H-DCA Net: Hierarchical dual-branch coordinate attention framework for multi-scale gastric histopathology diagnosis.

BACKGROUND AND

OBJECTIVE

Early and accurate diagnosis of gastric cancer is crucial for improving patient prognosis. However, conventional histopathological diagnosis is constrained by subjectivity and low efficiency. Current computer-aided diagnosis (CAD) systems often struggle with a "scale mismatch" problem, failing to simultaneously integrate macroscopic tissue structures and microscopic cellular details. To address this bottleneck, this study aims to develop a novel deep learning framework, named H-DCA Net, that mimics the multi-scale diagnostic workflow of pathologists to improve the accuracy and reliability of gastric cancer classification.

METHODS

We propose the Hierarchical Dual-Branch Coordinate Attention Network (H-DCA Net), a framework featuring two heterogeneous branches. A high-level semantic branch (based on EfficientNetV2-S with Efficient Channel Attention) captures global tissue-level patterns, while a local detail branch (based on MobileNetV3-L with Coordinate Attention) focuses on fine-grained cellular features. These complementary features are integrated via a hierarchical attention mechanism. The model was trained and validated on the public GasHisSDB dataset across three different image patch sizes (80 ×80, 120 ×120, and 160 ×160 pixels) using an independent test set strategy.

RESULTS

H-DCA Net demonstrated highly competitive performance across multiple scales. Specifically, on the 120 × sub-dataset, the model attained an Accuracy of 99.27 %, Precision of 99.3 %, Sensitivity of 99.2 %, and Specificity of 99.4 %. Similarly, high performance was observed at 160 × (Accuracy: 99.04 %, Sensitivity: 99.2 %, Specificity: 98.9 %) and 80 × (Accuracy: 98.62 %, Sensitivity: 98.6 %, Specificity: 98.6 %). Comparative analysis confirmed that H-DCA Net outperforms or matches state-of-the-art single-branch, dual-branch, and hybrid architectures, particularly excelling in scenarios with limited field of view (e.g., 80 ×).

CONCLUSIONS

The proposed H-DCA Net, through its synergistic dual-branch architecture and hierarchical attention mechanism, effectively integrates multi-scale pathological information. It serves as a high-precision foundation for an efficient, reliable, and interpretable auxiliary tool, offering robust support for automated gastric cancer screening and helping to reduce diagnostic subjectivity while improving workflow efficiency.