Diseases such as pneumonia, which can be caused by bacterial or viral infections, can be life-threatening if not addressed promptly. Therefore, early diagnosis of lung diseases is crucial. The primary goal of this study is to leverage digital X-ray technology for the automated screening and classification of lung-related ailments. We propose a model architecture for image classification that integrates a convolutional neural network (CNN) with attentional feature fusion. The deep learning model, U-STGSFF, employs attention feature fusion and is structured as an encoder-decoder framework, comprising three essential components: feature extraction, a global attention feature fusion model, and a spatial attention feature fusion module. Additionally, we employ a risk cross entropy loss function to improve diagnostic reliability by lowering false-negative occurrences, thereby enhancing the model’s practical application in medical diagnosis. Experimental results demonstrate that, compared to six mainstream classification models, U-STGSFF achieved superior classification performance in terms of accuracy and reference metrics. Specifically, its classification accuracy for normal, viral, and novel coronavirus pneumonia was 98.19%, 97.86%, and 98.61%, respectively. The data used in the experiments were sourced from a partial pneumonia chest X-ray (CXR) dataset and a partial COVID-19 dataset.