TriNet-MoE: One New Neural Network Framework Based on Mixture of Experts for CXR-Based COVID-19 Detection.

As a low-dose portable imaging technology, chest X-ray (CXR) is widely used for the screening of lung diseases, including COVID-19. However, existing deep learning methods for CXR-based COVID-19 detection and common pneumonia recognition still face the challenge of limited discrimination, partly because it is difficult to effectively integrate fine-grained local lesion information with global spatial context. To address this problem, we propose a triple neural network framework based on Mixture of Experts (TriNet-MoE) for CXR-based COVID-19 detection. Specifically, TriNet-MoE integrates ResNet34 for basic local features, ResNet50 for higher-order local semantics, and a Vision Transformer for global context to form a mixture-of-experts model. Then, a three-stage cross-attention mechanism, Cross-Attention Synergy (CA-Synergy), is designed to enable hierarchical feature interaction among experts, including complementary local feature mining between the ResNet dual branches and bidirectional information exchange with the ViT, aiming to obtain more discriminative fusion features. In addition, a dynamic decision mechanism, MoE-Intelligent Linked Feature Collaborator (MoE-ILFC), is introduced to adaptively fuse expert features based on the input content. Experiments on the DLAI3 and COVIDx datasets show that TriNet-MoE achieves accuracies of 98.73% and 98.78%, respectively, consistently outperforming representative baselines under the same experimental settings. Additional statistical evaluations across multiple runs demonstrate stable performance improvements, while cross-dataset experiments further validate the generalization capability of the proposed framework under domain shifts. Visualization analyses, including Grad-CAM activation maps and gating weight distributions, provide qualitative insights into how TriNet-MoE for robust COVID-19 detection.