RAE-NetB7: Disease Classification from Chest X-Rays using Regularization-Aware EfficientNetB7

Abstract

Coronavirus disease of 2019 (COVID-19) is a global epidemic, impacting millions of people and causing numerous deaths in a short period. However, conventional classification approaches struggle to capture the complex pixel relationships in images due to their inability to handle nonlinearities. To address this issue, this research proposes a Regularization-Aware EfficientNetB7 (RAE-NetB7) approach for disease classification using chest X-ray (CXR) images. The nature of the proposed method allows for dynamic adjustment of regularization strength during training, thereby enhancing the network’s ability to generalize to unseen data. Initially, the benchmark COVID-19 radiography dataset is utilized for classifier’s performance estimation. In the preprocessing step, data augmentation approaches are employed for improving the image counts in an acquired dataset, thereby enhancing the robustness of the classification tasks. Next, Neural Architecture Search Network (NASNet) is utilized to extract meaningful features, and the Dynamic Probabilistic Elitist-based Ant Colony Optimization (DPE-ACO) approach is applied for selecting relevant features. Finally, the RAE-NetB7 approach is proposed for classifying CXR images into different categories such as normal, COVID-19, lung opacity, and viral pneumonia. The proposed method achieves a commendable accuracy of 99.53%, outperforming the COVID-19 Detection using Feature Reuse Residual Blocks and Depth-wise Dilated Convolutional Neural Network(CovidDWNet).