Medical Image Classification for Coronavirus Disease (COVID-19) Using Convolutional Neural Networks

The coronavirus is a family of viruses that cause different dangerous diseases that lead to death. Two types of this virus have been previously found: SARS-CoV, which causes a severe respiratory syndrome, and MERS-CoV, which causes a respiratory syndrome in the Middle East. The latest coronavirus, originated in the Chinese city of Wuhan, is known as the COVID-19 pandemic. It is a new kind of coronavirus that can harm people and was first discovered in Dec. 2019. According to the statistics of the World Health Organization (WHO), the number of people infected with this serious disease has reached more than seven million people from all over the world. In Iraq, the number of people infected has reached more than twenty-two thousand people until April 2020. In this article, we have applied convolutional neural networks (ConvNets) for the detection of the accuracy of computed tomography (CT) coronavirus images that assist medical staffs in hospitals on categorization chest CT-coronavirus images at an early stage. The ConvNets are able to automatically learn and extract features from the medical image dataset. The objective of this study is to train the GoogleNet ConvNet architecture, using the COVID-CT dataset, to classify 425 CT-coronavirus images. The experimental results show that the validation accuracy of GoogleNet in training the dataset is 82.14% with an elapsed time of 74 minutes and 37 seconds. Keyword: CTcoronavirus image, GoogleNet, COVID-19, Deep learning, Convolutional neural networks. ديفهك( انورهك سوريف ضرمل ةيبطلا رهرلا فينرت 91 مادختساب ) ةيفيفلاتلا ةيبرعلا تاكبذلا لهجم ندحم دعم 9 * ، ميرك يمع نادحا 2 1 قا رعلا ،دادغب ،ةعماجلا ةيدارتقلاا مهمعمل دادغب ةيمك 2 قا رعلا ،فرشلاا فجشلا ،ةفهكلا ةعماج ،ينا رسعلا طيطختلا ةيمك ةصلاخلا سوريفلا يجاتلا هى ةمئاع نم تاسوريفلا يتلا ببدت ضا رمأ ةريطخ ةفمتخم يدؤت ىلإ تهسلا . مت رهثعلا ىمع نيعهن نم اذى سوريفلا اقًباس : سراس SARS-CoV ، يذلا ببدي ةمزلاتم ةيدفشت ةداح ، و MERSCoV ، يذلا ببدي ةمزلاتم ةيدفشت يف قرذلا طسولأا . فرعيُ ثدحأ تاسوريف ةيجاتلا يتلا تأذن يف ةشيدم ناىوو ةيشيرلا مساب ةحئاج COVID-19 . ونإ عهن ديدج نم تاسوريفلا ةيجاتلا يتلا نكسي نأ يذؤت ساشلا متو وفاذتكا لولأ ةرم يف ربسديد 2112 . اقًفو تاءارحلإ ةسعشم ةحرلا ةيسلاعلا ( WHO ) ، لصو ددع ISSN: 0067-2904 Mijwil and Al-Zubaidi Iraqi Journal of Science, 2021, Vol. 62, No. 8, pp: 2740-2747 2741 صاخشلأا نيبارسلا اذيب ضرسلا ريطخلا ىلإ رثكأ نم ةعبس نييلام صخش نم يف عيسج ءاحنأ ملاعلا لصو ددع نيبارسلا يف قا رعلا ىلإ رثكأ نم 22 فلأ صخش نادين ريش ةياغل 2121 . يف هذى ةلاقسلا ، اشسق قيبطتب تاكبذلا ةيبرعلا ةيفيفلاتلا ( ConvNets ) نم لجأ ةقد فذكلا نع رهص ريهرتلا يجاتلا يعطقسلا بسهحسلا ( CT ) يتلا دعادت مقاطلا يبطلا يف تايفذتدسلا ىمع فيشرت ةرهص CTcoronavirus يف ةمحرم ةركبم . نإ ConvNets ةرداق ىمع ممعتلا جا رختساو تا زيسلا ايًئاقمت نم ةعهسجم تانايب رهرلا ةيبطلا . فديلا نم هذى ةسا ردلا هى بيردت ةيشب GoogleNet ConvNet مادختساب ةعهسجم تانايب COVID-CT فيشرتل 425 ةرهص CT-coronavirus . تريظأ جئاتشلا ةيبيرجتلا نأ ةقد ةكبش GoogleNet يى 42.14 ٪ يف بيردت ةعهسجم تانايبلا عم تقهلا يزقشسلا هى 44 ةقيقد و 34 ةيناث .


Introduction
The COVID-19 virus belongs to the coronavirus (CoV) subfamily [1,2] of the viruses family, and specifically, to the beta genus (beta-coronavirus). The SARS-CoV, that caused the outbreak that emerged in China in 2003 [3] [4], and MERS-CoV, that caused the outbreak that appeared in the Arabian Peninsula in 2012 [5], are of zoonotic origin (the original host is the bat), which at one point have evolved and crossed the barrier between species until causing the pandemic [6]. It should be noted that there are other coronaviruses that affect humans on a daily basis, causing mild diseases of the respiratory system, such as common colds (229E, NL63, OC43, or HKU1), all of which are from the alpha-coronavirus subgroup [7].
They receive this name (corona) for the structure they possess which, as seen under a microscope, gives them a kind of crown on their outer surface [8]. Viruses consist essentially of genetic material and structural proteins that encapsulate it. Figure-1 presents the structure of COVID-19 [9]. They consist of the nucleocapsid, where the genetic material is strictly contained (exclusively a simple ribonucleic acid (RNA) sequence of around 32,000 bases) and packaged thanks to protein N, and the envelope which is made up of various structural proteins such as the membrane glycoprotein or M protein, involved in virus assembly and in contact with the nucleocapsid, protein S, which forms the spikes responsible for adhesion to the host cell, and protein E, which interacts with protein M for the formation of the envelope [10,11]. These represent the most relevant structural proteins, although others are vital for the replicative process of the virus as well as infection and entry into cells. It can be observed that the genomic sequence of COVID-19 is relatively similar to that of SARS-CoV (79%) and somewhat more different from that of MERS-CoV (50%) [12,13]. COVID-19 is characterized by having a relatively high mutation capacity compared to other viruses, which makes the development of specific diagnostic methods, therapies, and vaccines somewhat complex.
The pandemic is still spreading with high speed and afflicts a huge number of people of different ages and genders, with no medicine developed. All institutions, schools, and universities have been closed and curfew and quarantine measures have been imposed in many cities all over the world. Tourism and travelling abroad have been banned. The critical question is when will a medicine be produced for this pandemic that hits the planet. Very immediately, many studies appeared in 2020 about the use of deep learning techniques to classify a set of coronavirus images. The study conducted by Butt et al. [14] used multiple CNN models to classify and compare CT images of infections with coronavirus, influenza virus, pneumonia, and non-infection. The results indicated 98.2% sensitivity and 92.2% specificity. Narin et al. [15] suggested automatic CNNs-based transfer models (InceptionV3, Inception-ResNetV2, and ResNet50) for the prediction of COVID-19 in chest Xray images. The ResNet50 model had the best accuracy result for training, which is 98%. Singh et al. [16] proposed a multi-objective differential evolutionbased CNNs models for classifying chest CT images for COVID-19 patients. The results showed accuracy of 1.9789%, F-measure of 2.0928%, sensitivity of 1.8262%, specificity of 1.6827%, and Kappa statistics of 1.9276%. Ozturka et al. [17] proposed DarkNet model as a classifier for YOLO real-time object detection system, with 17 convolutional layers and several filtering steps in each layer. This study obtained accuracy values of 98.08% and 87.02% for binary classes and multi-class cases, respectively. Amyar et al. [18] proposed a multi-task deep learning for the classification and segmentation of CT COVID-19 images. The study included a dataset of more than 10000 patients (449 patients with the pandemic, 100 not infected, 98 with lung cancer, and 397 of other cases). The results showed a Dice coefficient value of < 0.78 for the segmentation and an area under the ROC curve value of < 93% for the classification. Another investigation [19] proposed deep learning and support vector machine (SVM) to detect COVID-19 disease. The results, based on the X-ray images, achieved accuracy = 95.33%, sensitivity= 95.33%, FPR= 2.33%, F1= 95.34%.

2.
Chest CT-COVID-19 Medical Image This test can assist to diagnose and monitor diseases, such as lung cancer, heart failure, sarcoidosis, pneumonia, tuberculosis, lung fibrosis, and COVID-19 disease. Physicians use X-rays to find out if treatments are effective and to check for any complications after a procedure or surgery. The following figures demonstrate infections with COVID-19 for people of different ages.

Figure-3 [21] shows a coronavirus infection CT image for A:
A man (35yr) with COVID-19 who suffers from severe headaches and fever after one day. CT-scan explains pure ground-glass opacity in the right lower lobe (red frame). B: A man (47yr) with COVID-19 patient who suffers from fever after seven days. CT-scan displays consolidation in the right lobe subpleural area (red frame).

The Proposed Method
The proposed method includes 3 steps (pre-processing, retraining GoogleNet ConvNet, and classification). Figure-4 presents the steps of the proposed method, which are also described below. The CT-coronavirus image contains a set of unnecessary information when it is taken. In order to obtain an adequate and perfect image, the recorded CT-coronavirus images must be pre-processed. 2 nd step: In this step, some operations are executed on a CT-coronavirus image, e.g. the enhancement, normalization, noise reduction, filtering, binarization, and thinning. The binarization process is applied since it is the most important pre-processing step that is applied before retaining the image dataset. Besides, this step includes converting the image from 2D grayscale to 3D colour as well as saving these images. The goal of the conversion to three-dimensional images is that CNNs accept this type of images. The Matlab function was applied for this process, as follows:

Classification Pretrained GoogleNet CNN classifier
In the binarization process, the function is applied to create a binary image of grayscale (2D)/colour (3D). This function is performed by replacing all values above a globally determined threshold ( ) with and setting all other values to . The function uses Otsu's segmentation method, which determines the value to minimize the intraclass variance of the black and white pixels.
utilizes a 256-bin image histogram to compute Otsu's threshold. Applying the Otsu segmentation method, a robust is computed based on the enhanced histogram. The best is obtained using an exhaustive search based on all likely values from 0 to 255. 3 rd Step: Classification according to GoogleNet ConvNet architecture to retrain CT-dataset achieves high accuracy and reduces computing costs. GoogleNet is identified as InceptionV1. It presents the new theory of the inception block in ConvNet, in which multiscale convolutional transformations are integrated using split, transformation, and merge ideas. This block encapsulates filters of various sizes to obtain spatial information at various scales. Figure-5 shows the sample images that have been used in this implementation. Then, convolutional layers are replaced in tiny blocks, as in the idea of replacing each layer with micro-NN, as suggested in the NIN (Network in Network) architecture. GoogleNet's use of the idea of a split, transform, and merge assists solving a problem associated with learning different types of variations in the same class of several images. Besides, to increase learning capacity, GoogleNet focuses on making ConvNet parameters useful. 4 th step: After retaining GoogleNet, each newly entered COVID-CT image can be classified.

Experiment Results
ConvNets are one of the deep learning methods applied in image classification, detection, and object recognition. The essential ConvNet architecture involves the convolution layer, pooling layer, and fully connected layer. All images obtained from the Google Images are unmodified and nonduplicate, as presented in Figure-5. Actually, it was a very difficult and tedious task to collect these images in high-resolution. There are not many images related to this topic on the Google site. Most of the studies are recent, where a pre-print version is also available on the Medrxiv and ResearchGate sites. All of these images are uploaded to the authors' profiles on the Mendeley website and can be used and shared by everyone. The proposed model execution runs by MATLAB v.2018 with a computer that has specifications of GPU=GeForce GTX 1070, CPU= Core i7 (Intel), RAM=16 GB, and Win=10 home. The dataset includes 425 CT-COVID images. Figure-6 shows the results of classification by GoogleNet ConvNet. Table-1 presents the present results of the effects obtained by the classification, in terms of the values of accuracy, sensitivity, specificity, and precision. ReLu is an activation function that is applied in all layers. To explain the metrics used when applying the classifier and obtaining results, it is necessary to define some concepts: Accuracy: percentage of CT-scan images classified correctly. That is, we evaluate the percentage of success that the network has in our test set, as shown in equation 1:  The details in Figure-6 show the following: The validation accuracy is 82.14%. Elapsed time is 74 min and 37 sec (The more layers, the more accurate the results are, with 23 layers being applied in this paper). No. of iterations is 592. No. of epochs is 8. In addition, no. of iterations per epoch is 74.

Mijwil and Al-Zubaidi
Iraqi Journal of Science, 2021, Vol. 62, No. 8, pp: 2740-2747 2746 The training opportunities in the research were as in the following:

Conclusions and Future works
Artificial intelligence techniques play an important and effective role in uncovering the COVID-19 pandemic. Specialists employ artificial intelligence techniques to examine this virus, test possible treatments, diagnose infected and uninfected people, and analyse the effects on public health. In this paper, we apply a deep learning technique (GoogleNet ConvNet) for the classification of a set of CTcoronavirus images (425 images) with InceptionV1 model. The experimental results confirmed that the validation accuracy of this test is 82.14%. In addition, this model achieved a precision of 84.01%, sensitivity of 83.12%, and specificity of 82.11%. To the best of our knowledge, this study is the first to use this approach. There shall be other studies in the future that deal with the application of artificial intelligence techniques in classifying and detecting the X-ray and/or CT-scan of COVID-19 pandemic and comparing them with images of viral pneumonia and bacteria.