Artificial intelligence in chronic atrophic gastritis: A comprehensive review

Table 1 Artificial intelligence-assisted endoscopic detection of chronic atrophic gastritis

Ref.	Year	Aim	Study design	Algorithm	Training dataset	Validation/test dataset	Reference standard	Outcome (CAG focus only)
								Accuracy	Sensitivity	Specificity	Compare with endoscopists
Guimarães et al[6]	2020	Diagnosis of chronic atrophic gastritis	Single center	VGG16	200 images (10‑fold CV)	70 images (independent test)	Histopathology	92.9% (independent test)	100% (independent test)	87.5% (independent test)	Superior to endoscopists
Zhang et al[7]	2020	Diagnosis of chronic atrophic gastritis	Single center	DenseNet121	3829 images (70%, 5-fold CV)	1641 images (30%, independent test)	Histopathology	94.24%	94.58%	94.01%	Superior to endoscopists
Lin et al[8]	2021	Detection of gastric precancerous conditions	Multicenter	TResNet	5883 images	548 images	Histopathology	96.4%	96.2%	96.4%	Superior to endoscopists
Mu et al[9]	2021	AG vs non-AG classification	Multicenter	UNet++ + ResNet-50	5651 images (Layer 3)	453 images (internal); 258 images (external); 80 videos	Histopathology + expert consensus	88.78% (internal); 91.23% (external); 95.00% (video)	88.93% (internal); 93.24% (external)	88.61% (internal); 87.80% (external)	Superior to non-experts; comparable to experts
Mu et al[9]	2021	GA vs IM classification	Multicenter	UNet++ + ResNet-50	5651 images (Layer 3)	453 images (internal); 258 images (external); 80 videos	Histopathology + expert consensus	87.40% (internal); 85.81% (external); 92.86% (video)	91.56% (GA, internal); 82.43% (IM, external)	81.48% (GA, internal)	Superior to non-experts; comparable to experts
Xu et al[10]	2021	Detection of gastric precancerous conditions	Multicenter	VGG-16	3978 images	1052 images (external test)	Histopathology	86.4% (external)	89.8%(external)	79.4% (external)	Superior to non-expert and comparable to experts
Qu et al[11]	2021	Multi‑class diagnosis of 10 GI diseases	Multicenter	YOLO v3	28194 images (training); 9395 images (validation)	9398 images (multicenter test, CAG: 517); 5699 images (prospective test, CAG: 405)	Endoscopist evaluation	98.04% (multicenter CAG); 98.10% (prospective CAG)	70.79% (multicenter CAG); 75.31% (prospective CAG)	99.63% (multicenter CAG); 99.85% (prospective CAG)	NA
Luo et al[12]	2022	Detection of gastric antrum atrophy	Multicenter	ResNet50	2046 images (Data-set 1, 60%)	681 images (validation, 20%); 681 images (internal test, 20%); 300 images (external test, Data-set 2)	Histopathology	90.2% (internal); 89.0% (external)	89.1% (internal); 90.5% (external)	91.5% (internal); 89.0% (external)	Comparable to 3 expert endoscopists
Luo et al[12]	2022	Detection of CAG (all regions)	Multicenter	ResNet50	26514 images (Data-set 3, 80%, after augmentation; original 8285 images)	3313 images (validation, 10%); 1333 images (internal test, 10%); 200 images (external test, Data-set 4); 132 images (external test without antrum, Data-set 5)	Histopathology	85.9% (internal); 85.4% (external); 91.6% (non-antrum)	87.5% (internal); 87.0% (external); 91.2% (non-antrum)	85.4% (internal); 85.0% (external); 92.0% (non-antrum)	NA
Zhao et al[13]	2022	Diagnosis of CAG	Single center, prospective nested case-control	U-Net	3703 images (70%, 5-fold CV)	1587 images (30%, static test); 676 patients (matched cohort)	Histopathology	92.63% (test set); 90.53% (matched cohort)	92.73% (test set); 84.02% (matched cohort)	92.24% (test set); 97.04% (matched cohort)	Superior to endoscopists
Zhao et al[14]	2022	Diagnosis of CAG	Single center, prospective paired cohort	U-Net	3703 images	1587 images (static test); 268 patients (paired cohort)	Histopathology	92.63% (static test); 91.42% (paired cohort)	92.73% (static test); 88.07% (paired cohort)	92.24% (static test); 93.71% (paired cohort)	Superior to endoscopists
Chen et al[15]	2025	Differential diagnosis of AIG, HpAG, and NAG	Multicenter	ResNet101	18828 images (1070 patients)	233 patients (external)	Histopathology	92.1%	90.3%	93.1%	Superior to non-expert and comparable to experts
Yang et al[16]	2023	Detection of gastric precancerous conditions	Single center	LAG + DTL (SE-ResNet-50)	17137 original images (68548 after augmentation)	4283 images	Histopathology	97.12%	95.34%	98.90%	Superior to endoscopists
Chen et al[17]	2025	Diagnosis of chronic atrophic gastritis	Single center	Vision Transformer	1540 images (detection model)	130 images (CAG classification); 200 images (additional imbalanced test set)	Histopathology	91.0% (CAG, balanced); 98.4% (anatomy); 96.0% (imbalanced)	86.0% (CAG, balanced); 80.0% (imbalanced)	95.0% (CAG, balanced); 98.0% (imbalanced)	Superior to non-expert and comparable to experts
Chiang et al[18]	2026	Diagnosis of Helicobacter pylori infection and premalignant gastric conditions	Multicenter	Vision Transformer (Multi-step cascade, Step 5-premalignant conditions detection)1	17178 images (step 5)	853 images (step 5)	Histopathology (CAG)	88.6% (step 5)	79.4% (step 5)	74.7% (step 5)	NA
Zhao et al[19]	2023	Diagnosis of CAG and assisting OLGA staging	Multicenter	U-Net	3703 images	524 patients	Histopathology	89.89%	89.31%	90.46%	Superior to endoscopists
Tao et al[20]	2024	Diagnosis and risk stratification of gastric atrophy	Multicenter	UNet++ + ResNet50	5856 images	869 images	Histopathology	92.52%	92.72%	92.31%	Superior to endoscopists

¹Step 5 classifies premalignant gastric conditions, including atrophic gastritis and intestinal metaplasia. Only the atrophic gastritis results are extracted here to maintain the chronic atrophic gastritis focus of this table.

Studies on Helicobacter pylori, intestinal metaplasia, or general chronic gastritis are discussed in the text as indirect evidence. Only atrophic gastritis results are extracted here to maintain the chronic atrophic gastritis focus. Accuracy, sensitivity, and specificity values should not be directly compared across studies because of differences in datasets, reference standards, and validation settings. The table is intended to summarize study-level evidence, not to provide pooled estimates. CAG: Chronic atrophic gastritis; GI: Gastrointestinal; AIG: Autoimmune gastritis; HpAG: Helicobacter pylori-associated atrophic gastritis; NAG: Non-atrophic gastritis; NA: Not applicable; OLGA: Operative link for gastritis assessment; AG: Atrophic gastritis; IM: Intestinal metaplasia; GA: Gastric atrophy.