Artificial Intelligence for Tumor Tissue Detection in Stomach Cancer: A Retrospective Algorithm Development and...

Background: Gastric cancer remains one of the leading causes of cancer-related mortality worldwide, underscoring the need for more effective diagnostic strategies. This study aims to use annotated digitized histological slides of gastric cancer and precancerous lesions to develop artificial intelligence algorithms for the diagnosis of gastric lesions. Materials and Methods: We developed a deep learning tool using a training cohort of 970 digitized gastric biopsy slides. Convolutional neural networks (CNNs) were trained for histological recognition and ICD-10 code assignment. The model was validated on an independent test cohort of 250 cases, with expert consensus as the reference standard. Performance was assessed using sensitivity, specificity, and Cohen's kappa. Survival analysis used Kaplan-Meier, log-rank tests (SPSS 16.0; p < 0.05 significant). Results: Analysis of the training cohort led to a scoring system predicting fatal outcomes based on age and morphology (high-grade component > 70%, ulceration, absence of metaplasia/dysplasia). High-risk patients (4-5 points) had significantly worse survival than low-risk patients (0-3 points) (Log Rank = 14,754; p < 0.0001). One-year survival was 71% (low-risk) vs. 40% (high-risk); mean survival was 19.2 vs. 11.3 months. In the test cohort, the AI algorithm demonstrated 79.6% sensitivity and 86.7% specificity (p < 0.0001) for differentiating malignant from benign gastric lesions. Conclusions: A system combining AI-based analysis with a prognostic scoring model has been developed to reduce diagnostic errors and improve risk stratification in gastric cancer pathology.