Diagnostic accuracy of artificial intelligence models in the detection of gastric cancer and early gastric cancer: a systematic review and network meta-analysis

Background This study aimed to evaluate and rank the diagnostic accuracy of different Artificial intelligence (AI) models for GC and EGC detection across imaging modalities through a systematic review and network meta-analysis. Methods PubMed, Embase, Cochrane Library, and Web of Science were comprehensively searched from inception to September 2026. Studies assessing AI models against histopathological reference standards (endoscopic biopsy or surgical specimens) were included. Four main model categories (traditional machine learning (ML), basic convolutional neural network (CNN), object detection systems (DS), and multimodal deep learning (DL)) and five subcategories (ML, ResNet, Fusion, Det, CNN_other) were compared. Bayesian network meta-analysis was used to estimate pooled sensitivity, specificity, diagnostic odds ratio (DOR), superiority index, and the summary receiver operating characteristic (SROC) curves. Results 59 studies were included. In primary analyses, for GC detection, multimodal DL showed the highest SROC area (0.98) and DOR (166.02), followed by basic CNN (SROC, 0.95; DOR, 105.29) and traditional ML (0.87; 35.83). For EGC detection, traditional ML showed the highest point estimates (SROC, 0.91; DOR, 182.13), but these were based on only three studies with wide confidence intervals, and no model category could be reliably ranked as superior. In subcategory analyses, CNN_other showed the highest GC point estimates and ML showed the highest EGC point estimates. Heterogeneity was high (I² > 90% for most DL categories). About 35% of studies had high risk of selection bias, and 8 had high risk of index-test bias, potentially inflating accuracy estimates. Conclusions This network meta-analysis revealed that AI models are generally effective in diagnosing GC and EGC. Multimodal DL and non-ResNet CNNs yield the highest point estimates for GC. For EGC, no model category clearly outperformed the others. Traditional ML had numerically higher point estimates, but this result was based on few studies with wide uncertainty. These rankings are limited by substantial heterogeneity, indirect comparisons, and a high prevalence of selection and index-test bias, and should be regarded as hypothesis-generating and require modality-specific validation. Prospective, multicenter head-to-head studies with standardized imaging protocols are needed to validate these findings and guide model selection. Systematic review registration https://www.crd.york.ac.uk/prospero/ , identifier CRD420261368020.