ, TG, JL), project administration (JZ, JL), and funding acquisition (JL). All authors have made significant contributions to this study and have approved the final manuscript.
Data sharing statement
The data that support the findings of this study are available from the corresponding author upon reasonable request.