MCNN: A Deep Learning Based Rapid Diagnosis Method for COVID-19 from the X-ray Images

As of yet reported, at the end of December 2019, an unusual and unidentified case of pneumonia originated from Wuhan city, Hubli Province of China. Initially, the symptoms of this disease were very similar to the known pneumonia. But, “Chinese Centre for Disease Control and Prevention (China CDC)” experts reported based on the analysis of various respiratory samples (throat swab samples) that this is a new kind of virus-infected disease. It spreads through a novel coronavirus and looks very similar to the native family of SARS-CoV and MERS-CoV [1]. Later on, the World Health Organization (WHO) officially announced the name of this disease as COVID-19 (coronavirus disease). As per the reports of WHO, till 23rd July 2020, the pandemic had covered 205 countries, 10,719,946 confirmed cases, and 517,337 confirmed deaths [2] were reported. COVID-19 has become a global threat to humanity. According to the research so far the coronavirus spreads among humans through facilitation via direct contact or inhalation of infected droplets and touching contaminated hand/places/objects/surface [1]. However, in some cases, the presence of gastrointestinal symptoms or live infectious virus in faecal might be another possible route of spread of infections in humans through faecal-oral transmission [3]. Droplets have two categories based on their sizes, known as respiratory droplets and nuclei droplets If the size of the droplet is >5-10 μm, then it is named to as respiratory droplet, and if the size is <5 μm, then it is named as nuclei [4] droplet. As per the current research, respiratory droplets are primarily responsible for the transmission of coronavirus in humans [5]. It has been proved by Phylogenetic analysis that the SARS-CoV-2, SARS-CoV, and MERS-CoV belong to Orthocoronavirinae, which further belongs to the Coronaviridae family [6]. But, both the viruses SARS-CoV and MERS-CoV are distinct from SARS-CoV-2 [7].

Generally, the symptoms of COVID-19 are dry cough, respiratory signs (like flu), acute fever, tiredness, and headache. The appearance of the disease may range from negative symptoms to positive symptoms. Here, positive symptom means the presence of the disease outbreak. In a recent study done over 41 patients who initially diagnosed with COVID-19, the most common symptoms available in the majority of the patients were cough, fever, and fatigue, while in some cases a severe and even fatal symptoms were noticed like difficulty in breathing, headache, hemoptysis, dyspnoea, diarrhea, and lymphopenia [8]. All patients were found to be affected by pneumonia after a chest CT scan. About 29% of the cases were of acute respiratory distress syndrome, 12% of cases of an acute heart injury, and 10% of cases of other infections were identified [8].

In recent studies, it is found that the automatic diagnosis of COVID-19 may help to implement different safety measures such as putting the patients in the quarantine centre on time, using specialized hospitals for incubation on time for the critical patients, and real-time monitoring of COVID-19 cases. In the current time, a real-time reverse transcription-polymerase chain reaction (RT-PCR) is being commonly used that takes respiratory specimens of the patients to diagnose the COVID-19 cases [9]. Thus, RT-PCR requires repetition in the test to make a confirm report [10]. Also, the sensitivity rate of RT-PCR is commonly between 60% -70% that can affect the diagnosis [11]. So, a valid environment needs to be developed, where proper medical care and quick treatment be available to take care of the patients against the COVID-19. Also, considering the fast-spreading nature of the COVID-19, it requires a diagnostic method that gives a quick response to the presence of disease in a human body. In this context, radiological images such as X-ray and Computed Tomography (CT) could be a better and faster way to diagnose COVID-19 cases [12]. It motivates us to work on radiological images through the identification of potential patterns that may help to diagnose the disease automatically. However, it requires some expert methods for automatic feature extraction and image classification. Diagnostic techniques based on Artificial Intelligence (AI) are being observed as a better alternative to provide expected results over the radiological images in terms of detecting the COVID-19 cases. Especially, deep learning, a sub-field of machine learning has gained acceptance in the field of artificial intelligence for mining, analyzing, and recognizing distinct data patterns. Deep learning creates an end to end model, which takes data as input and gives desired results, with automatic feature extraction and no manual intervention. The increased number of hidden layers of deep learning models help to potentially enhance the performance of the network as compared to the traditional models. Major application areas of deep learning are speech recognition [12], computer vision [12], and natural language processing [11]. In literature, we can see the significance of deep learning techniques in many applications such as breast cancer detection [13], segmentation of lung [14], X-ray images for pneumonia detection [15], and classification of skin cancer [16]. Recently, several deep learning models have been used on radiological images to detect DOVID-19. A deep learning model named COVIDX-net was proposed by Hemdan et al. [17] to diagnose the COVID-19 on X-ray images. Narin et al. [18] have proposed a deep learning model for the detection of COVID-19 cases on chest X-ray images. They found a 98% detection accuracy in the results.

Convolutional Neural Network (ConvNet or CNN), which is known as a class of deep neural networks in deep learning, uses convolution to perform feature extraction automatically. Learning of complex information is based on going deeper into the network and each layer is responsible for transforming and getting the abstract level of the data. CNN is widely applicable in computer vision. CNN has a significant role in image segmentation and classification in image processing. In recent years, CNN has been applied on chest X-ray images to diagnose pneumonia with high accuracy. COVID-Net was developed by Wang and Wong [19] for the classification of COVID-19, pneumonia, and normal images with an accuracy rate of 92.4%. Rajpurkar et al. [20] have proposed a CNN model named CheXNet. This model includes 121 layers of CNN for training over X-ray images to detect pneumonia. Another CNN method DenseNet-121 is proposed by Antin et al. [21] to detect pneumonia from the chest X-ray images. Recently, various models based on deep learning have been developed to analyze the COVID-19 from chest X-ray images, but, most of the methods are found to be less efficient in classifying the difference between COVID-19 and non-COVID-19 patients accurately and also take much time to analyze the cases in early stages. Also, in a study, it was observed that the recovered cases of COVID-19 were found positive again, during the re-diagnosis done after 5-13 days [22].

Some disadvantages of the conventional RT-PCR method are like the first, the unavailability of a sufficient number of RT-PCR test kits, and second, it includes the manual process for collecting respiratory specimens creating a chance of malfunctioning in diagnosis done by the experts. The third one is that it is a time-consuming process for getting test results, and fourth the detection rate of the test is very low, it commonly lies between 30% -50% so RT-PCR requires repetition of tests to confirm a positive or negative test result.

Artificial Intelligence (AI) based diagnostic techniques are a better alternative and provide expected results in terms of detecting the COVID-19 cases using radiological images. The deep learning model named COVIDX-net diagnoses the COVID-19 on X-ray images and another detects COVID-19 cases on chest X-ray images. They show 98% accuracy in the results while detecting COVID-19 cases. However, in most of the cases, these methods performed less efficiently in classifying the positive and negative tested patients accurately in the early stages.

Hence, it is required to propose such a method that can give an earlier diagnosis of COVID-19 with zero tolerance and higher accuracy. In the paper, a modified CNN named Modified Convolutional Neural Network (MCNN) is proposed for the earlier detection of COVID-19 from the X-ray images. The proposed method shows better results as compared to other existing state-of-the-art deep learning models that helped in earlier diagnosis of COVID-19 through X-ray image classification & observation, together under the clinical support.

3.1 Details of X-ray image dataset

In the study, chest X-ray images have been collected from various sources [29-35]. It includes images of basically two categories that are COVID-19 and Non-COVID-19. It consists of images of varying dimensions and various age groups. The dataset is categorized into two sets, which includes a training set and a testing set. There are 2000 images in the training set and 400 images in the testing set. All the images are in the JPEG format and occupy 622 MB of the system space. The dataset is a structured dataset consisting of an equal number of images of each category for training and testing intents. The train and test datasets are in the proportion of 10:2 (1000:200).

The distribution of image datasets for training and testing is shown in Table 1. For a particular class (COVID-19 or non-COVID-19), 1000 images have been used for training, and 200 images for testing. Since there are 2 classes i.e., class 1 (COVID-19) and class 2 (non-COVID-19), it can be seen that out of 2400 images, 2000 images are available in the training set, and 400 images are part of the testing set.

Table 1. Distribution of the dataset

In Table 2, the varying dimensions of the input images are shown in pixels where the minimum width of the class of COVID-19 images is 255, and the maximum width is 4248. The minimum and maximum height of these images are 249 and 3480, respectively. For the class of Non-COVID-19 images, the minimum and maximum width of the image is 976 &, 2890 respectively, and the minimum and the maximum height of the image is 616 & 2534.

Table 2. Distribution of the dataset

Figure 3 shows a sample of chest X-ray images of COVID-19 and non-COVID-19. Such type of images has been taken from the dataset used for the proposed model. Figure 3.a shows the chest X-ray images of Non-COVID-19 people, and Figure 3.b shows the chest X-ray images of COVID-19 infected people.

3.png

Figure 3. Sample Images of Non COVID-19 and COVID-19 patients [23-29]

3.2 Proposed method

The parameters are tuned in such a way that the proposed method, modified Convolutional neural network (MCNN) attains high accuracy with less time. The basic CNNs model is tuned by removing un-wanted layers that take more time. Some additional layers are added that can extract more vital and accurate features efficiently without putting the model into the overfitting condition. For that Rectified Linear Unit (ReLU) and Sigmoid activation functions are used. To optimize the proposed method, pool and dropout have been tuned enough that helps to achieve a 98.25% accuracy by the proposed algorithm. The number of layers and dropout layers have been evaluated during multiple times of training to record the loss and accuracy of the model. This record is used as feedback to train the model. Activation functions, drop out layers, convolutional layers, max-pooling layers, and dense layers are tuned enough such that the loss in training would be minimum. Also, the unnecessary layers and functions are removed so that the overall time is reduced, and the accuracy is not affected. Hence, by using the proposed MCNN an accurate diagnosis result of whether a person is COVID-19 positive or not can be achieved. The MCNN will save money and time, so the efficiency of the whole diagnosing procedure will increase by adopting this method.

Figure 4 shows the sequence in which the layers are used in the model. Conv is used to represent the Convolutional layers which are used in the training. Here x represents the number of the layer, e.g., Conv1, Pool2, etc. The Pool represents the pooling operation. Here, the Fully Connected (FC) layer represents that flatten, dense, and dropout operations have been performed and x represents the number of layers, e.g., Conv1, Pool2, etc.

4.png

Figure 4. Layer wise layout of the proposed model

3.2.1 Proposed algorithm

Algorithm 1: Training the model

Input: Image dataset

Output: Weights of the proposed model

Method: Model Training

Begin

Step 1. Load the training and testing images and convert them into the training set and validation set.

Step 2. Construct a model with 6 convolutional layers where the kernel size is increased in each layer. After each convolutional layer apply pooling layer to maintain dimensionality. After the 6^th convolutional layer flattens the output and pass it to fully connected layer.

Step 3. Apply sigmoid activation function at last to distinguish the images.

Step 4. Save the weight file to use it separately to distinguish the input image class.

End

Algorithm 2: Classification

Input: Input Image

Output: Label of input image

Method: Classification of input image

Begin

Step 1. Feed the input image to the neural network.

Step 2. Obtain output vector by matching the weight file and input image.

Step 3. Use the output vector to perform image annotations.

End

3.2.2 Architecture of the MCNN

The input image is given to the Convolutional layer. The image is split into different layers. Each layer contains different features of the image. Then the ReLU activation function is performed where all the negative values are replaced by zero. Pooling is done to reduce the size of the image matrix. Then again the image matrix is passed through the Convolutional layer. Then ReLU function is performed on the matrix generated. Pooling has been done after performing the ReLU activation function. The image matrix obtained now is passed through the Convolutional layer again for feature extraction. Now after performing ReLU function the dropout layer is performed. In the dropout layer, some neurons of the previous layer have not been connected. The dropout layer decreases the overfitting. Now a combination of the Convolutional layer, ReLU function, Dropout layer, and pooling layer are performed 3 more times. Then the flatten layer is used for a fully connected layer. The flatten layer has converted the image matrix into a vector form. Now ReLU function is used which removes any negative value. After that, the dense layer is applied four times which is followed by the ReLU function. The dense layer checks whether each neuron is connected with the previous layer or not. After applying a dense layer dropout layer is performed. This dropout layer has decreased the overfitting curve by reducing the training loss. At last sigmoid function is used for classification of the image into the categories. This procedure is shown in Figure 5.

5.png

Figure 5. Layer operations in the proposed method

The sequence in which layers have been used in the purposed model is given below. In the sequence, C is used for the Convolutional layer, R for the ReLU activation function and P denotes the pooling layer. DR indicates the dropout layer, F represents the flattening layer, D symbolizes dense and S is used for the sigmoid activation function. It is shown that there are 6 Convolutional layers in the model.

C1 – R – P1 – C2 – R – P2 – C3 – R – P3 – DR – C4 – R – P4 – DR – C5 – R – P5 – DR – C6 – R – P6 – DR – F – D – R – D – R – D – R – D – R – DR – D – S

The architecture of the MCNN is shown in Figure 6, where first the input image is passed through the Convolutional layer denoted as Conv2D where feature extraction is being done at various levels and layers of the image are split. Then the extracted features are given to the pooling layer denoted by MaxPool because at this layer the features extracted from the Convolutional layer are stored in a smaller matrix which is obtained by marking the matrix cell with the highest value of the kernel matrix when imposed over the Convolutional layer matrix. This process is performed several times with only a change in the number of features to be extracted from the previous layer. After that, when the feature extraction has been done and an increased loss was observed even by applying MaxPool then dropout is used. Now the two-dimensional MaxPool layer is flattened into a single column matrix form and a dense layer is applied to the flatten layer. Sigmoid activation function has been used in the last layer for classification, the highest percentage, in which the image is matching, is calculated, and the image is predicted to be belonging to that class.

6.jpg

Figure 6. Architecture of MCNN

MCNN is a fine-tuned neural network. As CNN has varying parameters for each layer. While in MCNN the parameters to be passed are fixed. As MCNN contains six convolutional layers. In CNN there is no limit on the number of convolutional layers. In MCNN every convolutional layer has been followed by a max-pooling layer to maintain the dimensionality of the image. In CNN pooling layer can be either a max-pooling layer or an average pooling layer. MCNN consists of seven dense layers with fix sequence of parameters to be passed. In CNN there is no limit on the number of dense layers. The dense layer can have different values of parameters. In MCNN dropout layers have also been applied to prevent the graph from going into an overfitting region. The dropout layer is not mandatory on CNN. In CNN there are various activation functions such as leaky ReLU, ELU, softmax, softsign, etc. In MCNN only ReLU and sigmoid activation function had been used. Every layer has been parameterized properly. The sequence in which layers should be applied is fixed. Parameters are so tuned that the graph goes neither in the under- fitting region nor in the overfitting region. MCNN graph curve falls in the good fit curve category. In this way, MCNN has a minimum number of layers and gives good accuracy.