The review, led by Chong Chen and colleagues, takes a deliberately different approach from earlier surveys that simply catalogued model performance. Instead, the authors applied a translational evidence framework, weighing each study according to its validation hierarchy, cohort diversity, reproducibility, interpretability, and readiness for real clinical workflows. They searched PubMed/MEDLINE, Web of Science, Embase, Scopus, and Google Scholar for English-language studies through January 2026, prioritizing primary gastric cancer studies for claims about diagnostic accuracy, prognostic stratification, or treatment-response prediction. Studies were rated as having lower translational maturity when they relied on retrospective single-center data, internal validation only, small or highly selected cohorts, or absent calibration analysis. Higher maturity required multicenter cohorts, independent external validation, prespecified endpoints, and evidence of clinical utility beyond raw discrimination metrics such as the area under the receiver operating characteristic curve.

Nowhere is the gap between laboratory performance and clinical reality more visible than in computational pathology. Whole-slide images provide extraordinarily rich morphological information, but their gigapixel scale, staining variability, scanner differences, and annotation burden create formidable methodological challenges. One deep learning model reported 100 percent accuracy and an AUC of 0.999 for binary stage classification from whole-slide images, yet the study used a limited, perfectly balanced dataset and included no independent external validation cohort, raising the strong possibility of overfitting or methodological inflation. More cautiously framed efforts include interpretable deep learning models that predict dMMR/MSI-H status from routine hematoxylin and eosin slides with AUCs up to 0.930, and models that infer Epstein-Barr virus status from biopsy specimens with an AUC of 0.87. The authors argue these should be treated as screening or triage tools pending prospective validation against established molecular assays across institutions, scanners, and patient populations.

Radiologic AI faces its own reproducibility problems. Radiomics, the high-throughput extraction of quantitative spatial and textural features from medical images, allows non-invasive profiling of tumor characteristics, but feature stability depends heavily on scanner type, acquisition protocol, and segmentation method. Dual-energy CT models integrating iodine concentration and monochromatic attenuation parameters have predicted serosal invasion with AUCs of 0.837 to 0.860 in multicenter validation and correlated with postoperative disease-free survival. Hybrid radiomics models combining features from the primary tumor and the lesser curvature on contrast-enhanced CT have improved prediction of lymph node metastasis, including in radiologically non-enlarged nodes, a persistent clinical dilemma. A 2.5D deep learning model analyzing venous-phase CT images has been proposed for preoperative prediction of lymphovascular invasion, an important marker of recurrence risk. All of these, the review stresses, remain research-stage evidence.

Endoscopy-based AI currently offers some of the strongest near-term translational evidence, largely because it can operate in real time and directly affect lesion detection. Early gastric cancer may present as subtle mucosal discoloration, minimal elevation or depression, or irregular microvascular patterns, making detection highly operator-dependent. Deep learning systems analyzing real-time video have achieved diagnostic performance comparable to or exceeding that of expert endoscopists in selected settings. Intriguingly, the review highlights the importance of video over static images: one static-image model saw its accuracy collapse from 82.7 percent to 56.6 percent when applied to video clips, whereas a purpose-built sequential video classifier maintained stable accuracy of 83.7 percent by incorporating dynamic information such as peristalsis. Hybrid explainable systems such as ENDOANGEL-LA, which mimic expert reasoning by integrating established endoscopic visual criteria with machine learning classifiers, achieved 88.76 percent accuracy versus 89 percent for experts, though the system requires magnification endoscopy that may limit deployment in lower-resource settings.

Emerging biological modalities are pushing AI beyond conventional imaging. An AI-analyzed epigenetic liquid biopsy using cell-free DNA methylation markers identified lymph node metastasis preoperatively in early T1 gastric cancer with an AUC of 0.86, potentially informing the choice between organ-sparing endoscopic resection and radical gastrectomy. A support vector machine-based stool DNA methylation test using a four-gene panel reported high specificity of 97.8 percent but a more modest sensitivity of 67.5 percent, meaning a substantial proportion of true cases would be missed in a screening context. Other exploratory approaches include a random forest model using 22 serum N-glycans to distinguish gastric cancer patients from healthy controls with AUCs above 0.90, and logistic regression models profiling bacterial extracellular vesicles from urine. The authors consistently flag these as preliminary, requiring larger validation cohorts and standardized pre-analytic procedures.

On the prognostic front, machine learning has shown genuine promise in a more mature setting: predicting postoperative complications. In a prospective multicenter study of 3,926 patients from the PACAGE database, random forest algorithms performed well for overall and infectious complications, identifying body mass index, age, and the number of examined lymph nodes as key risk factors, with SHAP values used for interpretability. Notably, a simpler dynamic nomogram built through machine learning feature selection achieved discrimination comparable to the best complex model, with an AUC of 0.856. For long-term outcomes, an interpretable model integrating clinical variables, CT radiomics, and pathology features reported an AUC of 0.903 for early recurrence prediction, and hybrid CNN-Transformer architectures combined with graph convolutional networks have generated prognostic signatures that outperform traditional TNM staging in retrospective analyses. Dynamic models that update recurrence risk using time-dependent changes in the Prognostic Nutritional Index have also improved three-year recurrence prediction compared with static models.

Treatment-response prediction represents perhaps the most ambitious frontier. A multicenter deep learning radiomics nomogram integrating contrast-enhanced CT features, radiomics, and clinical factors predicted neoadjuvant chemotherapy response with an AUC of 0.751 in external validation, a respectable but not yet decision-ready figure. Machine learning has also identified PAK6 as a candidate contributor to oxaliplatin resistance via the homologous-recombination repair pathway, and attention-aware models tracking dynamic circulating tumor cells have been proposed for non-invasive monitoring of therapeutic efficacy. For immunotherapy, deep learning models analyzing pathology slides may pre-screen for MSI-H/dMMR or EBV-positive tumors, subgroups enriched for response to immune checkpoint inhibitors, though confirmatory molecular testing remains mandatory. Multi-omics signatures, including ferroptosis scores associated with immune-inflamed tumor microenvironments, and hypotheses linking CDH1 loss to ferroptosis-based therapies remain exploratory. Evidence for AI-guided radiotherapy personalization in gastric cancer is so limited that the review declines to treat even related gastrointestinal data as more than hypothesis-generating.

The barriers to clinical translation are as much organizational as technical. Data heterogeneity, siloing, and inconsistent validation are identified as primary obstacles: models optimized for specific scanners, staining workflows, or population structures may fail when deployed elsewhere, a phenomenon known as domain shift. Biological heterogeneity compounds the problem, since gastric cancer encompasses genomically stable, chromosomally unstable, microsatellite-unstable, and Epstein-Barr virus-associated subtypes, and H. pylori virulence factors such as CagA EPIYA motifs vary geographically. Models trained predominantly on Western or East Asian cohorts may not generalize to underrepresented populations, including Alaska Native patients with distinct molecular landscapes and higher rates of EBV-associated tumors, raising genuine concerns about algorithmic bias. The opaque black-box nature of deep networks further threatens clinical trust, which the authors argue demands widespread adoption of explainable AI techniques such as SHAP and Grad-CAM, alongside strict adherence to data privacy and informed consent protocols.

The path forward, the review concludes, requires the field to move beyond high AUC values toward prospective, multicenter evaluation of clinical utility, safety, cost-effectiveness, and workflow impact. Future studies should report calibration, decision-curve analysis, subgroup performance, and failure modes, not just accuracy. Federated learning and secure multicenter evaluation offer ways to overcome data silos while protecting privacy, and multimodal foundation models may eventually reduce annotation burdens, though gastric cancer-specific evidence for these approaches remains limited. For integration into practice, AI systems must be embedded into existing endoscopy, radiology, pathology, and oncology workflows with user-centered interfaces that clarify uncertainty and recommended next steps while minimizing alert fatigue. Regulatory pathways, reimbursement models, medicolegal responsibility, and post-deployment monitoring of model drift should be planned from the outset rather than bolted on afterward. The strongest current candidates for clinical adoption are image-based tasks with clearly defined inputs and endpoints, endoscopic lesion detection, pathology-based biomarker pre-screening, and CT-based staging, while liquid-biopsy AI, multi-omics signatures, and large multimodal systems remain firmly in the exploratory category. What the review ultimately delivers is a realistic roadmap: artificial intelligence is unlikely to transform gastric cancer care overnight, but with rigorous validation, transparent reporting, and attention to equity, its most mature applications could reach patients within the current decade.

Subject of Research: Artificial intelligence applications for diagnosis, prognosis, and treatment-response prediction in gastric cancer

Article Title: Artificial Intelligence in Gastric Cancer: Diagnostic, Prognostic, and Predictive Developments, Evidence Maturity, and Translational Challenges

Article References: Chen, C., Xu, W., Wu, T., Zhao, J., Xu, X., Xiao, M., & Xu, K. (2026). Artificial Intelligence in Gastric Cancer: Diagnostic, Prognostic, and Predictive Developments, Evidence Maturity, and Translational Challenges. Cancer Reports, 9(10), Article e70665. https://doi.org/10.1002/cnr2.70665

Image Credits: AI Generated

DOI: 10.1002/cnr2.70665

Keywords: gastric cancer, artificial intelligence, deep learning, radiomics, computational pathology, endoscopy, liquid biopsy, lymph node metastasis, immunotherapy, clinical validation, explainable AI, precision oncology