Gastric cancer (GC) ranks as the fifth most prevalent malignancy worldwide, and its early detection is crucial for reducing mortality. The current clinical gold standard relies on histopathological image examination, which despite its diagnostic reliability remains manual, labor-intensive, and time-consuming. Consequently, there has been increasing interest in computer-aided diagnostic (CAD) systems to assist pathologists in decision-making. Although deep learning approaches have demonstrated considerable potential, conventional models often capture only a restricted set of discriminative image features, thereby limiting classification performance. To address these challenges, this study introduces a comprehensive computational framework for automated GC classification using 245,196 histopathological images divided into two categories. The pipeline integrates data preprocessing with multiple filtering strategies, feature extraction, feature selection, and a CNN-Attention-LSTM (CALNet) classifier. For feature extraction, three pre-trained CNN models were fine-tuned under diverse parameter configurations through transfer learning to obtain discriminative representations. Subsequently, Local Interpretable Model-Agnostic Explanations (LIME) was employed as a feature selection strategy to discard non-informative attributes and enhance generalization. Finally, the proposed CALNet model was utilized to ensure reliable and interpretable predictions. Our framework achieved an accuracy of 98.74%, AUC of 98.47%, specificity of 91.00%, precision of 98.15%, and MCC of 91.21%. Experimental outcomes demonstrate that the proposed approach substantially outperforms existing methodologies, underscoring its potential to serve as a clinical decision-support system for gastric cancer diagnosis.