key: cord-0826515-gpm24h8f
authors: Perumal, Murukessan; Nayak, Akshay; Sree, R. Praneetha; Srinivas, M.
title: INASNET: Automatic identification of coronavirus disease (COVID-19) based on chest X-ray using deep neural network
date: 2022-03-03
journal: ISA Trans
DOI: 10.1016/j.isatra.2022.02.033
sha: 3cdae03b2c427a1769baf0487bc77a7e29c7bf3e
doc_id: 826515
cord_uid: gpm24h8f

Testing is one of the important methodologies used by various countries in order to fight against COVID-19 infection. The infection is considered as one of the deadliest ones although the mortality rate is not very high. COVID-19 infection is being caused by SARS-CoV2 which is termed as severe acute respiratory syndrome coronavirus 2 virus. To prevent the community, transfer among the masses, testing plays an important role. Efficient and quicker testing techniques helps in identification of infected person which makes it easier for to isolate the patient. Deep learning methods have proved their presence and effectiveness in medical image analysis and in the identification of some of the diseases like pneumonia. Authors have been proposed a deep learning mechanism and system to identify the COVID-19 infected patient on analyzing the X-ray images. Symptoms in the COVID-19 infection is well similar to the symptoms occurring in the influenza and pneumonia. The proposed model Inception Nasnet (INASNET) is being able to separate out and classify the X-ray images in the corresponding normal, COVID-19 infected or pneumonia infected classes. This testing method will be a boom for the doctors and for the state as it is a way cheaper method as compared to the other testing kits used by the healthcare workers for the diagnosis of the disease. Continuous analysis by convolutional neural network and regular evaluation will result in better accuracy and helps in eliminating the false-negative results. INASNET is based on the combined platform of InceptionNet and Neural network architecture search which will result in having higher and faster predictions. Regular testing, faster results, economically viable testing using X-ray images will help the front line workers to make a win over COVID-19.

Acute Respiratory Syndrome coronavirus 2 (SARS-CoV2) [1] . It has been firstly identified in December 2019 in Wuhan, Hubei province, China. And the first case has been reported in November, 2019 [2] . The virus initially named as novel coronavirus but later it has been changed to SARS-CoV2 after studying its internal structure by world health organization. Virus has created so much panic among the masses [3] . We are not seeing it for the first time as SARS and MERS are the virus from coronavirus family which affects the humans severely.

We still had not been developed the efficient vaccine or particular effective cure. SARS and MERS were the previous outbreak caused by the members of the coronavirus family in 2002 and 2012 respectively and still showing their terror at regular time intervals [4] [5] . Symptoms in the infected persons are similar to the patient getting infected with common cold, influenza, and pneumonia which makes the identification of the infection based on the symptoms. And fever, cold, fatigue, loss of appetite, loss of smell and taste, lung fibrosis are the major symptom seeing in a COVID-19 infected patient [6] [7] . In severe cases, the person is feeling shortness of breath, uneasiness in breathing, acute respiratory syndrome, and even may have severe pneumonia. The incubation period is ranging from two to fourteen days. WHO has declared the diseases as pandemic Testing is the main weapon in order to prevent the spread of COVID-19

infection. All countries are working on finding a suitable testing method which will be both efficient and cheaper [9] . There are various testing methods that are present which are not as efficient and are taking very much time in giving the results [10] . RTPCR testing is the most commonly used testing where the sample has been collected from the nasal and throat swabs. In this type of testing, RNA has been first converted into DNA [11] . This amplified DNA has been analyzed to predict the infection in the patient [12] [13] . Antigen based testing method is also available where the antigens are being analyzed in the sample. This testing would not be beneficial in giving the results of predicting whether the patient is currently infected with the virus or not and is also generating false-negative results [14] [15] . Testing plays an important part in combating the infection as if a state tests more and more people, after that only they will be able to identify the infected patients. COVID-19 infected patients are isolated and then their symptoms can only be treated as still there is no particular treatment is available in the market. Separation and isolation of infected person reduce the risk of community transfer which in result reduce the burden over the healthcare system. Artificial Intelligence has already proved its ability in testing various diseases like tuberculosis [16] , pneumonia [17] , and cancer. Artificial Intelligence is flying its flag in solving medical and healthcare problems like disease detection, drug detection, affected site detection, and many more [18] [19] . Deep learning J o u r n a l P r e -p r o o f Journal Pre-proof techniques and neural networks are being used for detecting pneumonia and tuberculosis using X-ray images [20] . Similar to which, we have developed a testing system which will detect the COVID-19 infected patient using X-ray images. The X-ray images are being analyzed and the system will be able to predict whether the concerned person is COVID-19 positive or not. It is very necessary to separate out the COVID-19 patient from the patients having pneumonia and influenza as symptoms are pretty much similar. We have taken care of this issue and our system will be able to identify the difference between the X-ray image of COVID-19 infected person and the pneumonia affected person.

We have been proposing a neural architecture that will be able to give efficient results in no time and is a way cheaper than any other testing method. This method does not require extra specialists and can be easily operable without any requirement of additional material.

There is an invisible competition going among various big research labs and institutes for finding a testing and treatment mechanism for curing the COVID-19 infected person. Digital image analysis has been made significant progress in the past recent years. There are various software and image processing techniques that are available to analyze the image and being able to gain the specific knowledge from those captured images. Deep learning techniques involving image analysis using a convolutional neural network is very significant in gaining the key features from the images. Convolutional neural networks have been used in various healthcare applications like detection of cancerous cells, pneumonia detection, localization of pneumonia patches in the X-ray image of lungs, detection of tumors in the brain, and heart disease detection [21] . Researchers and scientists are working on finding the best suitable method for testing the has been looking for in the sample in molecular testing [22] . Antibody testing is a best-suited example for serological testing where the blood sample has been collected from the patient's body and checked whether the antibodies are being present in the serum sample or not [23] . Based on the analysis of whether the person has been distinguished as COVID-19 infected patient or not. RT-PCR testing is one of the examples of the molecular testing method. Another testing method is, Reverse transcription-polymerase testing method is used for finding out the person having some COVID-19 related symptoms that whether he is COVID-19 positive or not [24] . In this method, they are analyzing the genetic part of the viral structure that has been present in the sample taken from the suspected person. On determining the desired genetic material, the person has been confirmed as COVID-19 infected patient [25] . There are other testing methods like Quidel test, qPCR testing, Antigen based testing, Isothermal amplification testing, plasmonic photothermal sensing-based testing, and CRISPER based testing [26] method which have been used by the doctors in order to test the COVID-19 suspected patients.

Having all of these different types of testing methods available in the market, why is there a need to developing a new testing mechanism for COVID-19. Innovation is everything and we should have to optimize every part of the process.

There are both pros and cons of these testing mechanisms. But cons surpass over pros in terms of economy, efficiency, and time taken to give the results.

RT-PCR testing method takes more time and results are not very efficient. Antibody testing is not giving the confirmation whether the person is still carrying the virus. Researchers are working on optimizing the testing mechanism in terms of reducing the time taken to provide the results which helps in early and accurate identification of the infected patients and indirectly will help in contact tracing. Deep learning has been considered as the best suitable option opted by the research fellows working in the computer vision field. CheXnet has been used for the classification of various pulmonary activities including pneumonia. creating an AI framework for detecting COVID-19 patients using smartphone embedded sensors [29] . We are proposing a deep learning model which will be based on InceptionNet and Nasnet combination where we are analyzing the X-ray images of the COVID-19 suspects. Our model is successfully predicting the COVID-19 suspect's persons whether they are normal or have been infected with the SARS-CoV2 virus. In the next section, we are discussing our proposed mechanism in detail.

Deep Learning techniques are widely used in the detection of various diseases like pneumonia. Digital image processing has proved its strength in healthcare, bioinformatics, drug design, and medical image analysis. These techniques have been automated most of the process in the medical area. We are using supervised deep learning techniques where we first train our system using a dataset. Based on the experience and the knowledge received from the training process, the system will work in finding the infection in the person when a new X-ray image has been supplemented to it. The artificial neural network used in our system is mainly based on the visual human cortex that can be used in gaining the information. In this work, we are proposing a model for classification of X-ray images from a person having COVID-19 infection, pneumonia infected person, and normal person. Our model is based on deep learning functionalities and is following supervised learning along with reinforcement learning. In supervised learning, the model has been trained over the labeled dataset so that the network is able to learn the features gaining from the corresponding input. We have As the X-ray images can easily be analyzed using an ordinary convolution neural network. The overall architecture of the proposed model has been shown in 

The CNNs can be stacked together to get a deeper network which may increase the performance of predicting the disease. Gradient upgrades are very difficult to transverse throughout the entire network. Deeper convolutional neural networks have a huge affinity towards over-fitting as the kernel size should be large due to the more global spread of information. And choosing the right kernel is seriously a tedious job. As in the COVID-19 pandemic and because of a large population, speedy inhibition on the infection spread rate, and large randomized dataset availability, inception block will be the best which has provided speedy, efficient and accurate results as compared to other deep learning models.

The use of the inception block is computationally less expensive in computing J o u r n a l P r e -p r o o f Journal Pre-proof the results due to much wider network. Therefore, instead of having a deeper network, we are creating a wider network applying varied size kernels over the image simultaneously at a particular level. The representation bottleneck has been removed which will reduce the computational expenses and also helps in preventing the loss of information. Label smoothing will be incorporated which will help in preventing over-fitting problems.

The reinforcement learning supports the neural architecture search space technique which will help in optimizing the regularization parameters and hyperparameters. The resulting concatenated output from the inception blocks and the NAS layer then be further supplemented to the dense blocks which will help in learning the features and be able to give accurate results. NAS has been added in automating the convolution features in the inception layer. Adding NAS block will help in optimizing the hyper-parameters. Nasnet will helps in learning the features from the large dataset by applying two featured convolution cells [30] .

These convolution cells are the normal cells and the reduced cells which return the feature maps. Normal cells return the feature maps of the same height and width whereas the reduced cells return the feature maps whose height and width can be reduced by some factor. The Nas block architecture for imagenet has been shown in the Fig. 2 .

Hill climbing technique has been used that applies network morphism where the network has been optimized based on the local search by finding the best solution based on an incremental approach. As we should have to be fully confident about the results that the system will be able to accurately identify.

We have included one of the NAS blocks parallel to the inception layer for getting an accurate result. The NAS network has been introduced in parallel to the convolutional blocks and max pool layer as it will add more CNN feature maps to the network and is acting like a skip connection from input to fully layer in the network will compliment with the locally converged features. All of this aids Nasnet in improving the metrics.

The output is fed to the dense layered block and then the soft-max classifier which will then predict the person category from analyzing the corresponding X-ray image. The dense block consists of a dense layer, batch normalization layer, relu and finally the dropout layer. For increasing the training speed and reduction in co-variance shift of the values, batch normalization layer and dense layer has been used. The dense layer has also been used for reducing the parameters used in the convolution layer. The dropout layer has been added in order to improve the generalization error where we are disabling some of the neurons in the network randomly. Using all these specific layers and developing a hybrid model introduced in this paper that has been used for giving better and faster results in testing on COVID-19 suspects.

Dense block is used in the hybrid model in order to extract the features. It is used for to deepen the network without using any additional parameters. All layers with their matching feature map sizes have been directly connected to each other with the help of dense block. Each layer is obtaining extra inputs J o u r n a l P r e -p r o o f Journal Pre-proof from all foregoing layers and the resulting feature map or we can say its own feature map has been passed to its following layers. This property will maintain the feed forward nature in the network. This will help in the reduction of the usage of number of extra parameters. Dense Block consists of dense layer, Batch Normalization layer, relu as an activation function, dropout layer.

Dense layer used is a deep network of neural layers. The output of the dense layer depends on the number of units specified in the dense layer. Using the dense layer will allow the largest potential function approximation in the limit of the provided width of the given layer. This will be used to avail the space under limited function approximations. In dense layer, each hidden unit in one layer has been mapped to every other hidden unit in the subsequent layer. On increasing the width of the network, the network could simultaneously approximate more functions resulting in the expansion of the solution space.

Network could use those parallel approximations resulting in making of more informed decisions by increasing the depth of the network and dense layer will help in that in order to increase the depth at the needed time.

Batch Normalization layer can be used to increase the training speed of the network. Using batch normalization we are actually decreasing the co-variate shift. Co-variate shift will help each layer of the network in learning on a more stable distribution of inputs. It is actually the variation in the dispersion of input values to a learning algorithm [31] . In hidden layers this is called internal co-variate shift which makes the training slow and will require very small learning rate and a good parameter initialization. Batch Normalization is a process 

(1) and adding a shift (β) and placing X as a non linearity's input.

To regularize our network we are using an additional dropout layer in the hybrid network. When hidden units or neurons of a specific layer have been always become influenced by the output of a particular hidden unit in the previous layer, it will reduce the efficiency of our model. To prevent this indirect biasing, we are make use of dropout layer which will helps in better training of our hybrid model ultimately leads to better efficiency and accuracy.The layer will reduce the extra dependent nature of particular hidden units in the neural network. Use of dropout layer finally results in reducing the over-fitting problem.

In this work, we use the open-source and publicly available chest X-ray dataset. In this, dataset consists of three different class image data such as normal, pneumonia, and COVID-19. The COVID-19 is a new pandemic disease, and the availability of this virus related images are very less.

For experimental analysis, we merged and modified three distinct datasets consisting of COVID-19 images. These three datasets are: 1) COVID-19 Chest X-ray Dataset Initiative [32] , RSNA Pneumonia Detection Challenge dataset [33] and COVID-19 Image Data Collection [34] . In this combined dataset 

The performance of our model have been measured on metrics like accuracy Accuracy(AC) = T P + T N T P + T N + F P + F N

The F1-score in terms of precision and recall is being calculated as follows:

Our novel hybrid approach has proved its capability and has made a step No additional equipment's are necessary for the proposed testing method.

In X-ray testing mechanism, the main problem that the doctors and the states are facing is about the availability of the specialized radiologist who have the necessary skills to identify the confirmed cases on analysing the X-ray report showing that our model has been performed well on the validation set as well.

The average validation loss for the different models have been shown in the Table 2 for better understanding of our proposed model's efficiency.

The performance of our hybrid proposed model for COVID-19 detection using X-ray images can be visualized using confusion matrix. The matrix will

give a clear understanding of our results and will help in analyzing the statistics of our classification task. Instances in a predicted class has been represented by the rows whereas the columns represents the instances of the actual class.

Besides, we have analyzed the performance of the other deep learning models using the confusion matrix. The comparison results and the confusion matrix of INASNET and other models has been shown in Fig.8 . Based on this we are proposing that our hybrid model works best on the COVID-19 detection and is way more efficient than any other deep learning model.

We have developed a hybrid inception model which will help the doctors and complex datasets and is optimized by having batch normalization and dense layer which will help in increasing the efficiency for detection. This will also helps the hospitals in having a better understanding of COVID-19 by having a significant amount of X-ray's for a person which makes them to have a regular check on the area that have been affected in the lungs due to SARS-CoV2.

The authors would like to thank the anonymous reviewers for their thorough review and valuable comments. This work was supported in part by grants Gov- 

A novel coronavirus from patients with pneumonia in China

A novel coronavirus genome identified in a cluster of pneumonia cases -Wuhan

Severe acute respiratory syndrome-related coronavirus: The species and its viruses-a statement of the Coronavirus Study Group

Global surveillance, national surveillance, and SARS

Middle East respiratory syndrome coronavirus transmission

Evidence for limited early spread of COVID-19 within the United States

The airborne lifetime of small speech droplets and their potential importance in SARS-CoV-2 transmission

How to Protect Yourself & Others. Centers for Disease Control and Prevention Website

Laboratory testing of 2019 novel coronavirus (2019-nCoV) in suspected human cases: interim guidance

Centers for Disease Control and Prevention

Archived from the original on 30

There's only one way to know if you have the coronavirus, and it involves machines full of spit and mucus

Update:FDA Authorizes First Antigen Test to Help in the Rapid Detection of the Virus that Causes COVID-19 in Patients 2020 US Food Drug Administration

FDA issues emergency approval of new antigen test that is cheaper, faster and simpler

Automatic detection of tuberculosis using deep learning methods

Deep convolutional neural networks for chest diseases detection

Analysis on novel coronavirus (COVID-19) using machine learning methods

Content based medical image retrieval using dictionary learning

Chexnet: Radiologist-level pneumonia detection on chest x-rays with deep learning

Coronary heart disease diagnosis using deep neural networks

New blood tests for antibodies could show true scale of coronavirus pandemic

The promise and peril of antibody testing for COVID-19

Helicase-dependent isothermal DNA amplification

Chemically modified guide RNAs enhance CRISPR-Cas genome editing in human primary cells

CRISPR-Cas12-based detection of SARS-CoV-2

An ensemble classifier system for early diagnosis of acute lymphoblastic leukemia in blood microscopic images

Lung infection quantification of COVID-19 in CT images with deep learning

A novel AI-enabled framework to diagnose coronavirus COVID-19 using smartphone embedded sensors: design study

Learning transferable architectures for scalable image recognition

Batch normalization: Accelerating deep network training by reducing internal covariate shift

Tackling the radiological society of North America pneumonia detection challenge

Covid-19 image data collection: Prospective predictions are the future

chest x-ray data initiative; 2020

Manuscript title: INASNET: Automatic Identification of Coronavirus Disease (COVID-19) Based on Chest X-ray Using Deep Neural NetworkThe authors whose names are listed immediately below certify that they have NO affiliations with or involvement in any organization or entity with any financial interest (such as honoraria; educational grants; participation in speakers' bureaus; membership, employment, consultancies, stock ownership, or other equity interest; and expert testimony or patent-licensing arrangements), or non-financial interest (such as personal or professional relationships, affiliations, knowledge or beliefs) in the subject matter or materials discussed in this manuscript.

Dr. METTU SRINIVAS Assistant Professor National Institute of Technology Warangal Telangana, India-506004