AugCXR Dataset: An Augmented Chest X-Ray Image Dataset for Robust Deep Learning Pneumonia Diagnosis

Chest X-ray (CXR) imaging is commonly used to detect pneumonia, but reliance on expert radiologists may delay diagnosis and increase costs. The shortage of radiologists and the risk of diagnostic errors highlight the need for automated solutions. Deep learning models using Convolutional Neural Networks (CNNs) have shown potential in computer-aided diagnosis of pneumonia from CXR images. Most studies use the publicly available dataset from Guangzhou Women and Children’s Hospital, which includes CXR images of children aged 1 to 5. However, the dataset is imbalanced, with more pneumonia cases than normal. This imbalance may affect model performance and generalizability. This study proposes a geometric data augmentation method using five transformations: rotation, width-shift, height-shift, zoom, and brightness to balance the dataset and improve model accuracy. The proposed Augmented Chest X-ray (AugCXR) dataset was validated using three widely adopted architectures: Improved Visual Geometry Group-13 (IVGG13), MobileNetV2, and EfficientNetV2L. The results demonstrate that the proposed augmentation method enhances classification performance across all three pretrained deep learning models.