Pneumonia Detection in Chest X-Rays Using XGBoost Based Meta-learner with Deep Feature Extractors

Objective: Pneumonia is one of the deadliest diseases on the planet. Proper medical intervention in a critical time-constrained environment often makes the difference in the lives of the patients. The assessment of post-pandemic world has strongly necessitated the implementation of robust healthcare systems, paving the way for the inductance of artificial intelligence (AI) tools in this domain. This work proposes an architecture based upon meta-learning guided ensemble of ResNet50 and Vision Transformer (ViT) models for the classification of images into Pneumonia, COVID-19 and Normal image samples. Methods: Two transfer learning based pre-trained models, namely ResNet50 and ViT have been utilized for feature extraction from input images. The vitality of transfer learning enabled training regiment for the models stems from the extreme scarcity of availability of largely annotated datasets, which is an essential resource for such relevant practises. Feature reduction procedures and its influence on the model’s performance has also been studied. The implementation of Principle Component Analysis (PCA) has been instrumental in providing strategies to filter the most informative dimensions of the concatenated features, resulting in provisions to nullify the effect of noises. The final decision on the type of the image is selected under the aegis of supervision of the transformed features provided by the Extreme Gradient Boosting (XGB) model. Considering it to be an ensemble of multiple decision trees, it produces reliable and robust observations. The results of the operations on the reduced feature dimensions of the image samples make the proposed architecture considerably lighter and suitable for use in a resource-constrained environment. Results: The base-learners ResNet50 and ViT have produced 95.99% and 95.80%, classification accuracies, respectively, and with the proposed architecture the accuracy reaches to 96.19%. Experiments have been conducted on the benchmark Mendeley dataset. which consists of chest X-ray images. Conclusion: The model performs satisfactorily over unknown testing samples, with the loss curves and the confusion matrix showing evidence of minimal bias and overfitting towards any particular class. This approach may be helpful for medical practitioners for the detection of pneumonia and COVID-19 in patients. Code: https://github.com/SutirthaMukherjee97/PneumoniaDetection_CXR_Images .