CAXNet: Coordinate Attention-Enhanced ResNeXt with Multi-Level Feature Fusion for Multiclass Chest X-Ray Classification

Rapid and reliable COVID-19 detection from chest X-rays can aid clinical screening and radiologists' decisions. Due to overlapping radiographic features in COVID-19, lung opacity, viral pneumonia, and normal chest X-rays, multiclass categorization is difficult. This study offers CAXNet, a Coordinate Attention-enhanced ResNeXt COVID-19 chest X-ray classification model. This architecture enhances ResNeXt50-32x4d by preserving early and intermediate residual stages and incorporating Coordinate Attention modules for improved spatial feature representations. Before categorization, attention-enhanced low- and intermediate-level characteristics are fused using multi-level feature fusion. The model was tested using the COVID-19 Radiography Database for four-class classification: disease, lung opacity, normal, and viral pneumonia. The experimental findings demonstrate that CAXNet exhibited 92.00\% accuracy, 92.15\% precision, 92.00\% recall, 92.00\% F1-score, 0.8852 Matthews Correlation Coefficient, and 0.9329 ROC-AUC on the test set. The confusion matrix and t-SNE-based feature visualization show that the proposed model learns discriminative representations, notably for COVID-19 and Viral Pneumonia. The results show that coordinate-aware attention and multi-level feature fusion enhance ResNeXt-based chest X-ray image classification models.

COVID-19 classification; chest X-ray; deep learning; ResNeXt; Coordinate Attention; feature fusion; medical image classification