ARCHIVES

Pneumonia continues to rank among the principal contributors to global morbidity and mortality, and its management depends on prompt, reliable radiological assessment. Reading chest X-ray (CXR) images by hand is labour-intensive, observer-dependent, and subject to variability between readers, an effect that is amplified where radiological resources are scarce. Although deep learning can supply automated diagnostic assistance, many published models are constrained by weak feature discrimination or excessive computational demand. The present work introduces a hybrid Convolutional Neural Network (CNN) that pairs a lightweight EfficientNetB0 backbone with a Convolutional Block Attention Module (CBAM) to sharpen both spatial and channel-wise feature representation. Training was performed end-to-end on a class-balanced subset of 1,000 CXR images drawn from the Kaggle Chest X-Ray dataset, and the model was assessed on 624 held-out test images. On this test set the proposed hybrid CNN reached 90.71% accuracy, 93.01% precision, 92.05% recall, 92.53% F1-score, 88.46% specificity, and an area under the receiver operating characteristic curve (AUC-ROC) of 0.964, surpassing the VGG-16, ResNet-50, and DenseNet-121 baselines. Allowing every layer—batch normalization included—to update during fine-tuning was essential for adapting the network to the medical domain. Overall, the results indicate that coupling attention-driven feature selection with an efficient backbone yields competitive pneumonia detection well suited to clinical decision support.