Dual-Path Multi-Scale Model Based on Local-Global Feature Aggregation in Gastrointestinal Metaplasia Grading.

Accurate grading of gastrointestinal metaplasia (GIM) is crucial for early gastric cancer detection, but CNNs fail to capture global context, and Transformers often miss subtle pathological cues, leading to misclassification. To address these challenges, we propose a dual-path multi-scale model based on local-global feature aggregation (DMSTNet), which integrates the complementary strengths of CNNs and Transformers for efficient and precise GIM grading. First, we build a dual-branch framework with CNN and Swin Transformer feature extractors. Within this framework, we embed a newly designed multi-scale channel calibration attention (MSCCA) module between the two branches. The MSCCA module serializes feature maps into a multi-scale token pyramid and applies channel attention for feature calibration, enabling effective interaction between CNN-based local features and Transformer-based global modeling for synergistic local-global representation. Second, we propose the hierarchical gated parallel semantic aggregation (PSA-HG) module. Through the collaborative mechanism of hierarchical cross-attention and gated multiple instance learning, PSA-HG dynamically fuses multi-scale features from cross-layer and instance-level perspectives, achieving efficient aggregation and enhanced discrimination of global and local semantics. Third, grading prediction is performed on the fused semantic features. To validate the effectiveness of the proposed DMSTNet, comprehensive comparative experiments were conducted on our self-constructed gastrointestinal metaplasia dataset. Specifically, the DMSTNet improves classification accuracy by 4.19% over typical CNN models and by 2.05% over state-of-the-art Transformer models, achieving an overall accuracy of 85.96±0.28%. These results clearly validate the effectiveness and robustness of DMSTNet in the task of gastrointestinal metaplasia grading.