1. Introduction

In December 2019, Wuhan City in the Chinese province of Hubei reported an unfamiliar virus, which was later identified as a new type of coronavirus. The initial coronavirus was Severe Acute Respiratory Syndrome (SARS), which emerged in 2002; in 2012, Middle East Respiratory Syndrome (MERS) emerged, resulting in respiratory illness outbreaks in various Middle Eastern countries, Europe, and Asia. Symptoms of MERS include fever and difficulty breathing. After the World Health Organization declared COVID-19 a pandemic, the number of cases rose exponentially. By 12th April 2020, the number of cases reached 1.8 million while deaths across the globe reached 114,698 . The virus affected the entire globe, severely affecting the USA, Spain, and Italy, with the number of active cases at 560,433, 166,831, and 156,363, respectively, while deaths reached 22,115, 17,209, and 19,899, respectively ^[1]. According to research, the coronavirus family has dozens of different viruses, but only seven of them are a threat to humans. It should be emphasized that these viruses are transmitted to humans by animals. A recent outbreak of Ebola in the Western African country of Guinea was traced back to bats. It is therefore important that people take precautions when traveling to countries where these viruses exist. Precautions include avoiding contact with animals (especially bats) and taking steps to protect oneself from infections (e.g., washing hands regularly, and wearing gloves and masks when necessary). The COVID-19 virus that caused an acute respiratory ailment in humans emerged in a new and potentially deadly form. Early indications were that it spread more easily than past strains and posed a significant risk to public health. However, the complete extent of the virus’s potential hazards remains unclear, and further research is necessary.

The deep learning algorithm called Xception net has the ability to scan X-ray pictures of the lungs and determine the extent of COVID-19 infection. Xception net can also be used to diagnose diseases such as lung cancer. Analysis is done by a computer system that is able to learn from large amounts of data. The deep neural network (DNN) has proven particularly effective in medical image analysis applications among different deep-learning classifiers. The DNN’s results have demonstrated its efficacy in mapping picture data into predictable output. The DNN Xception network architecture is able to learn representations of data that are more discriminative for COVID-19 detection. The use of a large number of layers and filters allows the network to capture more intricate details from the data. This leads to improved performance in detecting COVID-19. The main outcome of this study is creation of a DNN driven model capable of training with images of both healthy and coronavirus-infected lungs. The proposed approach may detect COVID-19 infections more quickly by recognizing characteristics of infected individuals in X-ray pictures of the lungs, such as grey or shadowed areas.

In this study, we are going present a variation of the Xception net deep learning neural network technique for identifying COVID-19 by evaluating patient X-rays and CT scans, looking for visual markers in COVID-19 patients’ chest radiographic imaging.

2. Literature Review

After being infected with COVID-19, a patient may start showing multiple symptoms, including cough, fever, and respiratory distress (similar to the flu). However, under severe conditions, the infection may result in pneumonia, trouble breathing, multiple organ failure, and even death ^[2,3].

Most of the research is oriented on deep learning techniques, so chest radiography of COVID-19 virus victims can be used for the detection of particular features. Diseases can be automatically detected and managed owing to accurate analysis, identification, and classification of patterns in medical images from deep learning applications. Screening infected patients is critical in order to contain COVID-19, because only then can positive cases be identified and treated. RT-PCR is currently the main screening methodology for detecting COVID-19 ^[4,5]. Patients’ respiratory samples are used to conduct the test. Patients can obtain results within a few hours, or up to two days later. Chest radiographic images can be used as an alternative to PCR screening. This has been documented by multiple research articles in the journal Radiology ^[6,7]. According to researchers, positive COVID-19 patients have a few perceptible marks on their lungs. These marks are like ground glass ocular views, such as hazy darkened spots that aid in differentiating between COVID-infected and non-COVID patients ^[8]. Recently, Wang and Wong ^[9] implemented a model for detection of the COVID-19 virus, which was 83.5% accurate in classification of normal, pneumonia-viral, and pneumonia-bacterial cases. Narin et al. ^[10] achieved a detection accuracy of 98% for two classes by training a ResNet50 model with the of X-ray images of the chest. But its performance is not known for multi-class classification.

In order to fight COVID-19, researchers from around the globe have dedicated many days and nights. Several researchers have published papers outlining methods to detect COVID-19 using chest radiography ^[11].

Diseases can be automatically detected and managed owing to accurate analysis, identification, and classification of patterns in medical images thanks to deep learning applications. The primary reason is that deep learning algorithms can automatically learn features from the data, rather than relying on manually detected features ^[12]. In order to detect pneumonia from an X-ray of the chest, the deep neural network model that is generally employed is ChexNet ^[13]. This model has surpassed the performance of the average radiologist with its exceptional results. Another deep neural network model is Covidx-net ^[14], which is used for diagnosing thorax diseases with the aid of radiographic images of the chest, comprised of 17 convolution layers and the Leaky ReLU activation function. Their model provided 98.08% accuracy, and other parameters were 87.02% accurate for multi-class cases. These techniques are crafted to perform three-class classification (normal vs. pneumonia vs. COVID-19) or binary classification (normal vs. COVID-19) except for COVID-Net.

Apostolopoulos and Mpesiana ^[15] used different deep learning models pre-trained on a dataset comprised of 224 images from COVID-19 cases. Their model was 98.75% accurate for two of the classes, and 93.48% accurate for multi-class cases based on a three-class classification. Sethy and Behera ^[16] employed multiple CNN models along with an SVM-based classifier to identify COVID-19. Hemdan et al. ^[17] employed chest X-ray images and multiple deep-learning models for diagnosing COVID-19. Their research proposed the COVIDX-Net deep learning model, which consists of seven CNN models. Not one of the aforementioned models addresses pneumonia-viral and pneumonia-bacterial cases as individual classes, except for COVID-Net.

3. Dataset Description

Data act as fuel in deep learning for the training of models. The COVID-19 virus initially appeared in December 2019, and it has affected people around the globe, so no effectively sized dataset is available to conduct research. To facilitate our research, we utilized a dataset ^[18] consisting of X-ray images from various publicly accessible databases that are regularly updated by researchers from different regions. The dataset includes 43 positive female cases and 82 positive male cases for a total of 125 X-ray scans. Complete metadata for all patients is not mentioned. The average age of these COVID-19 patients is 55 years. We used 500 uninfected cases and 500 pneumonia cases selected at random from the dataset to avoid an unbalanced dataset. Examples are shown in Fig. 1.

Fig. 1. COVID-19 cases: (a) normal; (b) possible pneumonia; (c) infiltrated; (d) ground glass opacity; (e) infection; (f) patchy opacity.

4. The Proposed Approach

The aim of this study is to introduce a deep-learning approach to detecting COVID-19 infections using chest X-rays. To classify three types of pneumonia, we developed a deep neural network. Our research pertains to the categorization of viral, bacterial, and COVID-19 pneumonia, comprising the implementation of our proposed model for classifying binary and three-class versions and comparing results from other techniques. We named our proposed model CoroNet, and it will help differentiate the types of pneumonia. Doctors have found the model useful in diagnosing, quantifying, and following up on positive COVID-19 cases that are based on radiography images of the chest. The proposed model may not become an alternative to the current method completely, but a number of cases can still be identified where immediate testing or further review by experts is required. Our research can be fully grasped from the pseudo-code in Algorithm 1.

This section discusses the DNN layers, the proposed architecture, and its evaluation metrics.

A: Deep Neural Network

Deep learning image processing is mostly done using a DNN. Although the DNN is most widely used for image-based analysis, it can also be used for problems like data analysis or classification. Deep architectures allow these networks to learn several diverse and dynamic features that a basic neural network cannot learn. DNNs have many applications, including image recognition, video analysis, NLP, and time series forecasting.

A DNN consists of the following basic layers that help it to perform the abovementioned tasks.

1) Convolutional Layer: Instead of using matrix multiplication, this layer uses a convolutional operation. It has a number of learnable parameters called kernels. The key role of this layer is to detect features in the dataset and map them to a feature map in this layer. The operation is illustrated in Eq. (1).

where, I : input image

K : Filter of size m${\times}$n

F : Output of filter K

I*K is the convolutional operation. In order to introduce non-linearity, output from a DNN layer is pre-input to an existing activation function. Activation functions come in many types, but for classification the ReLU activation function is used.

2) Pooling Layer: In order to minimize the epochs of the parameters in the NN and minimize the spatial size of the input, a downsampling layer is used in the DNN. The most commonly used downsampling technique is Max Pooling, which takes a region as input and returns only the maximum value from it. Eq. (2) shows this operation.

3) Fully Connected Layer: DNNs are neural networks that excel in pattern recognition. They consist of multiple layers, with the initial layers being responsible for learning basic features, while the later layers classify or predict objects. The intermediate layer is known as the fully connected layer, which connects all neurons in the previous layer to all neurons in the successive layer, aiding in learning more complex features. The final layer of most DNNs uses a sigmoid activation function, which may encounter issues when distinguishing between classes with similar scores. For instance, if two classes have scores of 0.99 and 0.98, a sigmoid activation function would consider them equally likely. To address this issue, many DNNs employ a softmax activation function in the final layer. Softmax is a mathematical function used in conjunction with a neural network, and helps improve the accuracy of predictions made by the network. It calculates the probability distribution of the resultant class, as expressed in Eq. (3):

where x is input, and Z is output. Since the softmax function deals with probabilities, the sum of all probabilities is 1. All the abovementioned layers are lined up together to create a complete DNN architecture. In addition to abovementioned layers, the DNN can provide optional layers, such as batch normalization, to boost training time; to resolve the overfitting problem, a dropout layer is used.

B: The Model Architecture

Automated COVID Detection (ACDD) is a DNN-based architecture developed to automatically detect COVID-19 using X-ray scans. It uses the Xception model (basically a DNN model) that is trained on the ImageNet corpus. It is a modified version of the Inception architecture with depth-wise separable convolutions instead of standard inception modules. This reduces the number of operations by 1/k.

The vanishing gradient problem occurs when an activation function is used in almost every layer, which makes gradients of the loss function approximately zero, and makes the network difficult to train. In order to avoid this problem, residual connections are used to allow the flow of gradients without passing through the nonlinear activation function during backpropagation, as shown in Fig. 2.

In ACDD, the Xception net model is implemented as the base model featuring a dropout layer and two fully connected layers at the end. Due to insufficient data, we adopted a pre-trained model, utilizing transfer learning to prevent overfitting. The details of the model architecture are listed in Table 1.

Fig. 2. Residual Connections.

5. Experimental Results

A: Dataset-1 Classification Accuracy

Our proposed model was trained using a dataset of CT scans consisting of 349 COVID-19 images and 463 non-COVID-19 images in a binary classification problem addressed using ACDD. We allocated 80% of the data for model training and 20% for testing, utilizing a pre-trained model with the Adam optimizer, a learning rate of 0.0001, a batch size of 10, and 80 epochs. To enhance model performance, we incorporated data shuffling during preparation, reshuffling the dataset for each epoch. Figs. 3 and 4 show the training and validation set accuracy and loss plots, respectively.

Fig. 3. Training and validation binary class accuracy plots.

Fig. 4. Training and validation set binary class loss plots.

B: Dataset-2 Classification Accuracy

The outcomes demonstrated that the DNN model achieved superior performance on both the training and validation datasets. We allocated 80% of the dataset for model training and 20% for testing, utilizing a pre-trained model with the Adam optimizer, a learning rate of 0.0001, a batch size of 10, and 80 epochs.

Furthermore, we incorporated data shuffling during preparation, randomizing the data before each epoch to enhance the model’s performance. Figs. 5 and 6 show the training and validation set accuracy and loss plots, respectively, for binary classification.

In a similar fashion, Figs. 7 and 8 show the training and validation set accuracy and loss plots for multi-class classification. We assessed the model’s performance on a test set consisting of 29 COVID-19, 72 normal, and 120 pneumonia cases. The model demonstrated a satisfactory level of accuracy on the test set. It displayed remarkable precision in distinguishing COVID-19 cases from normal and pneumonia cases. Additionally, the AI model’s performance remained consistent across various data splits. Performance metrics for binary and multi-class classification are given in Table 2.

Fig. 5. Training and validation set binary class accuracy plots.

Fig. 6. Training and validation set binary class loss plots.

Fig. 7. Training and validation set multi-class accuracy plots.

Fig. 8. Training and validation set multi-class loss plots.

C: Dataset-3 Classification Accuracy

To validate our proposed model, we employed an additional dataset comprising 310 normal X-rays, 330 bacterial pneumonia X-rays, and 284 COVID-19 X-rays, for a total of 924 images. Prior to the classification task, we resized all the images to 224x224. After fine-tuning our proposed model, we obtained accuracy of 90%. Figs. 9 and 10 show the accuracy and loss plots, respectively, for binary classification. Similarly, Figs. 11 and 12 show training set and validation set accuracy and loss plots, respectively, for multi-class classification. In addition, precision, recall, and accuracy results for binary classification and multi-class classification are given in Table 3.

Fig. 9. Training and validation set binary class loss plots.

Fig. 10. Training and validation set binary class loss plots.

Fig. 11. Training and validation set multi-class accuracy plots.

Fig. 12. Training and validation set multi-class loss plot.

In order to detect COVID-19 virus cases from chest X-ray images and CT scans we deployed a deep learning architecture based on the Xception model. Two datasets were used for its validation, and the proposed model gave the best results. A comparison is shown in Table 4.

6. Conclusion

Many countries across the globe are witnessing the depletion of resources while COVID-19 cases surge daily. It inherently becomes important to identify positive cases, with no case going unidentified. With this perspective in mind and recognizing the urgency of the situation, we developed a DL-based technique for detecting the COVID-19 virus through chest radiography. Our proposed methodology employs a specialized DNN model called CoroNet, designed to identify COVID-19 cases. We constructed a dataset utilizing publicly available chest X-ray images of COVID-19 and pneumonia cases from various other databases. In our experiments, CoroNet demonstrated promising outcomes on the prepared dataset and proved to be computationally less expensive. If training data are enhanced, its performance can be enhanced further. Despite the promising results, CoroNet still needs to undergo clinical study as well as testing. Nonetheless, it can already be beneficial; it is highly sensitive and accurate in detecting COVID-19. Hence, health experts and radiologists can employ CoroNet for a better understanding of crucial aspects that are linked to COVID-19 cases.