Hybrid Whale Optimization and Canonical Correlation based COVID-19 Classification Approach

The COVID-19 worldwide pandemic has become a great challenge for medical systems. Early detection of coronavirus increases the cure rate and saves the lives of patients. Therefore, a computer-assisted diagnostic tool is necessary to assist radiologists to quickly classify cases of pneumonia as COVID-19 or not. In this paper, we developed an approach that automatically detects COVID-19 disease. This approach relies on an optimized convolutional neural networks (CNNs) to extract deep features from both the computed tomography (CT) scans and the X-ray images. Then, features selection process is applied on CT and X-ray features to eliminate the redundant and irrelevant features and also select more accurate, appropriate, and discriminative ones. The selected CT and X-ray features are combined together using features fusion method to form a final deep feature descriptor for classification. CT and X-ray features are extracted using an optimized CNN architecture model containing thirteen convolutional layers. Sparse Filtering (SF), Features Transformation (FT), Particle Swarm Optimization (PSO), Grey Wolf Optimization (GWO), Hybrid Whale Optimization (HWO), and Entropy Mutual Information (EMI) Techniques are utilized for features selection. Furthermore, various features fusion algorithms are evaluated such as; CNN-based Fusion, Fuzzy Fusion, Canonical Correlation, Neuro Fuzzy, Curvelet based Fusion, and Discrete Wavelet Transform (DWT). Besides, for both augmented and un-augmented CT and X-ray images, we evaluate various optimization algorithms (OA), mini-batch size (M-B) values, and learning rates (LR) to reveal the optimal parameters of the proposed CNN architecture model. Regarding optimization, adaptive moment estimation (Adam) algorithm showed better performance than root mean square propagation (RMS prop), and stochastic gradient descent with momentum (SGDM). The proposed approach achieves a remarkable accuracy (99.43%) for augmented images using Hybrid Whale Optimization (HWO) technique with Canonical Correlation based fusion approach. The proposed approach is superior to traditional pre-trained CNNs and the state-of-the-art classification techniques.