Gastric cancer (GC) remains one of the leading causes of cancer related deaths worldwide. Accurate classification of GC in H&E-stained histopathology images is challenging due to the high structural similarity between malignant and non-malignant tissues. As a result, robust automated diagnostic tools are essential for supporting pathologists in clinical decision making. In this study, we proposed a hybrid GC identification framework that combined multi encoder deep feature extraction, explainable feature selection and an attention-based hybrid classifier using the large scale GasHisSDB dataset. Raw images were processed and image augmentation to ensure consistent visual quality. Features were extracted by fine tuning DenseNet121, EfficientNetV2 and Swin Transformer. These encoders captured complementary characteristics. Their outputs were combined into a hybrid feature vector encoding both local texture and global structural information. SHAP based ranking was applied to select the 500 most informative features. It is improving model robustness and generalization. GC classification was performed using a hybrid ViT–SABiLSTM model. The Vision Transformer (ViT) divided the images into patches and applies self-attention to highlight the most relevant regions. Self-attention models long range spatial relationships and enabled the model to focus on subtle morphological differences. The SABiLSTM module captured sequential dependencies in the feature embeddings, refining contextual information and improving discrimination between malignant and non-malignant tissues. The proposed framework achieved 97.95% accuracy. It outperformed existing GC detection methods. These results demonstrated a strong potential to support clinical diagnosis and the development of reliable computer-aided tools.