Enhancing Early Gastric Cancer Diagnosis: Real-Time Explainable AI using Multimodal Graph-Attention Networks

Background: Early gastric cancer (EGC) detection remains a major challenge in oncology, especially in resource-limited settings. Artificial intelligence (AI) can improve detection accuracy, but most models are black-box and slow, limiting clinical use. Objective: The aim of the study was to develop and validate Gastric-GAT, a real-time, explainable AI model using endoscopic images and clinical data to detect Early Gastric Cancer (EGC). Methods: In this retrospective study, 400 patients (200 with EGC, 200 with benign findings) from two tertiary hospitals were analysed.. Images were processed through ResNet-50, and clinical features (age, haemoglobin H. pylori status) were integrated using a Multimodal Graph-Attention Network (GAT). Explainability was ensured through SHAP (SHapley Additive exPlanations) values and attention maps. Performance was evaluated using AUC, sensitivity, specificity, inference time, and clinician interpretability ratings. Results: Gastric-GAT achieved an AUC of 0.92 (95% CI: 0.90-0.94), sensitivity of 0.90, and specificity of 0.88. Inference time was 0.85 seconds. SHAP visualizations matched expert interpretations in 92% of cases. Conclusion: Gastric-GAT is a novel, explainable, and fast AI solution that enables real-time EGC risk prediction in clinical settings. Published by Avicena d.o.o. Sarajevo.

Keywords: Early gastric cancer, Explainable AI, Graph Attention Networks, Multimodal fusion, Real-time diagnosis..