Recent advances, comparative performance, and data requirements of artificial intelligence-assisted technologies for the diagnosis of gastric cancer.

Authors

Abstract

Early diagnosis of gastric cancer is critical for improving prognosis and increasing the 5-year survival rate. However, gastroscopy and pathological examination are invasive, resource-dependent, and associated with low adherence, while conventional serological and imaging screening methods have limited sensitivity for early lesions. In recent years, artificial intelligence (AI), especially deep learning, has improved the efficiency and accuracy of real-time endoscopic detection, digital pathological recognition, non-contrast CT-based opportunistic screening, and multimodal diagnostic decision support. This review summarizes recent advances in AI-assisted technologies for gastric cancer diagnosis and, in response to the need for a clearer methodological and quantitative framework, compares representative models according to data modality, model architecture, dataset size, validation strategy, and reported diagnostic performance, including AUC, accuracy, sensitivity, specificity, precision/positive predictive value (PPV), and F1 score when available. We further discuss how data size, modality diversity, annotation granularity, class imbalance, and missing modalities influence model performance in real-world multicenter settings. Finally, we propose a structured multimodal AI-clinical decision support system (AI-CDSS) framework integrating CT imaging, endoscopic images and videos, whole-slide pathological images, serological indicators, molecular biomarkers, and clinical reports. Although AI-assisted diagnosis has shown considerable promise, clinical translation still requires prospective multicenter validation, standardized reporting of performance metrics, interpretable model outputs, privacy-preserving data governance, and regulatory-compliant deployment.

Tags

Topics