View of CNNCD Screening model to distinguish Covid-Virus by using chest X-ray and Transfer Learning

878

CNNCD Screening model to distinguish Covid-Virus by using chest X-ray and Transfer

Learning

Dr. Pritee parwekara_{, Swathi Rapeti}b_{, Parul Vats}c_{, and Megha Sharma}d

a

Associate Professor (CSE), SRM IST, NCR Campus

b,c,d _{M.Tech Student (CSE), SRM IST, NCR Campus}

Article History: Received: 10 January 2021; Revised: 12 February 2021; Accepted: 27 March 2021; Published online: 16 April

2021

_____________________________________________________________________________________________________ Abstract: Coronavirus disease has been announced as a pandemic by World Health Organization and till this date 2,683,536

are lost their lives due to Covid-19. The one and only way to reduce the cases is Quarantine the patients that who are tested Covid-19 positive. Researchers have done Different kind of design deep learning models to screen the Covid-19 pandemic. They discovered different deep learning models to detect the Covid-19 using chest X-Rays most of the methods having less accuracy rate. In few models Overfitting problem increasing difficulties in most of the models. In this Article an automatic Covid-19 Screening model is developed to identify the Covid Detection, Pneumonia and Normal. Different learning techniques used separately to learn the model like CNN, VGG16 and ResNet. From those three models VGG-16 is giving better performance.

Keywords: CNN, VGG16, ResNet, Custom Neural Network.

1. Introduction

Covid-19 is a virus causing by several kind of illness symptoms known as Serve acute Resipratory Syndrome Coronavirus(SARS-COV-2) and come up in Dec-2019 in Wuhan city in China. Later it spread by Person to person. Till Date Covid-19 has affected 121,547,884 from this 98,004,078 has recovered from Covid-19 Disease and 2,688,540 lost their lives due to this pandemic. The one and only quick solution to reduce Covid 19 Cases is to keep the patients in quarantine to stop spreading of this disease to other people. To keep the people in Quarantine that we need to identify the Cases. Number of infected people are increasing by day by day. To control, quarantine of the Covid-19 patients is the only way and then quick screening of the Covid-19 Virus is required. The method which is used to screening the Covid-19 is the detection of nucleic acid using RTPCR (Reverse transcription polymerase chain reaction). Sometimes it is giving false results and the patient should be tested again and again. It will cause waste of time and money.

Based on Clinical studies chest X-ray is an effective and fast screening technique, and it can be finding out the features of Covid-19 so in detecting of Covid-19 positive cases Chest X-ray is essential and Main. Number of radiologists are less than Number Positive Cases in the world so manual processing is very difficult to identify. Here we are doing automatic screening with the help of Artificial Intelligence it will speed up the detecting process as well as resolves the issues of waiting time, cost and unavailability of RT-PCR (reverse transcription polymerase chain reaction) Kits[1].

The automatic diagnosis of disease using machine learning (ML) & convolutional neural network have gained more popularity because fast diagnosis of Covid-19 using ML is possible. By providing the input data we can achieve best results here we don’t need any manpower. Deep learning and Convolutional neural network methods are involved in to increase the accuracy of the test results. The difficulties and complexity of the existing models are motivated me to propose a new Covid-19 automatic screening method. In this work a model called CNNCD, is proposed to screen COVID-19 by using Convulational Neural Network and Transfer learning technique like VGG16. Here the chest X-rays collected from the date sets of Kaggle.

2. Literature Survey:

A new tool to fight against of all odds to find the patients of COVID -19 and other respiratory viruses i.e., “Iteratively pruned deep learning ensembles for COVID-19 detection in chest X-ray” [7]. In this system they came with a new classification method which is CXR’s (Chest X-Rays) diagnosis. By make use of CXR group, they can tuned the transferred learned knowledge to upgrade the concert. This proposed system states about the strategies to improve classification performance of CXR’s. The result of this proposed system have an accuracy of 97.06 % and area under the curve is 0.9972 % . It has merge use of CX’R modality particular knowledge transfer; the main advancement of this Iterative model pruning ensemble learning will reduces the prediction

879

variance. The drawback of this proposed system is, it requires vast Dataset size to compare more classifications to get accurate results, inherent variability. In order to perform this model the installation process is complex[2]. 2. To detect COVID -19, University of Lakehead from Canada researchers came up with a new technology i.e., “DL-CRC: Deep Learning – based chest radiograph classification for COVID-19 detection: A Novel Approach” [8]. This proposed system discusses the leverages a data augmentation of radiograph images (DARI) algorithm with convolutional neural network (CNN). This DL – CRC also describes about the comparison of Dark CovidNet and Proposed system’s dataset. The accuracy of detection for this survey is 93.94 %. Some how every system has its own drawbacks, in this way, this DL – CRC system has also had some drawbacks. One of the major drawback is, this system has very low memory that’s why it had only few actual COVID – 19 chest X- ray image samples in the original dataset. Another drawback is installation process of CNN models, DARI GAN based augmentation[3].

3. University of Sidi Mohamed Ben Abdellah’s researchers, had introduced a noval system, which is “Development of clinical decision support system for the early detection of COVID -19 using deep learning based on chest radiographic images.” [9] This proposed system states that the deep learning builds on chest radiographic images in extensively learning procedure which will take out graphical characteristics of COVID – 19. This system will diagnose the 100 images of collected prototypical viral pneumonia cases with normal cases of 100 images. By introducing Deep Learning with a novel architecture of VGG (CNN) is very helpful to enhance the accuracy rate for classification. This proposed system had an accuracy rate of 92.5 % internal validation and 87.52 % in internal validation and also it have drawback like huge memory consumption[4].

4. Introduced an “IOT based deep learning framework for early assessment of COVID -19.” [10] This research introduces Internet of Medical Things (IOMT) along with data frame work and vast involvement of Machine learning, Deep Learning and Artificial Intelligence (AI) techniques. To swamp the constrain regional based Convolution Neural Network (CNN). This proposed system uses faster regions (RCNN) along with ResNet-101 is applied on X- ray images. This system not only works on Deep leaning and Artificial Intelligence, also uses RPN (Regional Proposal Network) to perform detect up to 7000 non COVID – 19 tests with 4000 COVID – 19 cases. On following the above scenario this research gives the results with an accuracy of 98%, this is one of the biggest and finest results in this area of research. The main aspect of this proposed system is to decrease the work pressure of medical experts. However, the proposed system is fine but even though it has some major drawbacks. One of the major drawbacks is assemblage of comprehensive dataset for the training of deep learning system is delicate. To compare the number of cases reported in all hospitals which are in the entire globe require freely available online dataset[5].

5. University of China-Japan union hospital researchers introduced a novel system “Intelligent detection for CT image of Covid-19 using deep learning”, which states that the new approach of classification methods to examine the CT scan images of Covid-19 by using deep learning target detection method on the base of time-Spatial sequence convolution. To get a sustain results with the above mentioned system has to satisfy the recurrent neural structure and a 2-D convolutional layer structure by the help of faster RCNN, YOLO3 and SSD algorithm models. In this way, the proposed system got more accurate results[6]. One of the major benefit of this entire proposed system is usage of FRCNN, YOLO3 and SSD to get speed and accurate results. This system has drawbacks like requirement of huge dataset rate and high cost effective .

6. Authors came up with a novel ideology in order to detect the Covid-19 patients that is Novel deep learning approach for classification of Covid-19 images” [15] . Which states about the classifications of x-ray and CT images of Covid-19 patients with a multiclass level algorithm. This proposed method have deep learning technique for Covid-19 patients versus non Covid-19 patients and multi-class of Covid-19 patients versus non Covid-19 versus pneumonia patients classifications from x-ray and CT images. To classify and enhance the patients disease is possible by using a 24 layer convolutional neural network. This CNN makes the input image are resized to 256 X 256 for making computization faster. By using this 24 layer CNN this proposed model got near about 99.68% of accuracy on x-rays and 71.81% of accuracy on CT scans. The major drawback of this model is the performance of classification in CT images can be improved with preprocessing. This may help to recognize the feature injected images[7].

3. Proposed Work:

In this project a Convolutional Neural Network Model CNNCD(Custom Neural Network for Covid Detection) is designed for the automated screening of Covid-19. In this CNNCD initially Convolutional Neural Network is used to extract the relevant features from the Convolutional layers by passing from the next layer which is pooling layer and convolutional layer and then It will form fully connected layer. Here we used new algorithm here it named it CNNCD(Custom Neural Network for Covid Detection). Here I am compared this model with VGG16 and VGG19. Here used the pretrained dataset that we collected from the Kaggle. This algorithm is like sequential model with 14 different layers. The workflow diagram of the proposed model is Fig.1 from that diagram can get the clear cut of this model that what is going to proposing here.

880

Figure1: Workflow Diagram

Data Preparation and Data Preprocessing:

We have Downloaded Dataset from the Kaggle the official website is Kaggle.com, from this Dataset Covid-19 and Pneumonia and Normal database is collected. in this there are three different kind of chest X-ray images, Covid-19 Positive, Pneumonia cases and Normal or no infection. Total 509 Covid-19 positive_images, 1342 Pneumonia_Images and 3875 Normal images.

881

Figure 3: Pneumonia chest X-ray images

Figure 4: Normal chest X-ray images

Classification of Dataset:

Here From the Dataset images are classified into 3 categories Covid, Pneumonia and Normal by using Convolutional Neural Network and CNNCD (Custom Neural Network for Covid detection) chest X-ray. Classification by using CNNCD:

CNNCD(Custom Neural Network for Covid Detection)By using this Model we can find Covid positive cases. In this model will provide input to the layer which is called input layer than the layer will read the preprocessed images of chest X-ray from the datasets the shape or size of the input is 224*224, then the first layer named convolutional layer which is having 896 parameters, in our model here the output shape from this layer is 222*222*32. This particular layer is used to find out the pattern of Covid-19 from those chest X-ray images and this layer will helpful to retrieve the features from the Particular Chest X-ray images by using different kind of filters. Then the layer will carry the filters with those images. Features which are retrieved from the layer will be the output to the forward or next layer of our model which is Max pooling Layer which is used to reducing the number of parameters by reducing the size of images.

Next Layer is ConV2D_1 this Convolutional Layer which is used to enhance edges and emboss. It will helpful to the sharpen images and here the output shape is 109*109*32 with 9248 parameters and after this there is a Max pooling layer again that is max_pooling2d_1 which is used to reduce the size of parameters and images here the output shape will be 54*54*32, Now Conv2d_2(Conv2D) in this layer the previous output will be the input to this layer after performing Convolution function the output shape is 52*52*32 with 9248 parameters. Again max_pooling2d_2 with the output shape of 26*26*32, again Convolutional layer which is Conv2d_3 with the output shape of (Conv2D) 24*24*64 with 18496 parameters, max_pooling2d_3 with the output shape of

882

12*12*64, Conv2d_4with the output shape of 10*10*64 with 36928 parameters, again max_pooling2d_4 with the output shape of 5*5*64 with 64 filters. This flatten layer is a function which is used to convert the feature map into single column which is passed to fully connected Layer.

Layer Construction of CNNCD Model:

Implementation of CNNCD Model:

Results:

883

Fig: Train Validation Confusion Matrix:

Accuracy: Accuracy is a ratio of total no.of records that correctly classified to total no.of all the records in Dataset.

• Accuracy = True positive + True Negative o True pos +True Neg + False Pos +false Neg

Precission: It is the ratio of positive records correctly classified to the total num of predicted positive in dataset o Precession = True Positive

True Pos + False Pos

Recall: Ratio of true positive records to the true positive + false Negative. o Recall= True Positive

True pos + Fal Neg Comparative Studies:

VGG16 (Visual Geometry Group): This is one of the most preferred CNN architecture developed by Simonyan and Zisserman in the year of 2014. This model having 16 layers. It is have unique architecture only point to worry about the VGG16 is it has 138 million parameters which is certainly difficult to handle.

ImageNet: Used to classify thousand different Objects.

VGG19: This is the one of the convolutional Neural Network but it does amost as well as the VGG19 so most of the people are using VGG16. VGG 16 is giving best results compare with VGG19.

ResNet: One of the Covulational Neural Network which is having 25.5 million parameters only. Compare with VGG16 this is having very less parameters. RedNet is Faster than VGG16. Residual Network allow us to train extremely Deep Neural Network with 150 Layers.

Table 1

884

Comparison of performance between different Classifiers Table 2: Multi class comparison of 3 different Classifiers

Classifier Disease Precisio

n Recall F-1 score CNN Covid-19 91% 95.97% 95.27% Normal 95.95% 92% 95.81% Pneumoni a 96% 92% 94% VGG- Covid-19 93.98% 95.67% 94.82% 16 Normal 96.62% 97.34% 97.48% Pneumoni a 97.35% 96.62% 97.49% ResNet- Covid-19 85.19% 95.52% 89.35% 50 Normal 95.11% 88.68% 93.55% Pneumoni a 90.76% 93.42% 93.97% Conclusion:

In this Project we have developed a new model which is CNNCD Custom Neural Network for Covid Detection. By using this Model we can detect Covid-19 Virus from Chest X-ray of the Covid-patients. We have done Multi level Classification by using this model on three different Classes Covid-19, Pneumonia and Normal or No Infection. We got the 95.6% Accuracy By using this CNNCD Model. I have compared this model with other Transfer learning Models like VGG16, VGG19 and Residual Net. Here CNNCD model is 97.2%Accuracy

885

with 35 epochs 95.68% accuracy with 25 epochs and 92.38% with 5 epochs. This model will be helpful to the medical institutes and hospitals to screen Covid Virus.

References:

Chen, J., Wu, L., Zhang, J., Zhang, L., Gong, D., Zhao, Y., Zhang, K. (2020). Deep learning-based model for detecting 2019 novel coronavirus pneumonia on high-resolution computed tomography: a prospective study. MedRxiv. 10.1101/2020.02.25.20021568. [Europe PMC free article] [Abstract]

Sivarama krihnan Raraman, Jenifer Siegelman, Lucass folio, Les r.folio and Sameer k. Antani.” Iteratively pruned deep learning ensembles for COVID-19 detection in chest X-ray”. Volume 8, Journal article, IEEE publisher.

Sadman Sakib, Tahrat Tazrin, Mostafa M. Fouda, Zubair MD. Fadullah and Guizani.” DL-CRC: Deep Learning – based chest radiograph classification for COVID-19 detection: A Novel Approach”. Volume 8, Journal article, IEEE publisher.

Sidi Mohamed Ben Abdellah’s researchers, M. Qjidaa, A. Ben – fares, Y. Mechbal, H. Amakdouf, M. Maaroufi, B. Alami and H. Qjidaa.” Developmet of clinical decision support system for the early detection of COVID -19 using deep learning based on chest radiographic images”. IEEE conference paper, IEEE publisher.

Imran Ahmed, Awais Ahmed and Gwanggil Jeon. “IOT based deep learning framework for early assessment of COVID -19”.Early Access Article, IEEE publisher.

Jingxin Liu, Zhong zhang, Lihuizu, Hairihan wang and Yutong zhong.” Intelligent detection for CT image of Covid-19 using deep learning”. IEEE conference paper, IEEE publisher.

Malaya kumari nath, Aniruddha kanhe, Madhusudhan mishra.” A Novel deep learning approach for classification of Covid-19 images”. IEEE conference paper, IEEE publisher.

Das AK, Kalam S, Kumar C, Sinha D,” TLCoV- An automated Covid-19 screening model using Transfer

Learning from chest X-ray images.”, DOI: 10.1016/j.chaos.2021.110713,

https://europepmc.org/article/MED/33526961, Elsevier public health emergency Collection. http://doctorpenguin.com/