Deep Learning Models for Multi-Class Pneumonia Detection in Chest X-Rays: A Comparative Study of VGG16, MobileNet, and ResNet152

Pneumonia is a major global health threat, with diagnosis becoming increasingly complex due to emerging respiratory viruses such as SARS-CoV-2. This study explores the use of deep learning models—VGG16, MobileNet, and ResNet152—for classifying chest X-ray images into three categories: COVID-19, viral pneumonia, and normal. Models were fine-tuned using transfer learning on a retrospective dataset collected between 2010 and 2021 at a medical center in Guangzhou, China. The dataset contains 5,863 X-ray images (JPEG format), obtained from routine clinical care categorized into pneumonia and normal classes, organized into train, test, and validation folders. Data augmentation techniques, including rotation, scaling, translation, shearing, and flipping, were applied to improve model robustness. ResNet152 achieved the highest accuracy (89%) and showed perfect precision and recall in detecting COVID-19 and viral pneumonia cases, though its performance on normal cases was lower. The superior performance of ResNet152 is attributed to its deep residual learning architecture, which enables the extraction of complex image features while mitigating gradient vanishing. These findings demonstrate the potential of AI-driven systems in supporting pneumonia diagnosis and emphasize the importance of using larger, balanced datasets for improving diagnostic performance in real-world clinical settings, particularly in low-resource environments.