Pneumonia detection in chest x-ray images using an optimized ensemble with XGBoost classifier

Pneumonia is regarded as the top killer of children amongst all other infectious diseases by causing nearly 700,000 deaths to children aged under five every year. The list also includes the elderly (aged 65 and over) in addition to people with pre-existing health issues. However, detection of pneumonia at early stage has a huge impact on saving lives. Chest X-rays imaging technique is typically used for the identification of this disease. Nonetheless, examination of pneumonia is not a straightforward task, not even for an expert radiologist. Consequently, there is always an imperative need for automated diagnosis of pneumonia to assist radiologists confirm their diagnosis. This paper introduces a novel pneumonia classification methodology of Chest X ray images by integrating three optimized pretrained CNN models with XGBoost algorithm where the learned features from the three models are combined and fed as input to the XGBoost classifier. It is employed as an ensemble strategy method to learn the inherent structure of the combined features from each pretrained model and provide the final classification. Furthermore, Bayesian optimization is utilized to unlock the ultimate feature representation of each CNN model by searching for the optimal structural and learning-based hyperparameters, including the number of initial layers that should remain frozen to avoid loss of the acquired generic features as well as modifying the structure of the classification part together with the last activation and pooling layers responsible for delivering the necessary features for the XGBoost classifier. The obtained results demonstrate that the proposed methodology, defined by the modified versions of different CNN models with XGBoost achieved promising performance in comparison with state-of-the art methods with a correct classification rate of 99.15%, 99.53% for precision, 99.30% for sensitivity and AUC equals 0.9972%. © The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2024.