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

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Pathology

Background:

• Accurate grading of gastrointestinal metaplasia (GIM) is vital for early gastric cancer detection.
• Convolutional Neural Networks (CNNs) struggle with global context, while Transformers may miss subtle pathological details, leading to misclassification.
• Existing models lack a synergistic approach to integrate local and global features effectively for GIM grading.

Purpose of the Study:

• To propose a novel dual-path multi-scale model (DMSTNet) for efficient and precise GIM grading.
• To leverage the complementary strengths of CNNs and Transformers for improved pathological feature representation.
• To enhance the accuracy of GIM grading for better early gastric cancer detection.

Main Methods:

• Developed a dual-branch framework utilizing CNN and Swin Transformer feature extractors.
• Introduced a multi-scale channel calibration attention (MSCCA) module for synergistic local-global feature interaction.
• Implemented a hierarchical gated parallel semantic aggregation (PSA-HG) module for dynamic multi-scale feature fusion.

Main Results:

• DMSTNet achieved an overall accuracy of 85.96±0.28% on a self-constructed GIM dataset.
• The model demonstrated a 4.19% improvement in classification accuracy over typical CNN models.
• DMSTNet outperformed state-of-the-art Transformer models by 2.05% in GIM grading accuracy.

Conclusions:

• The proposed DMSTNet effectively integrates local and global features for robust GIM grading.
• DMSTNet offers superior performance compared to existing CNN and Transformer models in pathological image analysis.
• The model shows significant potential for improving early gastric cancer detection through accurate GIM assessment.