SA-ResNet50: A Four-Class Classification Model for Lung Diseases in Chest X-rays

Deep learning-based medical image analysis plays an increasingly vital role in disease diagnosis, particularly in identifying pulmonary diseases such as COVID-19 and pneumonia from Chest X-Ray (CXR) images. However, conventional Convolutional Neural Networks (CNNs) often encounter performance limitations due to the inherent challenges in CXR images, where lesion areas exhibit blurred edges, irregular shapes, and low contrast with surrounding normal tissue. These characteristics frequently lead to inaccurate localization and the loss of critical information during feature extraction, thereby constraining the performance of classification models. To address these challenges, this paper proposes the SA-ResNet50 model architecture, which embeds a spatial attention module into the residual bottleneck blocks of the ResNet-50 network. This work contributes in several aspects. First, it enhances classification accuracy by strengthening the localization and representation of key lesion areas. Second, on a comprehensive pneumonia dataset, the model achieves an Accuracy of 0.9847, Precision of 0.9826, Recall of 0.9827, F1-Score of 0.9826, and Specificity of 0.9939. Through rigorous comparative experiments against other baseline models and an in-depth ablation study, the effectiveness and superiority of the proposed model are strictly demonstrated. The experimental results not only validate the model's efficacy but, more importantly, offer a novel perspective and a baseline reference for addressing similar challenges in other medical image recognition tasks.