Gastric cancer is a common and aggressive malignant tumor that poses a significant threat to global health, ranking among the leading causes of cancer-related morbidity and mortality [1]. Its clinical prognosis is highly dependent on the timing of diagnosis. Epidemiological studies reveal that early-stage (T1) patients can achieve five-year survival rates exceeding 90 %, while this proportion plummets to below 30 % for those diagnosed at advanced stages (III/IV) [2]. This disparity underscores the urgent need for technologies that can objectively and accurately identify precancerous lesions and early-stage cancer.

The current "gold standard" for gastric cancer diagnosis is the histopathological assessment of H&E-stained whole-slide images (WSIs) by pathologists [3]. However, this manual process is laborious and time-consuming. Furthermore, its high subjectivity is evidenced by inter-observer concordance rates varying by up to 35 % in early gastric cancer diagnosis [5]. This conflict between efficiency and reliability highlights the clinical urgency for robust and efficient intelligent computer-aided diagnosis (CAD) systems [4], [5].

In this context, deep learning, particularly Convolutional Neural Networks (CNNs), has revolutionized medical image analysis [6]. CNNs can autonomously capture multi-scale features, from local textures to global tissue structures, simulating the cognitive workflow of pathologists [7], [8]. This workflow typically involves a multi-scale analysis process: pathologists first scan slides at low magnification (macroscopic view) to identify structural anomalies, then switch to high magnification (microscopic view) to examine cellular morphology for malignant features.

This cognitive transition reveals a fundamental limitation of standard single-branch CNN architectures like ResNet and DenseNet. These models struggle to simulate this collaborative multi-scale analysis within a single framework; they either focus on global features at the expense of cellular details or vice versa. This inherent structural deficiency leads to a "scale mismatch" problem, creating a significant research gap. Therefore, a unified, efficient, and interpretable framework is needed to synergistically integrate pathological features across multiple scales.

To address this "scale mismatch," the research community has explored several avenues. Initial efforts focused on benchmarking and fine-tuning standard architectures like VGG16 and ResNet-50, which established strong performance baselines when combined with transfer learning [9]. Other studies explored fusing features from different receptive fields or training custom lightweight CNNs from scratch [10], [11]. While these efforts laid a solid foundation, they also revealed that simple architectural modifications are insufficient to overcome the inherent scale limitations of single-branch designs.

Consequently, attention mechanisms were introduced to guide networks toward critical regions. Early methods involved complex, multi-stage pipelines, such as pre-segmenting "attention regions" before classification [12]. More recent approaches have focused on designing sophisticated, integrated attention modules like HCCANet [13] or validating the effectiveness of standard modules like the Squeeze-and-Excitation Network (SE-Net) [14]. While effective, these custom designs often increase architectural complexity and computational costs, motivating the search for new foundational architectures with built-in attention capabilities.

This demand coincided with the rise of Vision Transformers (ViT), which offer powerful global context modeling. Research has progressed along two main paths: improving pure Transformer architectures and designing CNN-Transformer hybrids. For instance, CSF-Transformer demonstrated that sophisticated cross-scale fusion within a pure Transformer can effectively address multi-scale challenges [15]. However, the current mainstream approach involves hybrid models, such as the dual-branch DCET-Net, which uses a Transformer to inject global information into a parallel ResNet branch, compensating for CNNs' deficiencies in global context modeling [16]. These Transformer-based works mark a significant evolution, elevating attention mechanisms to the level of global modeling.

In the broader landscape of medical image analysis, architectures have evolved significantly. Advanced CNNs, such as SAFP-Net [17], were proposed to enhance subtle feature detection through optimized feature extraction. To further capture global dependencies, hybrid models like ADCFormer [18] integrated adaptive convolutions with self-attention mechanisms. More recently, State Space Models (SSMs) have gained attention for their linear complexity; notably, SSR-Mam2MIL [19] applied the Mamba2 framework to computational pathology, demonstrating superior efficiency in processing gigapixel whole-slide images.

Specifically, within the domain of gastric and gastrointestinal pathology, recent studies published in Pathology - Research and Practice have also emphasized the utility of computational methods for morphological analysis. For example, Martos et al. focused on the optimized detection and segmentation of nuclei in gastric cancer images, highlighting the critical role of stain normalization and artifact removal in ensuring diagnostic accuracy [20]. Furthermore, Dong et al. proposed an interpretable machine learning framework for gastrointestinal stromal tumors (GIST), which utilized a multi-scale approach and SHAP analysis to align the model’s decision-making with pathologists' diagnostic criteria [21].

Despite these promising advancements, effectively integrating macroscopic tissue structures and microscopic cellular details—the "scale mismatch" problem—remains a critical bottleneck, particularly in creating an end-to-end framework that is both highly accurate and clinically interpretable. Consequently, this paper proposes the Hierarchical Dual-branch Coordinate Attention Network (H-DCA Net), a novel framework specifically designed for gastric histopathological image classification. Constructed upon a heterogeneous dual-branch structure, our method aims to organically integrate high accuracy, efficiency, and interpretability within a single end-to-end model. The objective is to achieve performance comparable to complex ensemble models while avoiding computational redundancy.

Heterogeneous Dual-Branch Architecture: We introduce a novel network that mimics pathologists' workflow. A global branch (EfficientNetV2-S) captures tissue-level structures, while a local branch (MobileNetV3-Large) focuses on cellular details. This parallel design enables comprehensive analysis by synergistically integrating macroscopic and microscopic features for superior classification accuracy.

Hierarchical Attention Mechanism: A task-specific attention strategy is proposed. The global branch uses Efficient Channel Attention (ECA) to select informative tissue patterns, while the local branch employs Coordinate Attention (CA) for precise localization of critical cellular regions. This enhances the model's focus on diagnostically relevant features at each scale.

State-of-the-Art Performance and Robust Validation: Through extensive experiments on the public GasHisSDB dataset, our H-DCA Net demonstrates superior performance, achieving up to 99.27 % accuracy. Ablation studies and further validation on the TCGA dataset rigorously confirm the effectiveness and strong generalization capability of our proposed architecture and attention mechanism.