A Deep Learning Approach to Multiclass Pneumonia Detection in Chest X-ray Images

Abstract

Pneumonia is a significant cause of mortality worldwide, particularly in children under five years, with detection from chest X-ray (CXR) images remaining challenging due to diagnostic errors common in manual radiographic analysis. This study develops a deep learning model for multiclass pneumonia classification in CXR images that is computationally inexpensive and suitable for resource-constrained settings. The proposed approach utilises pretrained models, including EfficientNet, MobileNet, RegNet, and ViT, fine-tuned using the PyTorch framework, with data augmentation and regularisation techniques applied to address class imbalance and overfitting. Using a dataset of 5,863 CXR images classified as normal, bacterial pneumonia, or viral pneumonia, the fine-tuned models achieved high classification accuracy, with ConvNeXt and EfficientNet attaining accuracy scores of 83% and 82% respectively. The findings demonstrate how data augmentation and regularisation significantly improved the models' generalisability, reducing overfitting and improving predictive performance. This work provides healthcare professionals with an efficient tool for multiclass pneumonia detection from CXR images suitable for resource-constrained settings.