**5. Discussion**

In the presented D-HL architecture, the softmax layer was replaced with a machine learning classifier. The CNN learning algorithm utilized empirical risk minimization as a method to reduce false positives and false negatives during training. When the backpropagation algorithm reaches the first hyperplane that separates, the training phase ends, and progress generally stops as a result. Another limitation of CNN is that it frequently assigns one output neuron a high value (around +1) while assigning low values to the other neurons (close to 1).

This makes it very difficult to reject implementation errors. Softmax classifiers provide us with likelihoods for each class label. On the other hand, conventional ML techniques help us develop a robust rejection strategy. The generalization ability of CNN is weaker compared to that of SVM DL approaches, in contrast to conventional ML methods, are the least understandable from an AI aspect and are assumed to as a black box.

We performed classification with three classes as well as with two classes. In three classes, pneumonia, normal, and COVID-19 were included and in binary classifications, we used COVID-19 and pneumonia. SVM gave the highest accuracy with three classes. We achieved outstanding accuracy in binary classification with SVM and KNN. In the case of three classes, SVM gave an accuracy of 96.7%, recall of 96.6%, precision equal to 96.7%, and F1-score equal to 96.7%. Table 3 shows the detailed overview of the proposed CNN architecture with all four conventional ML classifiers. Confusion matrixes based on the performance analysis of classifiers are demonstrated in Figure 5. We compared results obtained by the proposed method with other existing works based on different performance metrics. Apostolopoulos et al. [40] used five different CNN pre-trained architectures to classify between three classes(COVID-19, normal, and pneumonia) and gave a sensitivity of 98.66%, accuracy of 94.72%, and specificity of 96.46%. H.S. Maghdid et al. [24] used the transfer learning technique with the AlexNet model and got 94.1% accuracy, 72% sensitivity, and 100% specificity. S.S Khan et al. [41] applied a convolutional auto-encoder to achieve 0.7652 area under a curve. A. Narin et al. [19] also used five CNN models (ResNet50, ResNet101, ResNet152, inception-ResNetV2, and InceptionV3) to perform binary classification of four classes and achieve an accuracy of 96.1%, recall of 91.8%, specificity of 96.6%, F1-score of 83.5%, and precision of 76.5% with COVID-19 & normal binary classification. R. Kumar et al. [42] performed an experiment with DenseNet & GoogleNet and attained an F-score equal to 0.91, AUC: 0.97. Similarly, Makris A. et al. [43,44] used five different pre-trained CNN models and achieved 95% accuracy. Arora, R. et al [45] proposed stochastic deep learning model using ensemble of slandered convolutional models and evaluate developed model on standard dataset contain three classes: COVID-19, normal and pneumonia and attain an accuracy and AUC of 0.91 and 0.97, respectively. A detailed comparison is illustrated in Figure 6 and Table 6.

**Figure 6.** Confusion matrix-based performance analysis of competitive ML classifiers.

Experimental results show that our proposed models outperform all these experiments and achieved 96.69% recall, 96.72% precision, 96.73% accuracy, and 96.71% F-score, as shown in Table 5 and Figure 7.

Certain limitations still apply to our research investigation. The training period for feature extraction was lengthy due to the tiny batch sizes employed to extract the runtime features, which would have typically required a large amount of GPU RAM. Second, the proposed framework must go through a thorough clinical trial before radiologists' professional judgment may be utilized to resolve the patient data.


**Table 6.** Proposed hybrid learning method comparison with existing techniques on publicly available dataset. The Bold shows results of proposed method.

**Figure 7.** Comparative analysis of proposed COV-Net and D-HL with existing literature using accuracy and F-score [19,24,40,43,45].

#### **6. Conclusions**

Well-timed identification of COVID-19 infection is vital to preserve the patient's life and control the further spread of this life-threatening disease. In this study, a new CNN-based scheme for the detection of COVID-19 is proposed. COVID-19 analysis is performed using chest X-ray images containing three categories (pneumonia, COVID-19, and normal). Experimental results proved that the hybrid learning-based framework has shown improved performance compared to other methods. When the proposed framework's performance is compared with the state-of-the-art deep models', it shows that the proposed deep hybrid learning-based method achieved 96.69% recall, 96.72% precision, 96.73% accuracy, and 96.71% F-score for multi-class classification, and for COVID-19 and pneumonia we achieved 99.21% recall, 99.22% precision, 99.21% F-score, and 99.23% accuracy. The proposed COV-Net is less complex than pre-trained and custom-designed networks, and it is feasible to run it on ordinary current PCs. This is conceivable because the algorithm requires fewer resources for both training and execution. Performance analysis is carried out to attain the generalized model and it is likely to assist radiologists in making decisions in their clinical practice.

**Author Contributions:** Conceptualization, supervision, M.M.Z. and R.N.; methodology, M.M.Z.; software, validation, A.F. and A.S.; formal analysis, M.I. and A.A. (Ali Alqahtani); investigation, A.A. (Abdulmajeed Alqhatani); resources, M.Z. and A.A. (Ali Alqahtani); writing—original draft preparation, R.N., A.A.C. and A.F.; writing—review and editing, S.R. and A.A.A.; project administration,

M.Z. and M.M.Z.; funding acquisition, A.A. (Abdulmajeed Alqhatani), A.A. (Ali Alqahtani); and M.I. All authors have read and agreed to the published version of the manuscript.

**Funding:** The authors acknowledge the support from the Deanship of Scientific Research, Najran University, Kingdom of Saudi Arabia, for funding this work under the research group funding program grant code number (NU/RG/SERC/11/3).

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Not applicable.

**Data Availability Statement:** Data are available in publicly accessible repositories which are described in Section 3.1.

**Acknowledgments:** The authors acknowledge the support from the Deanship of Scientific Research, Najran University, Kingdom of Saudi Arabia, for funding this work under the research group funding program grant code number (NU/RG/SERC/11/3).

**Conflicts of Interest:** The authors declare that they have no conflicts of interest to report regarding the present study.
