An Ensemble Neural Architecture for Lung Diseases Prediction Using Chest X-rays

Abstract

Accurate diagnostic tools for disease control and treatment options is of immense importance, specially during pandemics, Coronavirus (or COVID) that drew global attention in late 2019. Early detection and seclusion are the cornerstone effective ways to prevent virus spread. Artificial intelligence (AI)-based diagnostic tools for COVID detection have surged dramatically using various diagnostic imaging techniques, among which Chest X-ray (CXR) have been extensively investigated due to its fast acquisition coupled with its superior results. We propose a hybrid, automated, and efficient approach to detect COVID-19 at an early stage using CXRs. One of the main advantages of the proposed analysis is the development of a learnable input scaling module, which accommodates various CXR with different sizes with the ability to keep prominent CXRs features while filtering out noise. Additionally, the suggested method ensembles several learning modules to extract more discriminative representation of texture and appearance cues of CXRs, thereby facilitating more accurate classification. Particularly, we integrated two sets of features (texture descriptors and deeper features) representing a rich concentration of local and global features. In addition to learnable scaling and information-rich features, an ensemble classifier using various machine learning models is used for classification. Our classification module included support vector machine, XGBoost and extra trees modules. Extensive evaluation, supported by ablation and comparison studies, is conducted using two benchmark datasets to evaluate the model’s performance in a cross-validation strategy. Using various metrics, the results document the robustness of our ensemble classification system with higher accuracy of 98.20% and 97.85% for the two data sets, respectively.