Artificial intelligence for endoscopic correlates of Correa’s cascade in gastric precancerous lesions and early neoplasia

Table 1 Cascade-aware framework linking cascade-related targets with visible endoscopic phenotypes, artificial intelligence task formulation, reference standards, clinical decision impacts, and comparability limitations in gastric endoscopy

Cascade-related target/histopathological entity	Visible endoscopic phenotype	Typical AI task	Reference standard	Clinical decision impact	Main comparability limitation
H. pylori-associated chronic gastritis	Diffuse or spotty redness; enlarged folds; altered RAC pattern	CADx for infection-status classification	UBT, RUT, histopathology, serology, stool antigen testing, or composite testing1	Supports infection diagnosis and guides eradication therapy	Post-eradication mucosal changes; heterogeneous reference standards
Gastric atrophy	Pale mucosa; visible submucosal vessels; atrophic border; reduced gastric folds	CADx; severity staging; segmentation	Histopathology, updated Sydney system, OLGA, Kimura-Takemoto, Kyoto classification	Supports risk stratification, biopsy targeting, and surveillance planning	Multifocal disease; sampling error; underestimation from single-image analysis
Intestinal metaplasia	Whitish plaques; villous pattern; LBC; MTB; patchy redness	CADx; extent grading; segmentation	Biopsy histopathology, EGGIM, OLGIM	Supports targeted biopsy, assessment of IM extent, and surveillance planning	Patchy distribution; sampling error; modality-dependent visualization
Gastric dysplasia/early neoplasia	Subtle color change; irregular surface or vascular pattern; flat or depressed lesion	CADe or CADx for dysplasia/neoplasia	Biopsy or resection histopathology; Vienna classification	Supports referral, targeted biopsy, and treatment planning	Interobserver variability; lesion coexistence; coarse reference labels
Early gastric cancer detection and delineation	Well-demarcated lesion; irregular MS or MV pattern; surface depression or elevation; fold convergence	CADe; CADx; lesion segmentation	Biopsy, expert endoscopic delineation, ESD or surgical histopathology, TNM staging	Supports lesion detection, margin delineation, and endoscopic treatment planning	Reference-margin heterogeneity; resection-enriched cohorts; verification bias
Invasion depth and curability assessment	Fold convergence; ulceration; wall stiffness; surface depression; non-lifting sign	Invasion-depth prediction; histologic differentiation prediction; treatment-curability assessment	Endoscopic resection or surgical histopathology	Supports selection of endoscopic resection vs surgical management	Verification bias; resection-enriched cohorts; invasion-depth label imbalance; limited external validation

¹Composite testing refers to the use of two or more diagnostic tests as the reference standard.

AI: Artificial intelligence; CADe: Computer-aided detection; CADx: Computer-aided diagnosis; EGGIM: Endoscopic Grading of Gastric Intestinal Metaplasia; ESD: Endoscopic submucosal dissection; H. pylori: Helicobacter pylori; LBC: Light-blue crest; MS: Microsurface; MTB: Marginal turbid band; MV: Microvascular; OLGA: Operative Link on Gastritis Assessment; OLGIM: Operative Link on Gastric Intestinal Metaplasia Assessment; RAC: Regular arrangement of collecting venules; RUT: Rapid urease test; TNM: Tumor-node-metastasis; UBT: Urea breath test; IM: Intestinal metaplasia.

Table 2 Model families commonly used in endoscopic artificial intelligence

Model family	Core idea	Practical implication
CNN	Learns local image features, including texture, edges, color, and mucosal patterns	Common backbone for classification, detection, and real-time applications
Transformer	Uses attention mechanisms to capture broader spatial or temporal context	Useful for context-aware analysis but more data-intensive and computationally intensive
Mamba	Uses state-space modeling for efficient long-sequence processing	Promising for video analysis and long-context tasks, although clinical validation remains limited
Traditional ML	Uses handcrafted features with classifiers such as support vector machines or random forests	Useful in selected applications but less flexible than end-to-end deep learning
Broad learning system	Uses a wide, expandable network structure	Enables rapid model updating in research settings, although clinical evidence remains limited
Automated deep learning	Automates model selection, hyperparameter tuning, or architecture search	Reduces development burden but may limit model transparency
Hybrid model	Combines multiple model families, such as CNN-transformer or CNN-ML architectures	May improve flexibility but remains dependent on the specific task, training data, labels, and validation strategy

CNN: Convolutional neural network; ML: Machine learning.

Table 3 Summary of artificial intelligence studies targeting endoscopic assessment of Helicobacter pylori infection

Tier1	Study	Modality	Task and unit	Reference standard2	Dataset and validation3	AI method4	Clinical maturity5	Key results
4A	Itoh et al[23], 2018	WLI	CADx; binary H. pylori infection-status classification; image-level	Serum H. pylori IgG antibody	179 images; retrospective image-based evaluation	CNN (GoogLeNet)	Retrospective image-based study	AUC 0.96; SEN 86.7%; SPE 86.7%
4A	Shen et al[24], 2023	WLI	CADx; binary H. pylori infection-status classification; video-derived case-level	Histopathology and/or UBT	113908 images from 599 videos and 456 patients	CNN (ResNet34 with weakly supervised MIL)	Retrospective image-/video-based study	AUC 0.95; ACC 89.9%; SEN 91.5%; SPE 88.8%
4A	Zou et al[25], 2024	WLI	CADx; binary H. pylori infection-status classification; patient-level	Concordant results from at least two tests (UBT, RUT, culture, or histopathology)	9457 images from 418 patients; randomized controlled trial	CNN (EfficientNet-B0 with SE-Net)	Randomized clinical trial	Test ACC 89.6%; RCT ACC 92.8% with AI vs 75.6% in control group
4A	Yan-Dong et al[26], 2025 (earlier:[82])	WLI	CADx; real-time binary H. pylori infection-status classification; patient-level	UBT	132297 images and videos from 3803 patients plus an additional 701 patients	CNN (site recognition[83] plus EF-CNN/HP-CNN pipeline)	Real-time clinical feasibility	AUC 0.918; ACC 86.3%; SEN 86.9%; SPE 85.9%; procedure duration 283.3 seconds
4A	Hu et al[27], 2025	WLI	CADx; binary H. pylori infection-status classification; case-level	Histopathology from biopsy specimens	308887 images and 197 videos; temporal and geographic external validation	Transformer (PVT-MIL with LSTM aggregation)	External validation; video-level evaluation	AUC 0.903-0.923; ACC 83.3%-85.3%; video ACC 85.3% with AI vs 76.7% for endoscopists
4B	Shichijo et al[28], 2019 (earlier:[84])	WLI	CADx; three-class H. pylori infection-status classification; image-level	UBT, stool antigen testing, blood or urine serology	122263 images; retrospective image-based evaluation	CNN (GoogLeNet)	Retrospective image-based study	Class-wise precision: Negative 80%; post-eradication 84%; positive 48%; runtime 261 s for 23699 images
4B	Nakashima et al[29], 2020 (earlier:[85])	WLI, LCI	CADx; three-class H. pylori infection-status classification; image-/video-level	Serum H. pylori IgG for uninfected and current infection; UBT for post-eradication	12887 images and 240 videos	CNN (separate WLI and LCI DCNNs)	Retrospective image-/video-based study	ACC for uninfected/current/post-eradication: WLI 75.0%/77.5%/74.2%; LCI 84.2%/82.5%/79.2%
4B	Li et al[30], 2025	WLI	CADx; three-class H. pylori infection-status classification with or without eradication history; image-level	UBT, stool antigen testing, serology, PG	44014 images; retrospective image-based evaluation	CNN (multistage DL model with SSL and SKD)	Retrospective image-based study; history-enhanced model	ACC for uninfected/post-eradication/current infection: Image-only model, 87.7%/83.6%/95.9%; history-enhanced model, 92.5%/91.1%/98.6%

¹Tier definitions and clinical decision impact: 4A: Binary Helicobacter pylori infection diagnosis, supporting infection assessment and eradication decision-making; 4B: Multi-class Helicobacter pylori infection diagnosis, supporting differentiation among uninfected, currently infected, and post-eradication phenotypes, which may guide eradication strategies and the interpretation of residual mucosal risk.

²Reference standard indicates the clinical verification method used to assign or confirm the dataset labels.

³Dataset counts refer to non-augmented original images, videos, or patients, when reported. Validation design is summarized as retrospective image-based evaluation, retrospective image-/video-based evaluation, external validation, prospective evaluation, randomized controlled trial, real-time feasibility, or other study designs reported in the original study.

⁴Artificial intelligence method denotes the primary artificial intelligence pipeline used for the main results and is reported at the architecture-family level, with key components noted where relevant. If a study evaluated multiple models, the best-performing or clinically emphasized configuration is reported.

⁵Clinical maturity is categorized according to the highest level of validation or deployment relevance reported, including retrospective image-based study, retrospective image-/video-based study, external validation, video-level evaluation, randomized clinical trial, real-time clinical feasibility, or history-enhanced model.

ACC: Accuracy; AUC: Area under the curve; CADx: Computer-aided diagnosis; CNN: Convolutional neural network; DCNN: Deep convolutional neural network; DL: Deep learning; H. pylori: Helicobacter pylori; IgG: Immunoglobulin G; LCI: Linked-color imaging; LSTM: Long short-term memory; MIL: Multi-instance learning; PG: Pepsinogen; PVT: Pyramid vision transformer; RCT: Randomized controlled trial; RUT: Rapid urease test; SE-Net: Squeeze-and-excitation network; SEN: Sensitivity; SKD: Self-knowledge distillation; SPE: Specificity; SSL: Self-supervised learning; UBT: Urea breath test; WLI: White-light imaging; AI: Artificial intelligence; HP: H. pylori; EF: Endoscopic finding.

Table 4 Summary of artificial intelligence studies targeting endoscopic assessment of gastric atrophy

Tier1	Study	Modality	Task and unit	Reference standard2	Dataset and validation3	AI method4	Clinical maturity5	Key results
4C	Guimarães et al[37], 2020	WLI	CADx; binary CAG classification; image-level	H&E biopsy histopathology; updated Sydney system	270 images; retrospective image-based evaluation	CNN (VGG16)	Retrospective image-based study	ACC 92.9%; SEN 100.0%; SPE 87.5%
4C	Zhang et al[38], 2020	WLI	CADx; binary CAG classification; image-level	Biopsy histopathology	5470 images; retrospective image-based evaluation	CNN (DenseNet121)	Retrospective image-based study	ACC 94.2%; SEN 94.5%; SPE 94.0%
4C	Shi et al[39], 2023	WLI	CADx; binary CAG classification; image-/video-level	Biopsy histopathology	10960 images and 118 videos; external image- and video-based testing	CNN (GAM-EfficientNet)	External validation; video-based evaluation	External image/video ACC 93.50%/92.37%; SEN 93.00%/96.23%; SPE 94.00%/89.23%
4D	Luo et al[40], 2022	WLI	CADx; antral atrophy severity grading and binary CAG classification; image-level	Biopsy histopathology; updated Sydney system; OLGA system	12325 images; retrospective image-based evaluation	CNN (two ResNet50 models)	Retrospective image-based study; severity grading	ACC 0.890 for antral atrophy; ACC 0.590 for severity; binary model ACC 0.854-0.916
4D	Kodaka et al[42], 2022	WLI	CADx; site-level GA classification and patient-level severity assessment	Biopsy histopathology; Kimura-Takemoto, Kyoto, and OLGA systems	19221 images; patient-level rule-based aggregation	CNN (ResNet34 with rule-based aggregation)	Patient-level risk stratification	Patient-level ACC 67.4% vs Kimura-Takemoto; 68.1% vs Kyoto; 62.2% vs OLGA; AUC 0.746 using the modified Kyoto classification
4D	Zhao et al[43], 2023 (earlier:[86])	NBI	Segmentation and CADx; site-level atrophy and patient-level OLGA risk assessment; video-level	Biopsy histopathology; Kimura-Takemoto classification	5290 images and 907 patients; test set and prospective cohort	CNN (U-Net)	Prospective evaluation; real-time video feasibility; patient-level risk stratification	Test/prospective ACC 92.73%/89.89%; SEN 92.24%/89.31%; SPE 92.63%/90.46%
4D	Tao et al[44], 2024	WLI	Segmentation and CADx; site-level GA and patient-level risk assessment; image-/video-level	Biopsy histopathology; Kimura-Takemoto classification	11538 images and 119 videos; image- and video-based evaluation	CNN (UNet++/ResNet50 plus site-identification model[87,88])	Video-based evaluation; patient-level risk stratification	Image/video ACC 92.52%/94.12%; SEN 92.72%/94.87%; SPE 92.31%/92.68%
4D	Liu et al[41], 2024	WLI, NBI	CADx; Kyoto gastritis score prediction; image-level	Kyoto classification of gastritis	29013 images; retrospective image-based evaluation	CNN (five GAM-EfficientNet models)	Retrospective image-based study; score prediction	Average ACC 78.70%; SEN 78.70%; SPE 91.92%

¹Tier definitions and clinical decision impact: 4C: Chronic atrophic gastritis diagnosis, supporting recognition of atrophic changes and biopsy planning; 4D: Gastric atrophy severity classification and risk stratification, supporting assessment of disease extent, surveillance planning, and patient-level risk stratification.

²Reference standard indicates the clinical verification method used to assign or confirm the dataset labels.

³Dataset counts refer to non-augmented original images, videos, or patients, when reported. Validation design is summarized according to the study design reported in the original article.

⁴Artificial intelligence method denotes the primary artificial intelligence pipeline used for the main results and is reported at the architecture-family level, with key components noted where relevant. If a study evaluated multiple models, the best-performing or clinically emphasized configuration is reported.

⁵Clinical maturity is categorized according to the highest level of validation or deployment relevance reported, including retrospective image-based study, external validation, prospective evaluation, real-time video feasibility, video-level evaluation, patient-level risk stratification, or score prediction.

ACC: Accuracy; AI: Artificial intelligence; AUC: Area under the curve; CADx: Computer-aided diagnosis; CAG: Chronic atrophic gastritis; CNN: Convolutional neural network; GA: Gastric atrophy; GAM: Global attention module; HE: Hematoxylin and eosin; NBI: Narrow-band imaging; OLGA: Operative Link on Gastritis Assessment; SEN: Sensitivity; SPE: Specificity; WLI: White-light imaging.

Table 5 Summary of artificial intelligence studies targeting endoscopic assessment of intestinal metaplasia

Tier1	Study	Modality	Task and unit	Reference standard2	Dataset and validation3	AI method4	Clinical maturity5	Key results
5A	Lin et al[52], 2021	WLI	CADx; three-class gastritis and IM classification; image-level	Biopsy histopathology; updated Sydney system	7037 images; retrospective image-based evaluation	CNN (TResNet)	Retrospective image-based study	CAG ACC 96.24%; IM ACC 97.16%
5A	Wong et al[53], 2022 (earlier:[89,90])	M-NBI	CADx; binary IM classification; image-level	Biopsy histopathology	1516 images; retrospective image-based evaluation	CNN plus BLS (ResNet18 with PAM and stacking ensemble)	Retrospective image-based study	AUC 0.931; ACC 93.2%; TPR 93.6%; TNR 91.2%
5A	Yang et al[54], 2023	WLI, LCI	CADx; binary IM or GA detection; image-level	Biopsy histopathology	21420 images; retrospective multimodal image-based evaluation	CNN (SE-ResNet50 with LAG and DTL)	Retrospective image-based study; multimodal comparison	IM ACC 98.49% with WLI and 99.18% with LCI; GA ACC 95.06% with WLI and 97.12% with LCI
5A	Ligato et al[55], 2024	BLI	CADx; patch-level IM classification with image-level voting	Biopsy histopathology	1384 images; patch-to-image thresholding	CNN (ResNet50 patch classifier with voting)	Retrospective patch-/image-based study	Patch/image ACC 74%/78%; SEN 76%/70%; SPE 72%/100% at a threshold of 0.8
5B	Lai et al[56], 2022	WLI, NBI, M-NBI	CADx; four-class IM severity grading; image-level	Not specified	1175 images; retrospective image-based evaluation	BLS (M3JLBN with multi-scale multi-feature learning and MIL)	Retrospective image-based study; severity grading	ACC 85.0%; SEN 85.1%; PRE 85.2%
5B	Li et al[57], 2023	IEE	CADx; three-class EGGIM grading; site-level	EGGIM-based endoscopic grading without biopsy confirmation	1115 images; retrospective image-based evaluation	CNN (ResNet101 with color-guided deformable convolution)	Offline EGGIM-oriented grading	ACC 84.17%; SEN 83.65%; SPE 92.06%
5B	Niu et al[58], 2024	M-NBI	CADe plus CADx; site-level IM grading and patient-level OLGIM risk stratification	Biopsy histopathology; updated Sydney system; modified OLGIM system	4473 images; internal and external validation	CNN (Faster R-CNN with ResNet50 and SE block)	External validation; patient-level risk stratification	IM grading AUC 0.872 (internal) and 0.803 (external); ACC 80.0% (internal) and 70.4% (external); OLGIM ACC 85.0% (internal) and 76.0% (external)
5B	Almeida et al[59], 2025	NBI	CADx; site-level EGGIM grading and patient-level EGGIM calculation	EGGIM-based endoscopic grading without biopsy confirmation	1280 images cropped into patches; image- and patient-level evaluation	CNN (ResNet50)	Patient-level EGGIM-oriented grading	Per-image ACC 87%; per-patient ACC 88%; SEN 100%; SPE 85%
5C	Wang et al[60], 2019	WLI	Segmentation; IM lesion delineation; pixel-level	Biopsy histopathology	200 images; retrospective image-based evaluation	CNN (DeepLabV3)	Offline segmentation	ACC 89.51%; mIoU 80.67%
5C	Pornvoraphat et al[61], 2024 (earlier:[91,92])	NBI	Segmentation; real-time IM lesion delineation; pixel-level	Biopsy histopathology	4951 images; inference speed reported	CNN (modified BiSeNet with switch adapter)	Real-time segmentation feasibility	ACC 96.31%; SEN 94.09%; SPE 97.07%; IoU 55.55%; 158 FPS
5C	Zhang et al[62], 2026	LCI	Segmentation; IM lesion delineation; pixel-level	Biopsy histopathology	1884 images; inference speed reported	Mamba (VM-UNetV2)	Deployment-oriented segmentation	ACC 86.05%; SEN 78.29%; SPE 88.46%; mIoU 57.02%; 34.61 FPS

¹Tier definitions and clinical decision impact: 5A: Intestinal metaplasia diagnosis, supporting targeted biopsy and identification of precancerous mucosal change; 5B: Intestinal metaplasia severity grading and risk stratification, supporting assessment of disease extent and surveillance planning; 5C: Intestinal metaplasia localization and segmentation, supporting spatial mapping, targeted biopsy, and endoscopic documentation of intestinal metaplasia extent.

²Reference standard indicates the clinical verification method used to assign or confirm the dataset labels.

³Dataset counts refer to non-augmented original images, videos, or patients, when reported. Validation design is summarized according to the study design reported in the original article.

⁴Artificial intelligence method denotes the primary artificial intelligence pipeline used for the main results and is reported at the architecture-family level, with key components noted where relevant. If a study evaluated multiple models, the best-performing or clinically emphasized configuration is reported.

⁵Clinical maturity is categorized according to the highest level of validation or deployment relevance reported, including retrospective image-based study, internal and external validation, image- and patient-level evaluation, real-time segmentation feasibility, deployment-oriented segmentation, and patient-level risk stratification.

ACC: Accuracy; AUC: Area under the curve; BLI: Blue-light imaging; BLS: Broad learning system; CADe: Computer-aided detection; CADx: Computer-aided diagnosis; CAG: Chronic atrophic gastritis; CNN: Convolutional neural network; DTL: Dual transfer learning; EGGIM: Endoscopic Grading of Gastric Intestinal Metaplasia; FPS: Frames per second; GA: Gastric atrophy; IM: Intestinal metaplasia; IoU: Intersection over Union; LAG: Local attention grouping; LCI: Linked-color imaging; MIL: Multi-instance learning; M-NBI: Magnifying narrow-band imaging; mIoU: Mean Intersection over Union; NBI: Narrow-band imaging; OLGIM: Operative Link on Gastric Intestinal Metaplasia Assessment; PAM: Parallel attention module; PRE: Precision; SE: Squeeze-and-excitation; SEN: Sensitivity; SPE: Specificity; TNR: True negative rate; TPR: True positive rate; WLI: White-light imaging.

Table 6 Summary of artificial intelligence studies targeting endoscopic diagnosis of early gastric neoplasia

Tier1	Study	Modality	Task and unit	Reference standard2	Dataset and validation3	AI method4	Clinical maturity5	Key results
6A	Cho et al[69], 2019	WLI	CADx; five-class neoplasia classification; image-level	Biopsy histopathology	5217 images; retrospective image-based evaluation	CNN (inception-ResNet-v2)	Retrospective image-based study	ACC 84.6%; elapsed time 00264 second/image
6A	Gong et al[70], 2023 (earlier:[93])	WLI	CADx and segmentation; six-class disease classification and IM/atrophy delineation	Biopsy histopathology	13672 images; external test set and prospective real-world clinical evaluation	CNN (YOLOv3, EfficientNet, and UNet++ multitask pipeline)	External validation; prospective evaluation	Classification ACC 85.3% (external) and 89.3% (prospective); segmentation dice coefficient 84.7%-84.9%
6A	Kim et al[71], 2024	WLI	CADe and CADx; six-class lesion detection and tumor-depth stage estimation	Biopsy histopathology; resection specimens for cancers, when available	10181 images; retrospective image-based evaluation	CNN (YOLOv7 detector plus T-stage model)	Retrospective image-based study; depth estimation	Detection ACC 73.43%; SEN 80.90%; SPE 83.32%; T-stage ACC 85.17%
6B	Wu et al[72], 2021 (earlier:[94])	WLI	CADe and CADx; focal lesion detection and binary neoplasia diagnosis	Biopsy histopathology; revised Vienna classification	1812 patients; randomized AI-first vs routine-first design	CNN (ENDOANGEL-LD)	Randomized clinical trial	Lesion miss rate 61% with AI-first vs 27.3% with routine-first; first-examination detection rate 49% vs 3.5%
6B	Du et al[73], 2023	WLI, WMI	CADx; binary neoplasia diagnosis; image-/video-level	Biopsy histopathology	4201 images, 7436 paired images, and 162 videos; image, video, and prospective evaluation	CNN (ResNet50)	Prospective evaluation; video-level evaluation; assistance study	Image/video/prospective ACC 86.54%/90.00%/93.55%; non-expert ACC improved from 70.75% to 85.75%
6B	Dong et al[74], 2023	WLI	CADx; binary neoplasia diagnosis; image-/video-level	Biopsy histopathology	4482 images and 296 videos; external image, external video, and consecutive video evaluation	CNN plus ML (DCNN and YOLOv3 features with random forest)	External validation; consecutive video evaluation	External image/video/consecutive video ACC 77.17%/88.24%/79.76%; SEN 87.30%/97.06%/88.24%; SPE 73.08%/82.35%/77.61%
6B	Feng et al[75], 2025	WLI	CADe and CADx; real-time EGC detection and binary EGC diagnosis	Post-ESD histopathology	13289 images and 130 videos; image- and video-based evaluation	CNN (CenterNet with ResNet18)	Real-time video feasibility	Image/video AUC 0.917/0.930; SEN 93.38%/96.92%; SPE 90.07%/89.23%; inference time 0028 second/image

¹Tier definitions and clinical decision impact: 6A: Dysplasia-explicit neoplasia diagnosis, supporting differentiation of clinically adjacent stages in the gastric carcinogenesis cascade and informing biopsy, referral, and therapeutic evaluation; 6B: Composite neoplasia diagnosis, supporting safety-oriented detection of clinically significant lesions and reducing missed neoplasia, but with limited granularity for treatment selection.

²Reference standard indicates the clinical verification method used to assign or confirm the dataset labels.

³Dataset counts refer to non-augmented original images, videos, or patients, when reported. Validation design is summarized according to the study design reported in the original article.

⁴Artificial intelligence method denotes the primary artificial intelligence pipeline used for the main results and is reported at the architecture-family level, with key components noted where relevant. If a study evaluated multiple models, the best-performing or clinically emphasized configuration is reported.

⁵Clinical maturity is categorized according to the highest level of validation or deployment relevance reported, including retrospective image-based study, external validation, prospective evaluation, randomized clinical trial, real-time video feasibility, video-level evaluation, assistance study, or depth estimation.

ACC: Accuracy; AI: Artificial intelligence; AUC: Area under the curve; CADe: Computer-aided detection; CADx: Computer-aided diagnosis; CNN: Convolutional neural network; DCNN: Deep convolutional neural network; EGC: Early gastric cancer; ESD: Endoscopic submucosal dissection; IM: Intestinal metaplasia; ML: Machine learning; SEN: Sensitivity; SPE: Specificity; WLI: White-light imaging; WMI: Weak magnification imaging.

Table 7 Summary of artificial intelligence studies targeting endoscopic delineation and depth staging of early gastric neoplasia

Tier1	Study	Modality	Task and unit	Reference standard2	Dataset and validation3	AI method4	Clinical maturity5	Key results
6C	Ling et al[76], 2021	M-NBI	CADx and segmentation; differentiation status and horizontal margin delineation	Histopathology for EGC differentiation; margin reference from EGC cases	5757 images and 2 videos; test set and man-machine comparison	CNN (VGG16 plus UNet++)	Retrospective image-based study; man-machine comparison	Differentiation ACC 83.3% in test set and 86.2% in man-machine comparison; margin ACC 82.7% for differentiated lesions and 88.1% for undifferentiated lesions at overlap threshold of 0.80
6C	Takemoto et al[77], 2023 (earlier:[95])	WLI	CADe and segmentation; EGC detection and heatmap-based extent demarcation	Histopathology from ESD specimen	1158 images; retrospective image-based evaluation	CNN (GoogLeNet patch classifier with sliding-window heatmap)	Retrospective patch/image segmentation	Image-level SEN 83.8%; SPE 77.5%; case-level successful detection 94.9%; mIoU 66.5%
6C	Ma et al[78], 2023	WLI, M-NBI, CE, NBI	CADx and segmentation; EGC diagnosis and lesion-region delineation	Biopsy histopathology	4697 images; retrospective multimodal image-based evaluation	CNN (GAIN-ResNet50 plus attention U-Net segmentation)	Retrospective image-based study; multimodal segmentation	Classification ACC 98.84%; SEN 97.38%; SPE 99.53%; segmentation PA 83.51%; IoU 0.64
6D	Bang et al[79], 2021 (earlier:[96])	WLI	CADx; mucosal vs submucosal invasion; image-level	Biopsy or resection histopathology	4702 images; two external test sets	CNN (Neuro-T AutoDL)	External validation; depth estimation	External test ACC 89.3% and 88.6%
6D	Wu et al[80], 2022	WLI, M-NBI	CADe and CADx; neoplasm detection, EGC diagnosis, invasion-depth estimation, and differentiation prediction	Biopsy histopathology; resection specimens for cancers, when available	68177 images and 100 videos; AI vs expert comparison	CNN (YOLOv3 plus ResNet50 models)	Video evaluation; expert comparison; multi-task treatment planning	AI vs experts ACC: WLI neoplasm 91.00% vs 76.91%; M-NBI EGC 89.00% vs 85.67%; invasion depth 78.57% vs 63.75%; differentiation 71.43% vs 64.41%
6D	Goto et al[81], 2023	WLI	CADx; mucosal vs submucosal invasion; image-level	Resection or surgical histopathology	700 images; AI, endoscopist, and human-AI cooperation comparison	CNN (EfficientNetB1)	Human-AI cooperation study; depth estimation	ACC 72.5% for AI, 70.0% for endoscopists, and 78.0% for human-AI cooperation; SEN 74.0%, 52.0%, and 76.0%; SPE 71.0%, 88.0%, and 80.0%

¹Tier definitions and clinical decision impact: 6C: Early gastric cancer delineation and extent mapping, supporting lesion-margin definition, biopsy or marking guidance, and endoscopic resection planning; 6D: Invasion-depth estimation and curability-related characterization, supporting selection between endoscopic resection and surgical management.

²Reference standard indicates the clinical verification method used to assign or confirm the dataset labels.

³Dataset counts refer to non-augmented original images, videos, or patients, when reported. Validation design is summarized according to the study design reported in the original article.

⁴Artificial intelligence method denotes the primary artificial intelligence pipeline used for the main results and is reported at the architecture-family level, with key components noted where relevant. If a study evaluated multiple models, the best-performing or clinically emphasized configuration is reported.

⁵Clinical maturity is categorized according to the highest level of validation or deployment relevance reported, including retrospective image-based study, external validation, prospective evaluation, randomized clinical trial, real-time video feasibility, video-level evaluation, assistance study, or depth estimation.

ACC: Accuracy; AI: Artificial intelligence; AutoDL: Automated deep learning; CADe: Computer-aided detection; CADx: Computer-aided diagnosis; CE: Chromoendoscopy; CNN: Convolutional neural network; EGC: Early gastric cancer; ESD: Endoscopic submucosal dissection; GAIN: Guided Attention Inference Network; IoU: Intersection over Union; M-NBI: Magnifying narrow-band imaging; mIoU: Mean Intersection over Union; NBI: Narrow-band imaging; PA: Per-pixel accuracy; SEN: Sensitivity; SPE: Specificity; WLI: White-light imaging.