Design of a lung region spatial attention network with wavelet features for COVID-19 chest X-ray image diagnosis

This paper introduces the Lung Region Spatial Attention Network with Wavelet Features (LRSANW) for multi-class classification of chest X-ray (CXR) images in COVID-19 detection. At the core of LRSANW is the Lung Region Spatial Attention Mechanism (LRSAM), designed to deliver explainable spatial attention by focusing exclusively on relevant lung regions within CXR images. To enhance lesion detection, high-frequency sub-bands from the wavelet-transformed CXR images are incorporated as auxiliary inputs, enabling the model to capture subtle image details. LRSAM is applied to both the original and wavelet-transformed CXR images to guide feature extraction. The feature extractor integrates the Convolutional Block Attention Module (CBAM) and Atrous Spatial Pyramid Pooling (ASPP), combining refined attention mechanisms with multi-scale contextual features to generate highly discriminative image representations. An ensemble strategy further strengthens performance by leveraging three classifiers trained on distinct feature sets: the original CXR image, its wavelet version, and their fusion. Extensive experiments conducted on three public datasets demonstrate that LRSANW outperforms state-of-the-art approaches in average case. The method achieves average performance scores of 92.84% Accuracy, 93.64% Sensitivity, 95.87% Specificity, 92.75% F1-score, 91.44% Precision, and 94.75% Balanced Accuracy. These results indicate that LRSANW can be effectively deployed in real-world scenarios to enable rapid and accurate COVID-19 diagnosis.