Synergistic applications of artificial intelligence and organoid technology in gastric precancerous lesion research: Mechanisms, translation, and challenges

Table 1 Diagnostic performance of artificial intelligence systems for gastric cancer and precancerous lesions

No.	Research focus	Ref.	Modality	Study design	Sample size/dataset	AI model/system	Gold standard	Key performance metrics	Results	Comparison with endoscopists
1	EGC diagnosis	Chen et al[20]	WLE, NBI	Systematic review (12 studies)	11685 cases	Various	Pathology	Pooled sensitivity, specificity, AUC	Sensitivity: 0.86 (95%CI: 0.75-0.92); specificity: 0.90 (95%CI: 0.84-0.93); AUC: 0.94	Not compared
2	Upper GI tumors	Arribas et al[21]	NBI	Meta-analysis (19 studies)	Not specified	Various	Pathology	Overall sensitivity, specificity, AUC	Sensitivity: 90%; specificity: 89%; AUC: 0.95	Not compared
3	Gastric neoplasia	Lui et al[22]	WLE, NBI	Systematic review and meta-analysis (23 studies)	969318 images	Various	Pathology	AUC	AUC: 0.96	Superior (AUC 0.98 vs 0.87, P < 0.001)
4	GPLs diagnosis	Dilaghi et al[23]	Not Specified	Systematic review and meta-analysis (4 studies)	Not specified	Various	Pathology	Accuracy	Accuracy: 90.3%	Not compared
5	CAG diagnosis	Shi et al[24]	Not specified	Systematic review and meta-analysis (8 studies)	25216 patients, > 90000 images	Various	Pathology	Pooled sensitivity, specificity, AUC	Sensitivity: 94%; specificity: 96%; AUC: 0.98	Significantly higher accuracy
6	CAG diagnosis	Zhang et al[25]	Not specified	Diagnostic study	5470 antral images	CNN	Pathology	Accuracy, sensitivity, specificity	Accuracy: 0.942; sensitivity: 0.945; specificity: 0.940	Exceeded three experts
7	CAG diagnosis	Shi et al[26]	Not specified	Diagnostic study	Not specified	GAM-efficient net	Pathology	Accuracy	External image: 93.5%; video: 92.37%	Outperformed endoscopists
8	GIM diagnosis	Yan et al[27]	Not specified	Diagnostic study	Not specified	Intelligent diagnostic system	Pathology	AUC, sensitivity, specificity, accuracy	AUC: 0.928; sensitivity: 91.9%; specificity: 86.0%; Accuracy: 88.8%	Not compared
9	Mucosal lesion DDx	Nam et al[28]	Not specified	Diagnostic study	Not specified	AI-DDx	Pathology	AUROC	AUROC: 0.86	Comparable to experts (0.89, P = 0.12); Superior to novices and intermediates
10	CAG and IM diagnosis	Lin et al[29]	WLE	Multicenter diagnostic study	7037 images (14 hospitals)	CNN	Pathology	AUC, Accuracy	CAG: AUC 0.98, Acc 96.4%; IM: AUC 0.99, Acc 97.6%	Not compared
11	Atrophy and IM detection	Yang et al[31]	WLE, LCI	Diagnostic study	21420 images	Novel DL method	Pathology	Accuracy	Atrophy: 97.12%; IM: 99.18%	Not compared
12	GA and IM diagnosis	Xu et al[32]	Image-enhanced endoscopy	Multicenter diagnostic study	6250 images, 98 videos (5 hospitals)	ENDOANGEL (DCNN)	Pathology	Accuracy	GA: 86.4%; IM: 85.9%	Comparable to experts; Superior to non-experts
13	Precursor detection	Xu et al[33]	Not specified	Prospective single-center clinical trial	Not specified	Not specified	Pathology	Detection Rate	IM: 14.23% vs 9.15%; atrophy: 22.76% vs 17.28%	Effect more pronounced in junior physicians
14	Invasion depth	Nam et al[28]	EUS	Diagnostic study	Not specified	AI-ID	Post-operative histology	AUROC	AUROC: 0.73	Superior to EUS experts (0.56, P < 0.001)
15	Cancer vs ulcer	Namikawa et al[36]	Not specified	Diagnostic study	Not specified	A-CNN	Pathology	Sensitivity, specificity, accuracy	Sensitivity: 99.0%; specificity: 93.3%; accuracy: 95.9%	Not compared
16	GISTs vs leiomyomas	Dong et al[37]	EUS	Multicenter diagnostic study	Not specified	Real-time AI-assisted EUS system	Pathology/histology	AUC, accuracy	AUC: 0.948; accuracy: 91.7%	Significantly outperformed
17	Pathology analysis	Yang et al[40]	Macroscopic specimen	Diagnostic study	Gastric cancer surgery specimens	AI algorithm	Histology	mAP, Accuracy	Lesion localization mAP: 95.90%; LN metastasis prediction Acc: 75.00%	Not compared
18	IM scoring	Iwaya et al[41]	Pathology slides	Diagnostic study	Not specified	AI system	Expert pathologist	Scoring difference	Difference with pathologists: 7.6%	Identified missed foci by pathologists

AI: Artificial intelligence; AUC: Area under the curve; CAG: Chronic atrophic gastritis; CNN: Convolutional neural network; DCNN: Deep CNN; DDx: Differential diagnosis; DL: Deep learning; EGC: Early gastric cancer; EUS: Endoscopic ultrasound; GA: Gastric atrophy; GIM: Gastric intestinal metaplasia; GISTs: Gastrointestinal stromal tumors; GPLs: Gastric precancerous lesions; IM: Intestinal metaplasia; LN: Lymph node; mAP: Mean average precision; WLE: White light endoscopy; NBI: Narrow-band imaging; AUROC: Area under the receiver operating characteristic curve; LCI: Linked color imaging endoscopy; GI: Gastrointestinal; CI: Confidence interval; Acc: Accuracy.

Table 2 Role of artificial intelligence in reducing inter-observer variability and improving diagnostic consistency

No.	Research focus	Ref.	Modality	Study design	Sample size/dataset	AI model/system	Key performance metrics	Results without AI	Results with AI assistance	Key finding
1	EGC diagnosis with ME	Li et al[42]	Magnifying image-enhanced endoscopy	Diagnostic study	Not specified	ENDOANGEL-LA	Accuracy	Novice: 71.63%	Novice: 87.45%	AI assistance bridged the gap between novices and experts
2	Real-time AI assistance	Dong et al[43]	WLE	Diagnostic study	Not specified	ENDOANGEL-ED	Accuracy	Endoscopists: 70.61%	Endoscopists: 79.63% (P < 0.001)	AI significantly improved endoscopist diagnostic accuracy
3	CAG diagnosis	Zhao et al[45]	Not specified	Prospective nested case-control	1306 patients	Not specified	Kappa, accuracy, sensitivity, specificity	Endoscopists' kappa: 0.291; Acc: 68.89%; Sens: 67.56%; Spec: 70.23%	AI kappa: 0.816; Acc: 89.89%; Sens: 89.31%; Spec: 90.46%	AI agreement with pathology was substantially higher
4	Pathology Dx (atrophy/IM)	Fang et al[46]	Pathology slides	Observer study (10 pathologists)	Not specified	GasMIL (SDL algorithm)	AUC, weighted kappa	Pathologists’ performance (baseline)	Pathologists’ performance significantly improved	AI assistance improved pathologists’ diagnostic metrics

“Not Specified” indicates that the information was not explicitly mentioned in the provided text. AI: Artificial intelligence; CAG: Chronic atrophic gastritis; Dx: Diagnosis; IM: Intestinal metaplasia; ME: Magnifying endoscopy; SDL: Semi-supervised deep learning; Acc: Accuracy; Sens: Sensitivity; Spec: Specificity; AUC: Area under the curves; WLE: White light endoscope.

Table 3 Major limitations and challenges in the clinical integration of artificial intelligence for diagnosing gastric precancerous lesions

Category	Specific challenge/issue	Supporting evidence/explanation	Ref.
Generalizability and robustness	Performance degradation across centers	Most studies are single-center, leading to models that may not perform well on data from different hospitals, equipment, or populations	[47]
		A 2023 external validation study in Singapore found endoscopists’ average accuracy was superior to the AI system, highlighting generalizability issues	[48]
Clinical validation	Lack of high-level evidence	The majority of studies are retrospective with inherent limitations (selection bias, overfitting). Large-scale, prospective multicenter RCTs are needed	[9]
	Scarcity of RCTs in gastric cancer	A 2025 systematic review found only 11% of GI oncology AI RCTs focused on gastric cancer	[49]
	Unclear impact on patient outcomes	The feasibility, effectiveness, safety, and long-term impact on patient outcomes require validation	[34,35]
	Inconsistent conclusions on utility	Studies comparing AI against endoscopists of varying experience levels have yielded inconsistent results	[48]
Data and algorithms	High heterogeneity	High heterogeneity in algorithms, imaging techniques, and study designs makes comparing models difficult	[50]
	Lack of standardized, public datasets	Absence of public, standardized large-scale databases (e.g., “EndoNet”) introduces selection bias, limits reproducibility, and hinders fair comparison	[1,9,13]
	Suboptimal training data ratios	The positive-to-negative sample ratio in training sets influences performance, with a suggested optimal ratio between 1:1 and 1:2	[51]
	Potential for missed diagnosis	Models trained primarily on typical lesions may miss rare, atypical, or minute early lesions	[41]
Interpretability and trust	“Black box” problem	The decision-making process of many AI models lacks transparency, hindering clinician trust, acceptance, and error troubleshooting	[1]
		Although techniques like heatmaps can help, the underlying logic remains insufficiently transparent, posing an obstacle to adoption	[47,52]
Regulatory and workflow	Underdeveloped regulatory framework	Lack of unified data standards, reliable evaluation systems, and comprehensive regulatory/ethical frameworks	[13]
	Undefined human-AI collaboration	The specific role of AI (e.g., primary screening, second opinion) and how to optimize human-computer interaction require further exploration	[48]

AI: Artificial intelligence; RCT: Randomized controlled trial; GI: Gastrointestinal.

Table 4 Proposed future research directions for artificial intelligence in diagnosing gastric precancerous lesions

Research direction	Specific goals/actions	Expected outcomes/rationale	Ref.
High-quality clinical trials	Conduct large-scale, multicenter, prospective RCTs	Evaluate real-world effectiveness, safety, generalizability, cost-effectiveness, and impact on patient outcomes to provide high-level evidence for clinical adoption	[34,49]
	Validate clinical utility in risk stratification and prediction	Bridge the gap between basic research and clinical application, strengthening translational research	[53]
Data standardization and infrastructure	Create public, standardized, large-scale image databases	Contain diverse images from different regions, equipment, and pathologies. Enhance research transparency, reproducibility, and facilitate fair algorithm comparison and improvement	[1]
	Develop international, multicenter, annotated databases (e.g., “EndoNet”)	Make high-quality data accessible to the research community to overcome a major current limitation
	Establish guidelines for data acquisition and model validation	Improve the reproducibility and reliability of AI research through academia-industry collaboration	[53,56]
Targeted algorithm development	Enhance detection of atypical/minute lesions	Employ augmented datasets enriched with rare cases for training to reduce miss rates and improve robustness
	Establish rigorous evaluation frameworks and ethical guidelines	Ensure model safety, reproducibility, and validate clinical translation potential	[54,55]
Comparative and optimization studies	Head-to-head comparison of AI algorithms and imaging techniques	Identify the optimal AI technical solution for specific clinical scenarios (e.g., screening vs depth assessment)	[20]
	Explore optimal human-AI collaboration models	Compare AI performance against endoscopists of varying experience; explore modes like real-time assistance, second reader, or quality control monitor to clarify AI’s best clinical role	[48]
	Optimize human-computer interaction interface and workflow	Maximize diagnostic efficiency and accuracy in clinical practice
Explainable AI	Develop interpretable AI models using heatmaps, attention mechanisms, etc.	Visualize the diagnostic basis of models, make the decision-making process transparent to clinicians, and enhance trust and efficiency in human-AI collaboration	[43,47,57]
Multimodal data fusion	Integrate endoscopic images, pathology, genomics, and clinical data	Build more comprehensive, robust, and accurate fused AI models for diagnosis, risk stratification, and prognosis prediction, enabling personalized medicine	[14,47]

AI: Artificial intelligence; RCT: Randomized controlled trial.

Table 5 Major limitations and challenges of organoid models in gastric precancerous lesion research

Category	Specific challenge/issue	Supporting evidence/explanation	Ref.
Research focus	Relative scarcity of precancerous lesion models	Current studies predominantly focus on advanced gastric cancers, with limited research on constructing organoid models for atrophic gastritis and intestinal metaplasia	[17]
Model fidelity and complexity	Lack of tumor microenvironment (TME) components	Existing models primarily consist of epithelial cells and lack critical TME components (immune cells, stromal cells, intratumoral microbiota), unable to fully replicate essential interactions (e.g., with H. pylori)	[16-18,54,64,65]
	Inability to recapitulate systemic physiology	Constraints in fully recapitulating vascular systems, innervation, and interactions with systemic physiological processes	[17]
Standardization and reproducibility	Lack of standardized protocols	Significant challenges exist in organoid culture methodologies, analytical techniques, and data interpretation, compromising reproducibility and reliability	[17,67]
	Model qualification gap	Absent standardized qualification processes undermine confidence in models' physiological relevance	[55]
Scalability and practicality	Challenges in scalability and cost	Limitations in scalability, reproducibility, cost-effectiveness, and time efficiency. Relatively long culture cycle, variable success rates, batch-to-batch variations, and high costs limit large-scale application	[54,55,68-70]
Model validation and comparison	Unclear representativeness	Whether organoids fully represent all characteristics of the original lesional tissue requires further validation. Inconsistent culture success rates and extended cycles are current drawbacks	[15]
	Lack of comparative studies	Notable absence of head-to-head studies comparing organoids against more established models (e.g., animal models, ALI models) to clarify their unique advantages and optimal applications	[6,67]
Clinical translation	Limited direct clinical evidence	Organoids are primarily used in basic research. Direct evidence for application in clinical diagnostics (e.g., predicting progression risk) remains scarce, and the technology remains distant from direct clinical implementation	[17,61,62,67]

ALI: Air-liquid interface; H. pylori: Helicobacter pylori.

Table 6 Proposed future research directions for organoid models in gastric precancerous lesion research

Research direction	Specific goals/actions	Expected outcomes/rationale	Ref.
Model development	Establish precancerous lesion organoid biobanks	Develop organoid models from patient tissues (e.g., with IM) and validate their ability to simulate malignant transformation in vitro, enabling study of key molecular events and driver genes	[17]
Enhanced complexity (co-culture)	Develop complex co-culture systems	Create co-culture organoid or “organoid-on-a-chip” systems incorporating vascular networks, immune cells (T cells, macrophages), stromal cells (fibroblasts), and H. pylori	[54,64,65]
		Utilize microfluidic technology for “tumor-on-a-chip” models to accurately mimic the complex TME for studying immune escape and prevention
		Employ 3D bioprinting to construct precise TMEs that better simulate in vivo responses to drugs, particularly immunotherapies
Standardization and biobanking	Establish SOPs and quality control systems	Promote the development of biobanks with detailed clinical/pathological information. Formulate standardized SOPs for culture, qualification, functional analysis, and data handling to enhance consistency and comparability	[54,55,68]
High-throughput screening (HTS)	Develop automated HTS platforms	Enhance the efficiency of organoid culture and drug testing through automation, microfluidics, acoustic manipulation, and high-content imaging	[70-72]
	Reduce costs	Explore low-cost alternative materials (e.g., synthetic hydrogels) to replace Matrigel, reducing the economic burden for large-scale screening and promoting translation	[70,72]
Comparative studies	Conduct multi-model comparison studies	Directly compare organoids vs GEMMs, chemical animal models, and ALI models in simulating the “Correa cascade” to determine the best model for specific research objectives and clarify organoid applications	[6,67]
Predictive diagnostics	Explore application in risk stratification	Use organoids from patients with different risk grades (e.g., LGIN vs HGIN) integrated with scRNA-seq to identify molecular features predictive of progression, enabling personalized risk assessment

ALI: Air-liquid interface; GEMM: Genetically engineered mouse model; HGIN: High-grade intraepithelial neoplasia; IM: Intestinal metaplasia; LGIN: Low-grade intraepithelial neoplasia; scRNA-seq: Single-cell RNA sequencing; SOP: Standard operating procedure; TME: Tumor microenvironment; H. pylori: Helicobacter pylori.

Table 7 Challenges and future directions for the integration of artificial intelligence and organoids in gastric precancerous lesion research

Category	Specific challenge/opportunity	Explanation/proposed Action	Expected outcome/goal	Ref.
Current challenges
Research gap	Lack of integrated studies	Almost no studies combine AI and organoid technologies. A complete lack of empirical data on their integrated application specifically to GPLs	N/A	[1]
Data dependency and validation	Limited training data	AI model performance depends on high-quality, large-scale data, yet public organoid pharmacogenomic data linked to clinical outcomes remain limited	N/A	[69]
Data standardization	Lack of unified standards	Effective integration requires standardized processes for data acquisition, processing, and annotation (imaging and organoid data). Currently, such standards are lacking	N/A	[1,13,73]
Data integration	Handling data heterogeneity	Integrating multi-omics and multi-dimensional data from different platforms/batches presents technical difficulties due to heterogeneity, sparsity, and high dimensionality	N/A	[53,76,77]
Clinical translation	Translational gap	Seamlessly integrating laboratory findings from AI and organoids into clinical workflows to improve patient outcomes remains a key challenge	N/A	[53,56,75]
Future directions
Direct integration	AI for organoid dynamic monitoring	Initiate research applying AI for real-time, label-free quantitative analysis of organoid morphological changes, proliferation, and differentiation	To high-throughput screen for factors or drugs influencing precancerous lesion progression	[38]
Prospective studies	Integrated clinical research projects	Design prospective, multicenter studies to collect imaging data and tissue samples. Build PDOs for screening/simulation, then use AI to integrate imaging and organoid data	To validate the value of synergy in risk prediction and intervention decision-making
Database construction	Multicenter standardized database	Establish a large-scale database containing clinical info, endo/pathology images, multi-omics, organoid culture/analysis, drug response, and outcome data	To provide a high-quality data foundation for developing and validating generalizable AI models	[1,69,91]
Targeted translational studies	Address specific unmet clinical needs	Use CRISPR-Cas9 in organoids to simulate mutations (e.g., abnormal folate metabolism, high RAMP1). Use AI-high-throughput platforms to screen/optimize targeted drugs	To validate efficacy for personalized treatment strategies
Algorithm development	Advanced data integration algorithms	Develop AI algorithms capable of handling data heterogeneity, imputing sparsity, and integrating multimodal data (imaging, spatial omics, scRNA-seq)	To build more comprehensive disease models and achieve better integration with organoid co-culture systems	[75,76]
Platform establishment	AI-driven HTS organoid analysis platform	Develop a dedicated AI image analysis platform for automated, quantitative dynamic monitoring of organoid growth, differentiation, and death	To enable high-throughput interpretation of drug screening results and accelerate translation from bench to bedside

AI: Artificial intelligence; HTS: High-throughput screening; PDOs: Patient-derived organoids; scRNA-seq: Single-cell RNA sequencing; GPL: Gastric precancerous lesion.