Detection of Pneumonia from Chest X-ray Images Using Transfer Learning and Ensemble Learning

Abstract

Pneumonia, characterized by lung inflammation and caused by various strains of bacteria and viruses, consistently ranks as one of the top three causes of death in Singapore. Despite chest X-ray being the standard imaging test for pneumonia, its diagnostic accuracy has limitations due to factors like low specificity and subjective variance among radiographers. This research aims to enhance the diagnosis by predicting pneumonia based on chest X-ray images sourced from a Kaggle dataset with 5856 images, primarily from pediatric patients at the Guangzhou Women and Children’s Medical Center. A novel approach modifies the Kaggle data split to introduce more validation data and employs data augmentation techniques such as random flips and rotations to stabilize the model. Convolutional Neural Networks (CNNs) form the core of the prediction methodology. Simple CNNs are evaluated, followed by transfer learning models using ImageNet architectures. The models are assessed based on their weighted F1 scores, with a special emphasis on recall to ensure critical cases aren't missed. Ensemble techniques, such as voting, are explored to further enhance prediction capabilities and robustness. The paper culminates by comparing these models against existing frameworks and provides insights into their potential application in real-world medical diagnostics to further enhance the medical field.