Dual-branch attention fusion network for pneumonia detection

Pneumonia is a severe respiratory disease caused by bacterial, viral, or fungal infections and remains a leading cause of morbidity and mortality worldwide, particularly among high-risk populations such as infants, young children, the elderly, and immunocompromised individuals. Early and accurate diagnosis is crucial for improving patient outcomes. In this study, we propose a novel Dual-Branch Attention Fusion Network (D-BAFN) based on transfer learning to enhance the accuracy of pneumonia classification in chest x-ray images. The proposed network adopts a dual-branch feature extraction architecture, combining a pre-trained convolutional neural network (ResNet-18) and a structural state-space model (Mamba Vision) to capture complementary local and global contextual features. A self-attention mechanism is further integrated to adaptively weight and fuse the extracted features, thereby improving the representation of key lesion regions. Experiments were conducted on two publicly available chest x-ray datasets: a pediatric pneumonia dataset for binary classification (normal versus pneumonia) and a multi-source dataset containing pneumonia, COVID-19, and normal cases for multiclass classification. Extensive data augmentation, transfer learning, and hyperparameter optimization were employed to maximize model performance. The proposed D-BAFN achieved a top accuracy of 97.78% ± 0.12 on the binary classification dataset, and an accuracy of 97.20% ± 0.15, F1 score of 0.972 ± 0.006, AUC of 0.997 ± 0.001, recall of 0.978 ± 0.005, and precision of 0.966 ± 0.004 on the multiclass dataset. These results highlight the model's effectiveness and robustness, offering a promising AI-assisted diagnostic tool for early and precise detection of pneumonia and other pulmonary diseases in clinical settings.