Computational Intelligence-Based Method for Automated Identification of COVID-19 and Pneumonia by Utilizing CXR Scans.

Chest X-ray (CXR) scans are emerging as an important diagnostic tool for the early spotting of COVID and other significant lung diseases. The recognition of visual symptoms is difficult and can take longer time by radiologists as CXR provides various signs of viral infection. Therefore, artificial intelligence-based method for automated identification of COVID by utilizing X-ray images has been found to be very promising. In the era of deep learning, effective utilization of existing pretrained generalized models is playing a decisive role in terms of time and accuracy. In this paper, the benefits of weights of existing pretrained model VGG16 and InceptionV3 have been taken. Base model has been created using pretrained models (VGG16 and InceptionV3). The last fully connected (FC) layer has been added as per the number of classes for classification of CXR in binary and multi-class classification by appropriately using transfer learning. Finally, combination of layers is made by integrating the FC layer weights of both the models (VGG16 and InceptionV3). The image dataset used for experimentation consists of healthy, COVID, pneumonia viral, and pneumonia bacterial. The proposed weight fusion method has outperformed the existing models in terms of accuracy, achieved 99.5% accuracy in binary classification over 20 epochs, and 98.2% accuracy in three-class classification over 100 epochs.

1. Introduction

Coronavirus disease 2019 (COVID-19) pandemic has been different in terms of severity in respiratory syndrome coronavirus [1]. The first case of COVID-19 occurred in the Wuhan city of China in December 2019 [2, 3]. The ever-increasing number of infectious patients globally since December 2019 made it declared as a pandemic [4]. Dated 26 December 2020, confirmed count of COVID-19 rises to 29356292 in various countries, it is not only the scenario, this count keeps on increasing day by day, and approximately 930260 people have lost their lives [5]. The increasing number of COVID-19 patients with each passing day expresses indirectly the need for early identification of such patients. Among all image modalities, chest X-ray (CXR) is a promising diagnostic tool for monitoring treatment outcomes of infected patients. Although CXR is affordable and most popular, the opinion of an expert radiologist is highly recommended for distinguishing COVID-19 patients from similar lung diseases like pneumonia. The literature survey reveals many attempts made by different researchers for solving numerous image processing tasks on X-ray images in the medical domain. Many cognitive and complex problems have been successfully solved using popular tools like machine learning and deep learning [6-8].

The proposed work uses transfer learning, and it is one of the techniques that comes under deep learning and takes advantage of past learning experience in the form of weights, for the early and efficient diagnosis of COVID-19 using CXR images. Due to nonavailability of large COVID-19 CXR image dataset, the proposed work has chosen not to provide training to the model from scratch, instead used transfer learning for providing the training model using the existing well-performed generalized models [9-12]. Additionally, learned weights from generalized images in the form of learning through transfer alleviate the assumption that the images used for training should be identically distributed and should not exhibit any dependency with the testing image data [13]. Number of attempts have been made for the identification of COVID-19 by employing CT images which makes wide use of variants of deep learning techniques. External validation and internal validation accuracy has been reported in literature by one of the models using deep learning technique for discovering COVID-19 which is 73.1% and 82.9%, respectively [14]. Another model for classifying COVID-19 from other similar kinds of infections like pneumonia caused by viral influenza and no infection category using computed tomography (CT) was introduced by Xu et al. [15] and achieved an accuracy of 86.7%. Zheng et al. [16] presented another model for detecting the COVID-19 by achieving 90.1% accuracy using CT scans.

Only few researchers have reported the development of an efficient and robust artificial intelligence-based classification model which claims low bias and low variance for COVID-19 by utilizing CXR images [17-22]. In [17], authors proposed a model for distinguishing the CXR and CT images into normal classes as coronavirus and bacterial pneumonia with an accuracy of 92.18%. A deep learning-based model with binary classifiers has been proposed by Ozturk et al. [18] for classifying COVID-19 and healthy patients into different groups. The researchers proposed another multi-class classifier for categorizing COVID-19/pneumonia/healthy patients and obtained the accuracies of 98.08% and 87.02%, respectively. A method for generating the synthetic CXR images with auxiliary classification model based on generative adversarial network was proposed by Waheed et al. [19]. This binary classifier has obtained an accuracy of 85% for CXR. As machine learning and deep learning-based models always perform better in the availability of large datasets, hence it is been reported that by augmenting original images and integrating both, the version accuracy has been increased to 95%. For the purpose of measuring the performance, one of the models was trained with original images and another was trained using synthetic as well as original images [19]. In order to predict pediatric pneumonia using CXR images, [23] has proposed a transfer learning-based technique and used accuracy and recall as performance evaluation parameters with values obtained as 96.4% and 99.62%, respectively. Another learning-based method variant of deep learning, RCNN mask version, has been proposed by Jaiswal et al. [24] to identify and localize the pneumonia in CXR. In this work, training and testing have been provided on the three types of CXR, namely, normal, lung opacity, and abnormal. Che Azemin et al. have made an attempt to predict the COVID-19 using CXR images and deep learning-based models. The binary classifier used was the same as [25].

Ucar and Korkmaz [20] have proposed the COVID diagnosis-Net to classify three-class CXR labeled as normal, pneumonia, and COVID-19 viral infection. COVIDiagnosis-Net has a test accuracy of 98.26% and is based on deep SqueezeNet with Bayes optimization. Oh et al. statistically investigated the possible CXR COVID-19 markers by studying various important parameters related to lungs and COVID-19 CXR images. The attempt has been made to analyze the important differences mathematically. The four distinct classes have been analyzed using a local patch-based approach resulting in 88.9% accuracy claimed to be the highest [21]. Apostolopoulos et al. classification is based on seven different classes where the concept of convolutional neural network (CNN) has been used from the beginning without taking the benefits of pretrained models. The results showed an accuracy of 87.66%. Two class discriminations, class one COVID-19 vs class two non-COVID-19, have been conducted, and values recorded for accuracy, sensitivity, and specificity are 99.18%, 97.36%, and 99.42%, respectively [22]. Multi-class and hierarchical learners have been developed by Pereira et al. [26] for identifying COVID-19 using X-ray images of chest, and performance parameters evaluated are macro-avg and F1-score with values of 0.65 and 0.89, respectively. A method has been proposed by Rahimzadeh and Attar [27] for multi-class classification of COVID-19 vs pneumonia vs normal images. This method has used accuracy as a performance parameter and average highest value, reported as 99.50%, for the case of COVID-19 and an average accuracy of 91.4% all the three classes. According to the recent reports [28-32], a lot of research is underway to develop chest X-ray classification models for better COVID-19 using a larger dataset. In [33], a three-phase transfer learning detection model is proposed to improve the detection accuracy by using stationary wavelet from CT scan images. As many deep learning models are developed in healthcare domain for identifying the disease in the lesser time [34], but still lot of improvements are required. CovXR: Automated Detection of COVID-19 Pneumonia in Chest X-Rays through Machine Learning—this model achieves an accuracy of 95.5% and an F1-score of 0.954 [35]. In [36], study applied support vector machine (SVM), k-nearest neighbor (K-NN), and deep learning convolutional neural network (CNN) algorithms to classify and detect COVID-19 using chest X-ray radiographs. In [37, 38], features extracted by convolutional neural network (CNN) models were united with conventional gray-level co-occurrence matrix (GLCM) and local binary pattern (LBP) algorithms.

The downside of these proposed models is the lack of systematic approaches including accurate data preprocessing and use of exhaustive augmentation techniques, although high classification accuracy has been reported. The gaps found in the recent models pushed us to come up with a classification system which can surpass and provide better solutions to remedy the deficiencies. Various challenges in this domain are as follows: (1) the cells available in the respiratory tracts and predominantly the lung tissues are more prone to COVID-19 virus attacks; the images of thorax can be used to detect the virus without an intervention of third-party test kit. If the virus is in its initial stages, CXRs are not very promising whereas chest CT scans can provide a clear picture even before the onset of symptoms. (2) Testing takes longer. It is really difficult to access the patient particularly at the time of initial stages of the development of such viruses. Radiologists too require long hours for physical examination of CT scans and CXR of many patients.

Therefore, there is a need for an automated system for patient data analysis without radiologist's intervention. In this work, authors have proposed and verified the protocol of reusing and fusing the weights of existing pretrained models on generalized dataset. A model has been developed for identifying the COVID-19 using CXR images. The features which display significantly different behaviors other than normal that are identified in the images can precisely be detected using the proposed approach.

This is the only work of weight fusion utilization for classification of healthy, COVID, and pneumonia sufferers using CXR to the best of the author's knowledge. The contributions of the proposed work are as follows: (i) a binary and multi-class classification transfer learning model has been proposed for classifying healthy, COVID, and pneumonia cases that outperformed the existing proposed models. (ii) Weights of existing pretrained model VGG16 and InceptionV3 have been taken in the proposed transfer learning model. (iii) A new base model has been developed using pretrained models (VGG16 and InceptionV3) by integrating the last FC (fully connected) layer as per the number of different categories for the classification of chest X-rays in binary and multi-class classification using transfer learning approach. (iv) Finally, combination layers are designed by integrating the FC layer weights of both the models (VGG16 and InceptionV3).

The paper is organized as follows: Section 2 describes materials and methods used for the proposed work. The results obtained from different experiments on the proposed model are presented and discussed in Section 3. Finally, Section 4 concludes the paper.

2. Material and Methods

2.1. Material

For the detection of COVID-19, chest imaging is used nowadays. Through the analysis of chest imaging, medical team can more precisely grasp the imaging modal features of COVID-19 cases, such as numerous minor patchy shadows and interstitial variations in the early phase, which are obvious outside the lungs. To the best of the authors' knowledge, COVID dataset used is the open-access benchmark dataset in terms of the number of COVID-19-positive patient cases. To generate the dataset, we combined 4 different publicly available data repositories: COVID-19 Radiography Database (Kaggle) [39], Italian Society of Medical and Interventional Radiology Public Database [13], Chest X-Ray Images (Pneumonia) (Kaggle) [40], and Joseph Paul Cohen-covid-chestxray-dataset (GitHub) [41]; number of samples per category are normal/healthy—1583 (testing—317, training—1266), pneumonia—4273 (testing—855, training—3418), bacteria pneumonia—1860 (testing—372, training—1488), viral pneumonia—2413 (testing—482 training—1931), and COVID-19—714 (testing—143 and training—571). All the images are in the Portable Network Graphics (PNG) file format and with a resolution of either 1024-by-1024 pixels or 256-by-256 pixels. The dataset distribution is relatively balanced. Before passing the images into a pretrained model for feature extraction, we resized all images to a size of 224 × 224 × 3 pixels. All images were normalized according to the pretrained model standards. If the training validation set is segregated by a ratio of 0.3, serious overfitting chances were confronted. Therefore, a division ratio of 0.2 is most appropriate and same is carried forward in the experimentation. Since the proposed methodology uses transfer learning approach, so it does not require large set of data as large part of the model is already trained on huge dataset.

2.2. Methods

In this study, the transfer learning technique was applied using ImageNet data to resolve the problems of inadequate data and preparation time. The weights trained on ImageNet were downloaded for each model. The feature maps were treated as input size in the applied layer training process. CNN model has been designed by using the architecture and weights of existing pretrained models VGG16 and InceptionV3. Base model has been created from a predesigned model, and a fully connected (FC) layer is appended as per the number of classes based on data for binary classification class 1 served by healthy images and class 2 served by combination of COVID-19 and pneumonia. For multi-class classification, class 1 served by healthy images, class 2 served by COVID-19 X-ray images, and class 3 served by pneumonia X-ray images. To check an efficiency of the designed classifier, another multi-class category has been considered.

2.3. Proposed Work

In this work, transfer learning has been used and the learned features are reused or transferred to another targeted network which is to be trained on a new dataset. Transfer learning advances learning by transferring information from related tasks that have been learned, i.e., transferring learned and trained parameters to a fresh model to support with its training phase [10]. The architecture of deep learning models is complex and data dependent, requiring much data to train them. COVID-19 data samples that are published online are less in size, making it tough to train a deep learning model from start to end. Transfer learning can facilitate the training of such a small sample dataset to achieve the research purpose. The work algorithm for the proposed methodology is described in detail in Algorithm 1.

Transfer learning is categorized as two common schemes: develop model and pretrained model. In the develop model approach, the model is evolved that can outperform naive models for ensuring that certain feature learning has been carried out, subsequently reusing and tuning the developed model for achieving the desired task. In the latter case, a pretrained model is chosen from the models at hand for fine-tuning and reuse. Figure 1 illustrates the proposed methodology. In the proposed work, the pretrained model approach has been used which works as follows:

Figure 1

Flow of the proposed methodology and architecture.

Base Model—A pretrained model has been selected from existing models. The top pretrained models available for image classification are VGG16, ResNet50, InceptionV3, and EfficientNet. In the proposed work, VGG16 and InceptionV3 have been utilized as these have been trained on ImageNet (1000-class photograph classification competition) and both the networks have 3 fully connected layers/dense layers. Though improved version of VGG19 is available for image classification, main shortcoming is that it is quite huge network in terms of the number of parameters to be trained and VGG16 does almost as well as the VGG19 so current study uses VGG16 for the current research.

Reuse Model—The pretrained model can then be used as the starting point for classifying CXR images based on pneumonia and COVID-19. In the proposed approach, first the model has been trained using VGG16 and InceptionV3 separately on the dataset. VGG16 is a convolutional neural network model proposed in [42]. The model achieves 92.7% top-5 test accuracy in ImageNet, which is a dataset of over 14 million images belonging to 1000 classes. The input to cov1 layer is of fixed size 224 × 224 RGB image. The image is passed through a stack of convolutional (conv.) layers, where the filters were used with a very small receptive field: 3 × 3 (which is the smallest size to capture the notion of left/right, up/down, center). The convolution stride is fixed to 1 pixel; the spatial padding of convolution layer input is such that the spatial resolution is preserved after convolution, in which the padding is 1-pixel for 3 × 3 convolution layers. Spatial pooling is carried out by five max-pooling layers, which follow some of the convolution layers. Max-pooling is performed over a 2 × 2 pixel window, with stride 2.

InceptionV3 was trained using a dataset of 1,000 classes from the original ImageNet dataset which was trained with over 1 million training images. Inception networks have proved to be more computationally efficient, in terms of both the number of parameters generated by the network and the economic cost incurred. An auxiliary classifier is a trivial CNN implanted between layers throughout training, and the loss experienced is added to the main network loss. In GoogLeNet, auxiliary classifiers were used for a deeper network, whereas in InceptionV3 an auxiliary classifier works as a regularizer [43].

The base model has been created using a pretrained model and adding the last FC layer as per the number of classes used in this work, and then, the model has been trained on the given dataset. At last, the model is saved in the hierarchical data format (HDF/H5) for both pretrained models that have been used in the proposed work. The saved H5 files are fetched and used for merging of dense layer weights in both the pretrained models.

The formula involved in calculating the output size from each convolution layer used in our work is given below:

where N is input dimension, f is the size of kernel used, and S is no. of strides in kernel. For values N  = 224, f  = 3 (kernel size—3 × 3), S  = 2, and keeping it in equation (1) we get below shape as an output 111 × 111.

The softmax function is used as an activation function which has been employed in this proposed work. The classification problems can be handled by the softmax function which is also a kind of sigmoid function and nonlinear function. It is usually utilized for handling multiple classes. It would pack the outputs between 0 and 1 for each class and also divide by sum of the outputs. To attain the possibilities for defining the class of each input, it is preferably employed in the output layer of the model.

Fine-Tune Model—This is the optional phase in which the model needs to be refined or adapted on the input-output data that is available for the task of interest. The final obtained model is fine-tuned using the adaptive moment estimation (Adam) optimizer. Adam optimizer provides much higher performance by giving an optimized gradient descent. SDG optimizer was also used in experimentation, but results were faster in Adam. The new model in the proposed work has been fine-tuned using the steps which are as follows: base model is unfreeze, recompiling of the model is done, and training of the model is done on training data.

3. Result Analysis and Discussion

In this, experimental results have been achieved by integrating the features that are extracted by CNNs, employing transfer learning and classifiers. Total number of experiments is 72 × 2 = 144 which have been executed. The system infrastructure used was an Intel i5, 7th gen, 8 GB of RAM, with Windows 10 operating system without a graphical processing unit (GPU).

3.1. Performance Analysis of the Proposed Model

The performance analysis of the proposed model has been carried out using recall, precision, F1-score, and support. F1-score can be determined as a harmonic mean of sensitivity as well as of precision which is used to determine an overall quality of the model.

The training time, extraction time, and test time have been analyzed in addition to the evaluation metrics. The training time is used to represent the duration it takes to provide training to the classifier to the moment it is likely to carry out the classification. Extraction time is used to measure how long the given model would take to produce the attribute vector after getting the CXR image. In addition, test time is defined as the duration it would take by the classifier for the prediction of image class from the moment it has received the attribute vector. Therefore, for the building of a model, training time plays an important role. Hereafter, the extracting time and testing time are more significant whose sum can be used to represent the classification time that is the time duration between getting the CXR and obtaining its class. The proposed model has been examined on binary classes, three classes, and four classes. Weights fused using pretrained VGG16 and InceptionV3 have given noteworthy results for binary and three-class classification. Table 1 represents the experimental results for binary classification of unhealthy and healthy patients of the proposed model from CXR images. The quantitative results have been shown in terms of recall, F1-score, and precision. From the experimental results, it has been noticed that the proposed model is able to classify between unhealthy and healthy patients well. The batch size is 10, no. of epochs is 20, and learning rate is 0.0005. Total parameters used in model are 16,812,35 and trainable parameters: 2,097,665.

Table 1

Binary classification of unhealthy vs healthy patient's results.

Category	F1-score	Recall	Precision	Support
COVID-19/pneumonia	0.93	0.88	1.00	24
Healthy	0.99	1.00	0.97	101

Furthermore, Figures 2(a) and 2(b) illustrate the plot of accuracy and loss of training and testing of the proposed model, respectively, over 20 epochs for binary classification of unhealthy (COVID-19/pneumonia) vs healthy patients. It is evident from the figure that 0.0563 test loss has been achieved and test accuracy achieved is 99.5% in two-class classification. Figure 3 illustrates the visual results of the binary classifier on two classes to classify healthy (no findings) and unhealthy (COVID-19/pneumonia) patients using CXR images. Table 2 lists the classification of healthy, COVID, and pneumonia diseases using the proposed model.

Figure 2

Plotting of (a) accuracy and (b) loss of training and testing of proposed model over 20 epochs for binary classification of unhealthy (COVID-19/pneumonia) vs healthy patients.

Figure 3

Classification of (a–c) healthy (no findings) and (d–f) unhealthy (COVID-19/pneumonia) class testing results.

Table 2

Classification of healthy, COVID-19, and pneumonia results.

Category	Precision	Recall	F1-score	Support
COVID-19	0.94	0.86	0.89	25
Healthy	0.84	0.96	0.93	90
Pneumonia	0.88	0.85	0.86	90

In addition to this, Figure 4 illustrates the plots of the accuracy, loss of training and testing of proposed models over 100 epochs for classification of healthy, COVID-19, and pneumonia diseases. In this, test loss achieved is 0.00453 and test accuracy achieved is 98.2% in three-class classification (healthy, COVID-19, and pneumonia classes).

Figure 4

(a and b) Plotting of accuracy and loss of training and testing of proposed model over 100 epochs for classification of healthy, COVID-19, and pneumonia.

Figure 5 shows visual results of the classification of healthy, COVID-19, and pneumonia class testing results, whereas Figure 6 illustrates the plot of accuracy and loss of training and testing of the proposed model over 20 epochs to classify healthy, COVID-19, bacterial pneumonia, and viral pneumonia. In this, test loss achieved is 0.6061 and test accuracy achieved is 62.5% in four-class classification (healthy, COVID-19, bacterial pneumonia, and viral pneumonia).

Figure 5

Classification of (a–c) healthy, (d–f) COVID-19, and (g–i) pneumonia class testing results.

Figure 6

(a and b) Plotting of accuracy and loss of training and testing of proposed model over 20 epochs to classify healthy, COVID-19, viral pneumonia, and bacterial pneumonia.

Transfer learning approach has the advantage of shrinking the time for training for a neural network model and contributing to lesser generalization error. While using this approach, we can either utilize weights or features extraction. Here in proposed work, we have utilized combined weights of two pretrained models (VGG16 and InceptionV3) and also optimization technique is used to propose approach, so resulting into the better outputs in terms of accuracy, precision, recall, and support parameters. The positive cases of COVID-19 are mostly reliant on immune identification technology, nucleic acid detection, clinical symptoms, and epidemiological history. Most of the schemes for the detection of COVID disease have many limitations such as costs, shortage of testing kits, equipment dependence, shortage of expert medical workers, time required, and intra-operator and inter-operator variabilities, making the cumbersome diagnostic procedures especially in a pandemic [44]. The assisting examinations like identification of nucleic acid detection technologies are suffering from false-negative rates that cannot be ignored too [45].

Globally, COVID-19 pandemic is declared as a medical emergency. Therefore, there is a need for a noninvasive, cost-effective, fast, user-friendly, and smart diagnostic scheme for the early diagnosis of diseases and rapid screening with less human intervention. Timely and early diagnosis of the COVID could help in the resources optimization such as all supportive measures and expertise human resources that are needed for taking care of positive patients.

Fully automatic artificial intelligence-based classification method using CXR has a wide potential for this unmet need which can be evidenced from the recent research. The most common diagnostic imaging used is CXR in comparison with magnetic resonance imaging (MRI) and computed tomography (CT) due to lower radiation exposure, little processing time, and lesser amount of cost [46]. In current epidemics, it is necessary to keep suspected patients in isolation for their appropriate and timely treatment. Moreover, rapid screening is necessary to diagnose such types of virus-infected patients in order to control the outbreaks. In the future, AI-based classifiers for the classification of diseases may also be integrated with laboratories for testing. Moreover, AI-based tools can act as an aid for rapid and less time-consuming prognosis and follow-up of the patients. Nowadays, several researchers have made various attempts in developing classification or identification methods for COVID detection using CXR images with different capabilities [47]. However, these studies have certain substantial restrictions that need to be resolved for the advancements of more accurate and reliable classification models. There are only few studies which have employed CXR images for different age groups [17, 20] and [22]. Moreover, models may get biased because of the differences in size of lungs of adult vs. pediatric age groups. The majority of studies did not consider the age group and postero-anterior view or latero-lateral view [17–20, 22, 26] and [27]. The absence of these considerations may provide unsuitable training to models which may not perform well.

In literature, many authors have used the augmentation techniques with less augmentation types for increasing the image dataset size and for the generalization of models [20, 22]. It has also been revealed that in some studies while providing the training to the models, non-COVID CXR images are wrongly mapped into the COVID images class [21], which can also influence the deep learning-based model accuracy. Moreover, it may lead to high false-positive rates in situations of real-life predictions. Moreover, most of the latest studies have claimed high accuracies of models on small CXR image datasets and may have overfitting problems. In spite of this, the proposed model has no overfitting issue.

3.2. Comparative Analysis of Proposed Model with Another State-of-the-Art Models

From the literature, it has been noted that several attempts have been made by various researchers in the area of COVID-19 detection using CXR images. Table 3 has listed the comparison of the proposed model with the studies of Nishio et al. [48], Sharma et al. [49], Wang et al. [50], Mahmud et al. [51], and Fawaz et al. [36]. Nishio et al. [48] have introduced the integration of conventional method with augmentation method with the use of transfer learning and have achieved 83.6% accuracy in detecting COVID patients on CXR images. Sharma et al. [49] have also utilized transfer learning on CXR image dataset and achieved 100% accuracy on 24 epochs. Further, Wang et al. [50] introduced the new model by the integration of ResNet101 and ResNet152 for the detection/classification of COVID and pneumonia from healthy ones and achieved 96.1% accuracy.

Table 3

Comparison of the proposed model with other state-of-the-art models.

Author	Dataset	Approach/method	Accuracy
Nishio et al. [48]	1248 CXR images—998 (training), 125 (validation), and 125 (test) images	Pretrained model is VGG16 integration of conventional method with augmentation methods	83.6% (non-COVID-19 pneumonia, COVID-19 pneumonia, and healthy)
Sharma et al. [49]	352 X-ray images (317—training and 35—testing)	Transfer learning used and external dataset validation performed	100% and 75% accuracy achieved for training and testing (24 epochs-based model)
Wang et al. [50]	Training (normal—7966, pneumonia—8620, and COVID-19—128)	Fusion of pretrained model ResNet101 and ResNet152	96.1%
	Testing (normal—885, pneumonia—956, and COVID-19—12)
Mahmud et al. [51]	Normal X-rays (1583), viral pneumonia (1493), and bacterial pneumonia (2780)	Utilizes depth-wise convolution with different dilation levels for the extraction of features from CXR	96.9% (COVID/viral pneumonia), 97.4% (COVID/normal), 94.7% (COVID/bacterial pneumonia)
Fawaz et al. [36]	The dataset consists of 21,165 X-ray images COVID-19-positive (3616), viral pneumonia (1345), images of normal patients (10,192), and lung opacity (non-COVID-19) (6012)	ResNet50 and AlexNet are used and features mined by convolutional neural network (CNN) using traditional methods	99% accuracy—COVID-19/viral pneumonia/lung opacity/normal
Proposed	Normal—1583	Weight fusion of InceptionV3 and VGG16 on FC layers	99.5% accuracy for (healthy/unhealthy) 20 epochs and 98.2% (normal/COVID/pneumonia) 100 epochs
	Pneumonia—4273
	Bacterial pneumonia—1860
	Viral pneumonia—2413
	COVID-19—714

Mahmud et al. [51, 52] utilized the depth-wise convolution with variable dilation ratios on CXR images and have achieved 97.4% accuracy on COVID/normal. In the proposed model, weights of InceptionV3 and VGG16 are fused and transfer learning has been applied on CXR images (1583 images are for normal conditions, 4273 images are of pneumonia disease, 1860 images are of bacterial pneumonia, and 714 images are of COVID-19) [53-56]. From results, it has been observed that 99.5% accuracy has been achieved using binary classifier when analyzing two classes with 20 epochs and 98.2% accuracy using multi-classifier for three classes with 100 epochs. Test accuracy achieved is 62.5% in four-class classification (healthy, COVID-19, bacterial pneumonia, and viral pneumonia). It can be understood that proposed transfer learning model used for 4-class classification is not appropriate because of overfitting issue as if we eliminate the first layers, and then it will disturb the dense layers as the number of trainable constraints will be altered. It can be said that model does not confuse COVID-19 and healthy images, but it confuses COVID-19 with viral pneumonia and bacterial pneumonia.

In the future, the training provided to the model will be improved additionally with the incorporation of a large number of CXR image dataset for the development of more scalable and robust models for classification of diseases. The proposed artificial intelligence-based classification models in this work are non-GPU-based and henceforth can be somewhat slow in the classification of diseases using CXR. Therefore, there is a need to develop GPU-based models for the classification of diseases as well as can provide bulk uploading of images to the users via an interactive web interface. The researchers are constantly making attempts to improve the deep learning-based classification models by making use of maximum available data as CXR image dataset on COVID-19 is increasing day by day. This will enhance the utility and reliability of the models in the real-world scenarios. Besides, the accessibility of different and large CXR image datasets will aid in developing more scalable and robust deep learning-based classification models.

4. Conclusion and Future Work

In this work, the weights of existing pretrained model VGG16 and InceptionV3 have been taken and a base model has been proposed using pretrained models (VGG16 and InceptionV3) by integrating the last FC (fully connected) layer as per the number of classes for classification of chest X-rays in binary and multi-class classification by appropriately using transfer learning and finally making the combination layers by integrating the FC layer's weights of both the models (VGG16 and InceptionV3). The proposed model has not been undertaken for a clinical study. Hence, it does not substitute a clinical diagnosis as more detailed exploration could be performed with a larger image dataset. Enhancement of accuracy over proposed work can be done on multi-class classification to distinguish between viral pneumonia, bacterial pneumonia, and COVID-19-affected lung CXR images. Also with increase and variety in dataset, the classification should be able to work on real-time basis. Furthermore, the proposed model has contributed to an automatic, accurate, inexpensive, and fast method in order to assist the diagnosis of COVID through CXR images. Moreover, the proposed model is also intended to integrate with a free online platform of lung disease classification. This would aid physicians and doctors to identify diseases in CXR images around the world, without developing their own classification platforms. Additionally, the proposed model has been compared with other models based on accuracy and training a network from scratch. The proposed models have performed better as binary and multi-class classifiers have achieved 100% accuracy. This model can be used in those places where there is a shortage of radiologists to take care of COVID-19-infected persons. The limitation of this study is the utilization of a limited number of COVID-19 CXR images, and it can be made more accurate and robust by using a large number of images from local hospitals.