Screening of COVID-19 Suspected Subjects using Multi-Crossover Genetic Algorithm based Dense Convolutional Neural Network

Fast and accurate screening of novel coronavirus (COVID-19) suspected subjects plays a vital role in timely quarantine and medical care. Deep transfer learning-based screening models on chest X-ray (CXR) are effective for countering the COVID-19 outbreak. However, an efficient screening of COVID-19 is still a huge task due to the spatial complexity of CXRs. In this paper, a dense convolutional neural network (DCov-Net) based transfer learning model is proposed for the screening of COVID-19 suspected subjects using CXR images. A modified multi-crossover genetic algorithm (MMCGA) is then proposed to tune the hyper-parameters of DCov-Net. Majority of the existing COVID-19 diagnosis models are not interpretable as they do not provide any transparency to the users. Therefore, the concept of heat-maps is used to achieve explainability and interpretability. MMCGA based DCov-Net is implemented on a multiclass dataset that contains four different classes. Experimental results reveal that MMCGA based DCov-Net achieves better performance than the existing models. The proposed MMCGA based DCov-Net can be utilized for initial screening of COVID-19 suspected subjects with an accuracy of 99.34 ± 0.51 %.


I. INTRODUCTION
Since December 2019, the entire world has been experiencing the epidemic of a novel coronavirus . It is due to severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) [1], [2]. The laboratory examinations of COVID-19 patients have elevated lactate dehydrogenase, reduced lymphocyte count, reduced or normal leukocyte count, elevated myoglobin, and elevated creatine kinase [3], [4]. The lesser lymphocyte count is found in severe cases. In some cases, the patients have a severe provocative storms that can cause death [5]. Therefore, the early screening of COVID-19 suspected subjects is desirable to decrease the mortality rate and ensure an efficient treatment of patients. For this, chest X-ray (CXR) images are utilized to monitor the common symptoms found in infected patients.
For efficient screening of COVID-19 suspected subjects, the reverse transcription polymerase chain reaction (RT-PCR) is utilized as a standard approach [6]. Due to timeconsuming and strict testing conditions of RT-PCR kit, fast and early testing of suspected subjects is limited. Moreover, RT-PCR kits suffer from high false-negative rates [7].
Recently, deep learning models are widely utilized by researchers for diagnosis of COVID-19 suspected subjects from radiological images. The deep learning-based localization and classification models have been designed for marking COVID-19 infected areas in ultrasound images [6].
Automated classification of COVID-19 suspected subjects from CT was developed by designing an deep 3D learning model. This model semantically produces deep-3D occurrences following the probably infected region in chest CT images to screen the COVID-19 suspected subjects [8]. A patch-based convolutional neural network (CNN) model was proposed for screening of COVID-19 by using CXRs. The patch-based CNN model used a significantly few trainable variables for COVID-19 screening [9].
COVID-19 screening and lesion localization can be achieved by using weakly supervised model from chest CT images. A pretrained UNet model was implemented for lung region segmentation. Thereafter, 3D-CNN was used to recognize COVID-19 suspected subjects [10]. More discriminative diagnosis of COVID-19 suspected subjects from CT Images can be achieved using prior attention residual learning [11].
An infection segmentation deep network (Inf-net) based COVID-19 Infected lung segmentation approach was implemented in [12] for chest CT images. To improve the learning ability, a semisupervised segmentation framework was also utilized. A COVID-19 pneumonia lesion segmentation (COPLE-Net) model was presented in [13] for pneumonia lesion segmentation from chest computed tomography (CT) images.
However, the existing COVID-19 screening models suffer from underfitting issues due to the lack of labeled datasets. Additionally, the existing models have focused on COVID-19 screening as a binary or three-class problem. However, it is found that there exists a similarity between CXRs of COVID-19 suspected subjects and CXRs of patients who have other lung diseases.
CT screening in the initial patient indicates significant sensitivity as compared with RT-PCR [14]. Also, it has been found that when COVID-19 infection is very low in patients (i.e., an early stage of infection), CT-based diagnosis shows significantly more sensitivity as compared to RT-PCR [15]. Due to the high increase in COVID-19 suspected subjects every day, the routine utilization of CT spots put a massive burden on radiologists and possible infection of CT suite hence, many researchers have started using CXR images for initial screening of COVID-suspected subjects [9], [16].
Therefore, the development of a fast and efficient automated COVID-19 screening model from CXR images is still a challenging issue. Compared with the competitive models, the proposed DCov-Net has the following advantages: 1) A dense convolutional neural network (DCov-Net) based transfer learning model is proposed for the screening of COVID-19 suspected subjects using CXR images. 2) Modified multi-crossover genetic algorithm (MM-CGA) is proposed to tune the hyper-parameters of DCov-Net.
3) The concept of heat-maps is also used to achieve explainability and interpretability. 4) Data augmentation is utilized to augment the size of the dataset.
5) The proposed MMCGA based DCov-Net is implemented on a multiclass dataset that contains four classes. The remaining structure of this paper is as follows: The related work is presented in Section II. Section III describes the proposed DenseNet based model. Section IV presents the experimental results and discussion Section V concludes the paper.

II. RELATED WORK
Deep learning models have been extensively utilized in the area of biomedical imaging [17], [18]. These approaches can be used in detecting abnormalities from CXR and CT scan [19], [20]. Due to low ionizing radiation exposure to patients, CXR is preferred over CT scans [21]. Das et al. [22] proposed a pretrained transfer learning architecture for diagnosis of COVID-19 suspected cases. The weights were fine-tuned through a pretrained Xception network. It has achieved 97.41% accuracy. Mahmud et al. [23] implemented a deep learning model namely CovXNet, for the screening of COVID-19 suspected subjects from CXRs. Depthwise dilated convolution was used for the extraction of diverse features from CXR. These extracted features were applied to gradient-based discriminative localization to discriminate abnormal regions in CXR. Their model attained an accuracy of 97.40%. Singh et al. [24] designed a differential evolution-based CNN (DCNN) for diagnosis of coronavirus infections using CXR. The hyper-parameters of CNN were tuned through multiobjective adaptive differential evolution (MADE). It has attained an overall classification accuracy of 94.48%. Togacar et al. [25] used the fuzzy based preprocessing approach for CXR. The preprocessed images are structured using the stacking approach. The efficient features were extracted through MobileNet and SqueezeNet models. A support vector machine (SVM) was then utilized to build the screening model from the extracted features. Their model attained an accuracy of 98.25% on 1357 CXR images.
Behzadi-Khormouji et al. [26] proposed a ChestNet model for detecting consolidation using CXR. ChestNet used a lesser number of max-pooling layers compared with VGG16 and DenseNet121. The accuracy obtained from their approach was 94.67%. Altan and Karasu [27] developed a hybrid model that consists of curvelet transformation, deep learning, and the chaotic salp algorithm for distinguishing COVID-19 imaging patterns from CXR images. The curvelet transformation was applied on a CXR for extracting coefficients. Chaotic salp swarm algorithm (CSA) was used to optimize the extracted coefficients. These optimized coefficients were applied to the EfficentNet-B0 model for screening of COVID-19. Their model was tested on 1596 CXR images. This model has shown 99% accuracy for screening of COVID-19 suspected subjects. Rahimzadeh and Attar [28] hybridized Xception and ResNet50V2 for the screening of COVID-19 from CXRs. 11302 CXRs were used to validate the performance of the hybrid model. The average accuracy obtained from the hybrid model was 91.40%. Ucar and Korkmaz [29] proposed a deep Bayes-SqueezeNet based COVIDiagnosis-Net for diagnosis of coronavirus infected persons. COVIDiagnosis-Net achieved 98.30% accuracy. Tuncer et al. [30] designed a screening model for COVID-19 using CXR images. They used grayscale conversion approach for preprocessing of CXRs. Residual exemplar local binary pattern (ResExLBP) was used for feature iteration. These characteristics were utilized in a SVM to distinguish COVID-19 from healthy subjects. It has attained 99.29% accuracy. However, a limited number of chest images were used for validation purposes. Alqudah et al. [31] employed artificial intelligence approaches for the identification of coronavirus infection in patients using CXR. The imaging features were extracted using CNN. SVM was then used for building COVID-19 screening model. This SVM based model has achieved 95.20% accuracy. Hemdan et al. [32] utilized seven different CNN architectures and implemented COVIDX-Net model for screening of COVID-19 from CXR images. VGG19 and DenseNet201 attained 90% accuracy using fifty CXRs. Wang et al. [33] developed a deep learning model namely COVID-Net for the diagnosis of COVID-19 from CXR. COVID-Net was trained using over 13975 CXR images. COVID-Net has shown the accuracy of 93.3% . It is observed from the literature review that there is scope for improvement in deep learning models for screening of COVID-19 suspected persons.
Yamac et al. [34] utilized a convolution support estimation network (CSEN) for COVID-19 diagnosis. CSEN utilized a noniterative mapping from query to ideally sparse representation coefficient support. Thus, it can handle critical information for class decision in representation-based approaches. CSEN has shown 98% sensitivity and 95% specificity to diagnose COVID-19 suspected cases. Abbas [35] designed a self-supervised super sample decomposition for transfer learning (4S-DT) model. The super sample decomposition of unlabeled CXRs was utilized to form self-supervised learning model. It has shown an accuracy of 99.8% and 97.54% on two different datasets. Ismael and Sengur [36] desinged a transfer learning based SVM (TL-SVM) model. TL-SVM has utilized various transfer learning models to extract the features from CXRs. The extracted features were classified using SVM. It has shown the accuracy of 94.7%. Hasan [37] has designed a hybrid two-dimensional empirical mode decomposition (2DEMD) based CNN (HT-CNN) model to diagnose COVID-19 suspected cases. Intrinsic mode functions were generated using 2DEMD. CNN model was utilized to diagnose COVID-19 suspected cases. It has shown the accuracy of 99.01%. Hasan et al. [38] proposed DenseNet-121 CNN model. Data augmentation was used to augment the dataset. It has shown the accuracy of 92% on CT images dataset. Sarker et al. [39] designed an explainability based COVID-DenseNet model. Transfer learning was used by utilizing the pre-trained CheXNet model. It has achieved an accuracy of 96.49% and 93.71%, for two-class and three-class datasets, respectively. Misra et al. [40] designed an ensemble based multi-channel pre-trained ResNet (MC-ResNet) for CXRs. Three ResNet based variants were used to build MC-ResNet. It has achieved an accuracy of 95.5% and 93.9% for raw and refined datasets, respectively.
It is clear from the literature that the development of early stage recognition of COVID-19 suspected subjects is an open area of research. In particular, the use of a deep transfer learning model can be explored further to improve the sensitivity and specificity of automated screening tools. The existing automated screening models are sensitives to their initial hyper-parameters. Majority of the existing COVID-19 diagnosis models are not interpretable as they do not provide any transparency to the users. Additionally, an efficient automated screening model can distinguish among the other kinds of lung diseases such as pneumonia and lung opacity.

III. PROPOSED MODEL
This section provides the mathematical formulation of the proposed DCov-Net-based screening of COVID-19 suspected subjects. Initially, the labeled dataset is obtained and then the dataset is decomposed into three sets for training, testing, and validation. As the obtained dataset is small in size, data augmentation is used. Thereafter, DCov-Net is proposed. The initial parameters of DCov-Net are then tuned using MMCGA. Finally, MMCGA based DCoV-Net is build to recognize the suspected patient as infected (i.e., COVID-19 (+ve) or pneumonia or lung opacity) or as a healthy person.

A. COVID-19 SCREENING MODEL
The primary objective is to design a model that can automatically screen COVID-19 suspected subjects from their respective CXRs. Figure 1 presents the block diagram of DCov-Net based COVID-19 screening model. Initially, Im-ageNet [41] dataset is used to obtain the transferable coefficients from DenseNet model. Thereafter, data augmentation is applied to the obtained dataset. Fine-tuning of DenseNet is then achieved using the training dataset with a pretrained DenseNet model. The trained DenseNet based COVID-19 screening model can successfully classify a suspected person as infected (i.e., COVID-19 (+ve), pneumonia or lung opacity) or as a healthy person.

B. DENSENET MODEL
In this paper, a pretrained DenseNet model is utilized. Figure  2 demonstrates the diagrammatic flow of DenseNet-based COVID-19 screening model. It contains three transition blocks (T B) and four dense blocks (DB), respectively. In Fig. 2, F i shows a composite function of three operators such as B N , F ReLU , and C V . X i represents the computed feature map by i th layer. Also, i = 1, 2, 3, ..., I. I shows number of layers. Like ResNet [42], DenseNet considers dense links to enhance the relationship between perspective information. The input of every layer includes the input and output of the preceding layer. Thus, it improves the information transmission to reuse the features efficiently. Features of CXRs are dense and redundant in nature. Combined with the constrained dataset, neural network-based model building is susceptible to gradient loss. Therefore, the trained model suffers from the overfitting issue. DenseNet integrates the feature map to overcome this gradient loss issue. Thus, it can efficiently be obtained the potential features from CXRs.
Although, there exist some similarities in the modules of DenseNet and ResNet but there are considerable variations in these modules. It is found that the skip connection of ResNet's residual module can easily aggregate x l−1 and the nonlinear transformation of x l−1 as: Here, x l denotes the output of layer l and F (•) defines a nonlinear transformation. Thus, the model can adapt the input's residual mapping to overcome the gradient loss issue.
Dense block (DB) module concatenates the feature map's connection instead of summation as: At i th layer, there are K × (i − 1) + K 0 convolution maps. K defines convoluted feature maps for every layer. In DenseNet, the input of every layer contains the feature maps of all preceding layers (see Figure 9). DenseNet develops a L layer architecture with L(L+1) 2 connections. In this paper, DenseNet264 is used. In the dense block (DB), both contain different numbers of convolution layers. Table 1 depicts the characteristics of both models.

C. TRANSFER LEARNING
Transfer learning is generally used for fine-tuning deep learning models to prevent local optimization issues. Instead of model initialization with random weights, it acts as an initial point of the new model [43]. Thus, it helps in the fast building of a new deep learning model with significant results even for small size datasets.
In fully connected layer, every neuron is associated with every neuron of the preceding layer. As the input of preceding layer is 2D feature map, therefore, Bernoulli function is utilized. It randomly computes a vector t l−1 following the [0, 1] distribution with a specific probability (p). It can be computed as [44]: Here, the vector dimension is c l−1 . k shows feature maps. o 1 and w 1 indicate the offset and weighting variables of the fully connected layer, respectively. Dropout is also used to prevent overfitting issue. Dropout blocks some neurons with specific probability on random basis. The computed feature map is then weighted and added to the bias offset. Thereafter, softmax activation function is used.

D. ACTIVATION AND LOSS FUNCTION
In this work, the softmax activation function is used at the output layer of the proposed DCov-Net for multiclass classification. The softmax function (P (c, s)) can be defined as: Here, c defines the total number of classes, P (c) and P (s, c) denote the class probability and conditional probability, respectively. s shows the probability of a sample that belongs to c class. P (c, s) can be rewritten as: where The binary cross entropy is utilized as a loss function. It can be evaluated as: VOLUME -, - Here, n shows the number of test CXRs and y i represents the output of the deep learning model. loss approaches 0, when y i approachesŷ i .ŷ i defines the accurate value of the desired output.

E. OPTIMIZATION OF HYPER-PARAMETERS
Proposed DenseNet model is sensitive to its initial parameters. Therefore, it is required to tune the hyper-parameters of DenseNet to achieve the efficient results. Table 2 depicts the search space for hyper-parameters of the proposed DenseNet model. In genetic algorithm, new solutions are searched using crossover operator and it also influence the optimization results. Therefore, many crossover operators are proposed in the literature to efficiently form the new solutions. However, the selection of the efficient crossover operator according to the problem is an open area of research. To resolve this issue, Xue et al. [45] has designed a multi-crossover genetic algorithm (MCGA). Five crossover operators were used to form the new solutions for next iteration. According to their performance during evolution process, a probability was assigned to each of them. Based on their probabilities crossover operator was selected using roulette wheel selection during the evolution process.
In this paper, a modified MCGA (MMCGA) is proposed to reduce the complexity and increase the computational speed. In MMCGA, three crossover operators are used that will be adaptively selected during evolution step. Also, elitist nondominated sorting [46] is used to evaluate the best solution for DCoV-Net. The step by step approach is discussed in Algorithm 1.
iPop (): Initially, population P is composed by randomly generating N solutions based upon their range values as reported in Table 2. iACS: Adaptive crossover selection function (ACS) and its associated parameters are also imitated. Initially, the probability of every crossover operator is 1/Q. Q indicates available crossovers used in MMCGA. iRWS(): Roulette wheel selection is used to select one crossover operator depending upon their respective probability values computed from ACS. iCros(): Depending upon ACS values (a v ), a specific crossover operator is then applied to obtain child solutions. Two parents are randomly selected. iMut(): Uniform mutation operator is also used to form new solutions. All new solutions are then stored in offspring population( P δ ). iCred(): Dominance relationship is then computed between parents and children and corresponding performance is stored in nR and nP . During N/2 th step, P δ is evaluated. iSel(): Elitist non-dominated sorting [46] and crowded distance [47] techniques are then utilized to select N individuals from R (P ∪ P δ ). iDomin(): dominance comparison is utilized to assign the reward or penalty to the selected parent and child solutions.
The reward and penalty of solutions in given step is stored in two matrices RW T H×Q and P L T H×Q , respectively. After every T H number of steps, ACS will be updated using RW T H×Q and P L T H×Q . The evolution process keep repeating until the fitness evaluations do not exceed F e . a v ← iRWS ( P ) 10: Randomly select two individuals as parents: P p

16:
end for 17: k ← k + 1 18: Append nR to k th row of RW T H×Q

19:
Append nP to k th row of P L T H×Q

20:
if k = TH then R ← P ∪ P δ

25:
P ← iSel(R) 26: Select non-dominated solutions in P as T P

27:
Tuned parameters ← T P 28: end while 29: return Optimal parameters The steps of proposed MMCGA are discussed in the succeeding subsections.

1) Multi-objective fitness function
The main objective of MMCGA is to obtain an optimal set of hyper-parameters with maximum performance in terms of sensitivity and specificity. Thus, multi-objective fitness function can be computed as: if P 1 ⊀ R i P 2 ⊀ R i then 13: nR q ← nRnR q + 1 where X is a solution. f 1 and f 2 computes the sensitivity and specificity values, respectively. These are used to obtain optimal hyper-parameter (X) of DCoV-Net

2) Crossover operators
In this paper, three well-known crossover operators namely single-point [47], Reduced surrogate [48], [49], and chaotic crossover operator [50] have been selected. Single-point [47] is selected as it is the most efficient operator concerning the computational effort [?].
Reduced surrogate [48], [49] crossover operator is selected because it avoids the unwanted crossover operations in case of the parents having same genes. This operator initially checks for the individual genes in the parents. It forms a set of possible crossover points where the genes of both parents are different. If there is no such crossover point then no action will be taken. The chaotic crossover operator [50] helps to form well-distributed and well-converged set of Paretooptimal solutions.

3) Reward and penalty
Rewards and penalties are assigned to the solutions by defining two vectors (i.e., nR and nP ) as: Pareto dominance among the solutions is utilized to update nR and nP .

a: Dominated parent
Assume that parent 1 (P 1 ) is dominated by P 2 , then the pareto dominance relationship of every child solution is compared with P 2 . If the child solution is not dominated by P 2 , then set nR q + 1. Otherwise set nP q + 1.

b: Non-dominated parent
Every child solution's pareto dominance is compared with its both parents. If it is not dominated by both parents, then set nR q + 1, otherwise set nP q + 1. Algorithm 2 shows the step by step updation of reward and penalty values.

4) Updation of ACS
ACS is used to update the selection probability of every crossover operator. It is updated after every T H iterations (for more details [51]). Two matrices, RW T H×Q and P L T H×Q , are utilized to store nR and nP values, respectively, by rows til T H iterations. Recent T H's n R and nP values are used to update ACS. To compute the probability for q th (q = 1, 2, ..., Q) operator, summation of q th column of RW and P L is performed, respectively: Here, S 1 q shows numbers of promising solutions obtained by using the q th crossover during last T H iterations. S 2 q shows the non-promising solutions. ACS for q th crossover can be evaluated as: Here, δ = 0.0001 is a small number used to prevent being divided by 0 if the q th crossover is never selected during last T H iterations.
Also, S 4 q is the probability allocated to q th crossover. To normalize the probabilities of each crossover, normalization is achieved as: Here, p q shows the normalized probability for q th crossover. Table 3 shows tuned hyper-parameters obtained from optimization algorithm. VOLUME -, -    parameters. It has shown the testing accuracy of 98.48% when number of iterations are set to be 580.
The proposed DCov-Net model is then trained on the optimized parameters obtained from MMGCA (see Table 3). In Figure 4, training and validation analysis of MMGCA based proposed model is presented. It has achieved the testing accuracy of 99.81% when number of iterations are set to be 580. Therefore, from Figures 3 and 4, it is observed that the optimized hyper-parameters based proposed model has achieved better performance with good convergence speed. MMCGA based DCoV-Net has shown an average improvement of 1.33% as compared to the accuracy obtained from trail and error basis based DCoV-Net.

A. DATASET
In this paper, COVID-19 radiography database [56] is used. The main characteristics of the database is presented in Table 4. Since the database is completely imbalanced in nature, therefore, data augmentation is utilized to balance the database. The data augmentation is achieved by considering random blurring, random cropping, and random rotation. Thus, it is used to increase the minority class CXR images for improving the strength of COVID-19 screening model and overcome the overfitting issue. Table 5 depicts the data augmentation approaches used along with their respective parameter settings. Figure 5 shows the sample view of the obtained CXRs from data augmentation approach. Also, all the images are resized to 224 × 224 resolution.     Table 6 shows number of images used for training, validation and testing of the proposed and competitive models. Training fraction also includes augmented images.

B. PERFORMANCE ANALYSIS
The median and uncertainty values (i.e., median ± IQR × 1.5) are utilized to evaluate the performance of MMCGA based DCov-Net. 65% of dataset is utilized for building the model. 10% of dataset is utilized for validation of proposed model. 25% of dataset is utilized for testing MMCGA based DCov-Net. The confusion matrix-based performance metrics VOLUME -, -  such as accuracy, specificity, sensitivity, area under curve (AUC) and F-measure are utilized for evaluating the performance of MMCGA based DCov-Net.
AUC -receiver operating characteristic (ROC) curve is a well-known measure for classification problems at different threshold settings. AUC represents a measure of separability and ROC defines a probability curve. AUC-ROC evaluates the performance of the model in terms of differentiating between classes. Maximum AUC values define that the model can efficiently differentiate between positive and negative classes. ROC curve is drawn with a true positive rate (TPR) against the false positive rate (FPR). It shows specificity and sensitivity analysis for every possible cut-off for a test or a combination of tests. Figure 6 demonstrates that the proposed MMCGA based DCov-Net achieves better AUC values as compared to the existing models. Therefore, the proposed MMCGA based DCov-Net can provide significantly better performance for the initial screening of COVID-19 suspected subjects. Tables 7 and 8   In this paper, a one-way analysis of variance (ANOVA) and boxplot based statistical analysis is also performed by considering the performance of testing results. For every performance metric, the hypotheses can be defined as: Here, µM i represents various COVID-19 screening models     models. Figures 11 and 12 show ANOVA and boxplot analysis of computed testing sensitivity values. It is clearly found that the sensitivity values obtained from the proposed model are significantly better than the existing models. Figures 13 and 14 demonstrate ANOVA and boxplot analysis of computed specificity values using the testing data. It is observed that the specificity values obtained from MMCGA VOLUME -, -  based DCoV-Net are significantly better than the existing models. Figures 15 and 16 show ANOVA and boxplot analysis of computed testing AUC values. It is clearly found that the AUC values obtained from MMCGA based DCoV-Net are significantly better than the existing models.
Majority of the existing models are not interpretable as they do not provide any transparency to the users. Therefore, the concept of heat-maps is used to achieve explainability and interpretability. Heat-maps allow us to understand how the proposed model achieves classification. The objective is to indicate the critical regions in CXRs so that we can understand that whether the proposed model has correctly classified the image or not. Figure 17 shows the heat-maps obtained using the proposed model. It is found that there are many dissimilarities between the heat-maps obtained from different class of patients and healthy subjects.
Additionally, the proposed model takes on average 8.94 minutes for building the multiclass COVID-19 screening model. Furthermore, it can test up to 500 images in one minute. Thus, DCov-Net can screen the COVID-19 suspected subjects at a reasonable time.

V. CONCLUSION
In this paper, DCov-Net was designed for screening COVID-19 suspected subjects from CXR images. However, DCoV-Net is senstive towards the initial control parameters. Therefore, MMCGA is then proposed to tune the hyper-parameters of DCov-Net. Majority of the existing COVID-19 diagnosis models are not interpretable as they do not provide any transparency to the users. Therefore, in this paper, the concept of heat-maps was utilized to achieve explainability and interpretability. MMCGA based DCov-Net was validated on a multiclass dataset that contains four different classes. Experimental results revealed that MMCGA based DCov-Net achieves better results than the existing models. Therefore, MMCGA based DCov-Net has an ability to be implemented in clinical applications for quick and accurate screening of COVID-19 suspected subjects and potentially other coronaviruses in the future.
In near future, the proposed model can be extended by designing an evolving deep transfer learning model. Since in real-time applications, CXRs may suffer from noise and poor visibility issues, therefore, image preprocessing techniques can also be integrated with the proposed model.