key: cord-0724182-80afi168 authors: Mohsin Ahmed, Hanaa; Wael Abdullah, Basma title: Overview of deep learning models for identification Covid-19 date: 2021-06-11 journal: Mater Today Proc DOI: 10.1016/j.matpr.2021.05.553 sha: 15542d47d52ddcd317150059cc489bba26b48330 doc_id: 724182 cord_uid: 80afi168 The well-being and health of global population is continuously and badly affected by COVID-19 pandemic. Thus, to prevent the spread the pandemic between individuals, there is high importance in implementing automatic detection systems as rapid alternative diagnosis. The virus is affecting the person’s respiratory system as well as creating white patchy shadows in the X-ray images of the lungs of individuals experiencing COVID-19. Also, deep learning can be defined as a useful and efficient AI technique used for analyzing chest X-ray images for reliable and effective screening of COVID-19; therefore, distinguishing people infected with COVID-19 and normal persons, and after that the infected individuals will be isolated for mitigating the virus spread. This study provides an overview regarding a few of the modern deep learning-based COVID-19, with design steps and types, also it compares the diagnostic method of COVID-19 with other methods of deep learning created with the use of radiology images. After a comparison between the most recent methods used in the previous works, it was found that RestNet50 pre-trained and DCNN model gives accuracy of 98%, which is the highest reported so far from among other proposed models were discussed in this paper. Coronavirus disease 2019 (COVID- 19) can be defined as one of the infectious diseases resulting from severe acute respiratory syndrome coronavirus-2 (SARS-CoV-2). Initially, COVID-19 has been identified in the year 2019 in Wuhan, China, and from that time, it was spread rapidly and globally, leading to the (2019-2020) coronavirus pandemic [1] . Many industries were affected and a lot of people were quarantined due to the spread of COVID-19, and this resulted in the fact that the life quality of human is devastatingly affected. COVID-19, majorly referred to as novel coronavirus turned into a pandemic disease causing a lot of mortalities. Later, this continuous outbreak was considered as a global public-health emergency by the World Health Organization (WHO) [2] . In many countries, legislatures are forcing flight limitations, fringe limitations, social-distancing and increasing awareness of hygiene. Yet, the spread of the virus is still rapid [3] . Because of the COVID-19 0 s high-transmissibility, its detection is of high importance in the way for controlling and planning to pre-vent the disease. At the same time, the limitations in experts and medical equipment and detective tools caused the detection of the disease to be slow. Therefore, it increased the number of casualties and patients. When quickly detecting the disease, there will be a decrease in its number of casualties and its prevalence [4] . The first phase is getting the detection, recognizing the disease symptoms, and using distinctive signs for accurately detecting the disease. Based on the disease type, symptoms might be shortness of breath, cough, acute respiratory problems, and common fever and cold, while for a few days, patients might also have a cough for no clear reason. Dissimilar to SARS, COVID-19 is not just affecting the respiratory system, yet also it affects other organs in the body, like liver and kidneys. Typically, symptoms of a new coronavirus resulting in COVID-19 start some days following an individual becomes infected, while in a few individuals, symptoms might show later [4] . It is assumed that the elderly individuals experiencing basic clinical issues such as diabetes, cardiovascular ailment, hepatic or renal maladies, ceaseless respiratory infection and malignant growth were bound to create genuine disease [3] . One of the major COVID-19 symptoms are the respiratory problems, which might be identified via chest's X-ray imaging. Also, a disease with mild symptoms might be detected via chest CT scans. Typically, the detection can be done by analyzing the signal data and images obtained with the techniques of medical imaging like MRI, CT and X-ray by means of deep learning models. Due to such analysis, one can diagnose and detect many diseases like brain tumor, diabetes mellitus, breast and skin cancer, and so on [7] . Therefore, examining such images might accurately detect the disease's existence in suspected individuals and even individuals with no initial symptoms. The use of such data might be also covering other issues, like the limitations of diagnostic kits as well as their production. The major benefit of CT scan is the availability of CT scan devices in the majority of laboratories and hospitals, while doctors are majorly using CT scan images for detecting the infections. Without the typical symptoms, like fever, using the chest CT scans has a fairly excellent capability for detecting the disease [4] . In laboratories, one of the major ways to detect COVID-19 is by human experts. Thus, injuries in chest radiology image and symptoms are examined via the specialist for detecting the existence of COVID-19 from a healthy person experiencing other diseases. However, such process has high costs and, especially, long-term detection. Recently, deep learning and computer vision were applied for detecting various lesions and diseases in the body in automatic way. A few instances are: detecting tumor volume and types in the brain, head, lungs and so on, considering that X-ray images and chest CT scans are the major approaches to diagnose COVID-19. In addition, the chest X-rays are showing multiple white patchy shadows in the lungs of a person affected with COVID-19. Using deep learning and computer vision might be of high importance in diagnosing COVID-19. Since the spread of the disease, a lot of studies utilized deep learning and machine vision techniques and they reported excellent results [4] . AI is used effectively in various tasks and fields [5] . Deep learning can be defined as one of the AI functions used for imitating the way that human brain works in data processing and generating patterns to be used in decision-making. Also, deep learning is one of the machine learning fields, while machine learning is a type of AI [6] . Deep learning is often referred to as deep neural network or deep neural learning. Recently, the models of deep learning were effectively utilized in many fields and resulted in excellent performances on many problems, like segmentation, classification, visual recognition, speech recognition, NLP and in medical image processing. The models of deep learning were effectively utilized in different areas like segmentation, classification and lesion detection of medical data [7] [8] [9] . The presented study provides an overview of the modern related works regarding the use of deep learning for detecting COVID-19 and other diseases, while also it provides a comparison between them. Besides, this study involves the deep learning image classifiers, and it discusses the major phases for diagnosing the disease. The remaining parts of this study are presented in the following way: the related works are provided in section (2), deep learning image classifiers are presented in section (3), deep learning steps are presented in section 4, lastly, the conclusions are presented in section 5. 1. In 2016, Srdjan Sladojevic and Marko Arsenovic [10] , provided a novel method to use deep learning for automatically classifying and detecting plant diseases from leaf images. In addition, the created model has the ability for detecting leaf existence and distinguishing between healthy leaves and 13 different diseases, which might be diagnosed visually. New plant disease image database has been formed, which consists of at least 3000 original images obtained from available internet sources and extended to no less than 30,000 with the use of adequate transformations. Furthermore, the experimental results reached a precision between (91-98) %, for separate class tests. The final overall accuracy related to the trained model has been 96.3%. Fine tuning didn't show considerable alterations in the overall accuracy, yet the process of augmentation showed more impact for achieving good results. 2. In 2017, U. Rajendra Acharya and Hamido Fujita and et al [11] , suggested a CNN method for detecting (automatically) the various segments of ECG. The signals of ECG have been acquired from a publicly-available arrhythmia database. Also, the study acquired V-Fib (Ventricular Fibrillation) ECG signals from Creighton University ventricular tachyarrhythmia, AFL (Atrial Flutter) and A-Fib (Atrial Fibrillation) ECG signals from MIT-BIH atrial fibrillation, and AFL (Atrial Flutter), A-Fib (Atrial Fibrillation) and NSR (Normal Sinus Rhythm) ECG signals from MIT-BIH arrhythmia database. The signals of ECG were utilized for 2 s and 5 s durations with no QRS detection. Also, the study achieved accuracy, specificity and sensitivity of 92.50%, 93.13% and 98.09%, for 2 s of ECG segments. They obtained sensitivity of 99.13%, accuracy of 94.90% and specificity of 81.44% for 5 s ECG duration. Furthermore, the algorithm is serving as one of the adjunct tools for assisting clinicians in verifying their diagnosis. The drawbacks of the suggested algorithm is that it requires a lot of data (big data) for training and takes more time for training the data. 3. In 2018, Fang Liu, PhD and Zhaoye Zhou, PhD and et al [12] , created a fully-automated deep learning-based cartilage lesion detection system by means of classification and segmentation CNNs. In addition, the fat suppressed T2weighted fast spin-echo MRI datasets regarding the knees of 175 patients experiencing knee pain have been analyzed (retrospectively) via the use of deep learning. Besides, Receiver Operating Curve (ROC) analysis as well as k statistic have been utilized for assessing the diagnostic performance and intra observer agreements to detect cartilage lesions for 2 individual evaluations done via the cartilage lesion detection system. Furthermore, the results of specificity and sensitivity associated with the cartilage lesion detection system at optimal threshold based on Youden index have been 85.2% and 84.1%, for evaluation 1 and 87.9% and 80.5%, for evaluation 2. Areas within the ROC curve have been 0.914 and 0.917 for evaluations 2 and 1, specifying high overall diagnostic accuracy to detect cartilage lesions. 4. In 2019, Ali Narin and Ceren Kaya and et al [13] [14] [15] [16] [17] [16] . They evaluated the efficiency of the state-of-art CNN models for the classifications of the medical images. In particular, the the process that has been referred to as the transfer learning has been implemented. With the transfer learning, detecting a variety of the anomalies in the small medical image datasets has been considered as one of the achievable targets, which usually yields outstanding results. The data-set that has been used in the present work represents a set of 1427 X-Ray images, 700 of the images have been with confirmed common pneumonia, 224 of them with confirmed Covid-19, and 504 of them with normal conditions have been included. A general 97.82% accuracy has been accomplished in detecting Covid-19. 9. In 2020, Sohaib Asif, Yi Wenhui, and et al [3] . they used deep convolutional neural networks (DCNN) for the automatic detection of the patients with COVID-19 pneumonia with the use of the digital x-ray images of the chest. The dataset includes 1345 viral pneumonia images, 864 COVID-19, and 1341 of the x-ray images of normal chest. The DCNN based model InceptionV3 with the transfer learning were suggested for detecting patients that have been infected by coronavirus pneumonia, with the use of the X-ray radiographs of the chest and gives an over 98% classification accuracy. 10. In 2020, L. Wang, Z. Q. Lin and et al [17] . They introduced COVID-Net, a DCNN design has been modeled for detecting the cases of COVID-19 from chest X-ray (CXR) images. They introduced as well COVID-x, which is a data-set that comprises 13,975 chest X-ray images over 13,870 patient cases. They have researched the way that COVID-Net makes predictions by using a method of explain ability as an attempt for not merely gaining more knowledge about the critical factors that are related to the cases of COVID-19 that may be helpful for the clinicians for improving the screening, however, also auditing COVID-Net as a transparent and responsible way for the validation of the fact that it makes the decisions according to the relevant information that has been obtained from chest X-ray images. In final the COV-IDNet has been capable of achieving high accuracy that reached up to 93.3% test accuracy. 11. In 2021, Ghulam Gilanie, and et al [18] They proposed an automatic approach of detecting Covid-19 with the use of the CNNs. There have been 3 data-sets obtained from the Radiology Department (i.e. Diagnostics). The utilized dataset included CT as well as X-Ray images (7021 images of both pneumonia and normal, whereas there have been 1066 images with Covid-19 infection). The suggested approach has been capable of achieving an average specificity (95.65%), accuracy (96.68%), and sensitivity (96.24%). 12. In 2021, Fudan Zheng, Liang Li and et al [19] . They have developed a computed tomography system of image diagnosis through the deep learning for the rapid diagnosis of COVID-19 through the integration of the Res-Net with the SE blocks. This architecture was successful in the identification of COVID-19 CT from the healthy individuals' CT, typical viral pneumonia patient CT, and bacterial pneumonia patient CT separately. They have obtained CT images of 262, 219, 100, and 78 individuals for COVID19, typical viral pneumonia [35] [36] [37] , bacterial pneumonia, and healthy individuals, respectively. The model has been capable of achieving a general recall of 0.94, precision of 0.95, and accuracy of 0.94. Table1, lists a comparison between the works that have been explained, by the utilized image types, number of the utilized cases, utilized methods, limitations and accuracy. It had discovered that the X-ray and the CT images are the only one may be utilized. The accuracy of the suggested system is dependent upon booth number of used images, and Deep CNN architecture. Finally, the limitations regarded to the number of the used images. Comparison of the results obtained from state-of-art approaches was presented in Table1. In [13] Covid19 and non-Covid19 images were classified with a general achieved 98% accuracy (with the use of the RestNet-50 pretrained model). Similarly, the [3] had classified Covid19 and non-Covid19 and Viral Pneumonia images that consist of small data-set, in other words, normal (1341), viral pneumonia (1345), and Covid-19 (864) with general achieved also 98% accuracy [42] , which has been defined as the highest reported yet. In the following some of available state-of-art deep learning classifiers of the images [15] . 1. VGG-19: which stands for Visual Geometry Group Network has been developed according to the CNN model that has been presented by Oxford Robotics Institute's A. Zisserman and K. Simonyan [20] . VGG-Net's performance has been quite beneficial on Image-Net dataset. For the purpose of improving the functionality of the image extraction, VGG-Net utilized smaller 3x3 filters, in comparison with the Alex-Net 11x11 filter. There are 2 versions of this network; which are: VGG-16 and VGG-19 have various layers and depths. 2. Inception-V3: Inception network or GoogLe-Net has been a 22layer network and won the 2014 Image net challenge with 93.30% top-5 accuracy. The InceptionV3 network includes a number of the symmetrical and asymmetrical building blocks, in which every one of the blocks had a number of the branches of the convolution, Pooling Layers also, could be called subsampling layers, is reduce the spatial size of the feature maps. concatenated, dropouts, this layer is used to prevent overfittings [21] , and fully-connected layers This is a traditional Multi-Layer Perceptron (MLP) in the fully connected layer all the neurons in the layer are connected to all the neurons in the next layer, this layer is used for classification task [22, 23] Can benefit from transfer learning by using one of this a pretrained model in the CNN layer because the transfer learning approach is a very effective method to build a model with high accuracy, especially when there is limited small data [25] . There are three main steps of deep learning to conduct the diagnostic procedure of disease (Fig. 1) , as [15] : Step1: Pre-processing All the X-ray images were obtained in 1 data-set and loaded for the scaling at a fixed 224 Â 224 pixel size for being proper for the additional processing within deep learning pipe-line. One-hot encoding [26] has been applied afterwards on image data labels for indicating the case for every one of the images in the data-set. Step2: Training Model and Validation For the purpose of starting the one deep learning model's training phase, the pre-processed data-set has been divided 80-20 based on Pareto principle. Which indicates 20% of the image data is utilized in the phase of the testing. Once more, the splitting of 80% data is utilized to construct equal sets of training and validations. Sub-sample arbitrary training image data selections for DL classifier, and apply after that metrics of evaluation for showing the performance that has been recorded on the set of the validation and Achieving high accuracy of the system comes from the high performance of the system with few errors [27] . Step3: Classification In the last framework's step, the data of testing is provided to tuned deep learning classifier for the categorization of all of image patches. Every classifier's accuracy has been viewed as a key parameter to evaluate the efficiency of every one of the classifiers [28] . At workflow's end, the general efficiency analyses for every one of the deep learning classifiers will be assessed according to the metrics that have been described in Table2. [29, 30] . TP, TN, FP, and FN that have been given in Eq. (1) -(5) are the number of the True Positive, True Negative, False Positive, and False Negative, respectively. Given a test data-set and model [31] . Where [32, 33] : True positive (TP): which specifies the number of the positive samples that were correctly classified. False Negative (FN): which specifies the number of positive samples that were incorrectly classified. Infectious COVID19 shocked the entire globe and remains threatening for the lives of billions of the people. In the present work, Comparison of results that have been obtained from stateof-art approaches was listed in (Table1) for identification best method of deep learning for the classification of the COVID-19 patients from the normal ones through the use of CXR and CT images. Show performance evaluations to methods in Table 1 that the RestNet50 pre-trained and DCNN model gives better results in comparison with other available approaches with about 98% accuracy. However, the studies (mentioned previously) carried out their experimentations on much smaller number of the images (size of Covid-19 samples is small), which possibly could not have enough variability, and Another limitation is that the algorithms in these studies mentioned above took more time to train the data. For the future works, we intend to make our model in the next article more accurate and robust by overcome limitations through using more Covid-19 images from the local hospitals, reduce the time used for training and build a model consisting of a large number of layers. CRediT authorship contribution statement Hanaa Mohsin Ahmed: Conceptualization, Methodology, Software, Data curation, Investigation, Supervision, Software, Validation. Basma Wael Abdullah: Writing -review & editing, Writing -original draft, Visualization. Table 2 The metrics questions. Equation Name Equation 1 Accuracy (TN + TP)/(TN + TP + FN + FP) 2 Recall TP/(TP + FN) 3 Specificity TN/(TN + FP) 4 Precision TP/(TP + FP) 5 F1-Score 2 Â ((PrecisionxRecall)/ (Precision + Recall)) Automatic Detection of COVID-19 Infection from Chest X-ray using Deep Learning Classification of COVID-19 from Chest X-ray images using Deep Convolutional Neural Networks A new modified deep convolutional neural network for detecting COVID-19 from X-ray images Improvement of genetic algorithm using artificial bee colony Hand gesture recognition of static letters American sign language (ASL) using deep learning Deep CNN based skin lesion image denoising and segmentation using active contour method An efficient approach for detecting and classifying moving vehicles in a video based monitoring system Documents classification based on deep learning Deep Neural Networks Based Recognition of Plant Diseases by Leaf Image Classification Automated Detection of Arrhythmias Using Different Intervals of Tachycardia ECG Segments with Convolutional Neural Network Deep Learning Approach for Evaluating Knee MR Images: Achieving High Diagnostic Performance for Cartilage Lesion Detection Automatic Detection of Coronavirus Disease (COVID-19) Using X-ray Images and Deep Convolutional Neural Networks Senior Member, IEEE, ReCoNet: Multi-level Preprocessing of Chest Xrays for COVID-19 Detection Using Convolutional Neural Networks A Framework of Deep Learning Classifiers to Diagnose COVID-19 in X-Ray Images Covid-19: Automatic detection from X-Ray images utilizing Transfer Learning with Convolutional Neural Networks COVID-Net: A Tailored Deep Convolutional Neural Network Design for Detection of COVID-19 Cases from Chest X-Ray Images Usama Ijaz Bajwa, Mustansar Mahmood Waraich. Coronavirus (COVID-19) detection from chest radiology images using convolutional neural networks Accurately Discriminating COVID-19 from Viral and Bacterial Pneumonia According to CT Images Via Deep Learning Very Deep Convolutional Networks for Large-Scale ImageRecognition Intelligent system for imposter detection: A survey, Mater. Today Proc Modified Siamese Convolutional Neural Network for Fusion Multimodal Biometrics at Feature Level, 2nd Scientific Conference of Computer Sciences (SCCS) The Impact of Filter Size and Number of Filters on Classification Accuracy A comparative study of fine-tuning deep learningmodels for plant disease identification Robust Real-Time Violence Detection in Video Using CNN and LSTM, 2nd Scientific Conference of Compter Sciences 3 -Sequential logic design Pupil Detection Algorithm Based on Feature Extraction for Eye Gaze Detection of confusion behavior using a facial expression based on different classification algorithms Diagnosis of lung cancer disease based on back-propagation artificial neural network algorithm Stream cipher with space-time block code Review on real time background extraction: models, applications, environments, challenges and evaluation approaches An efficient prediction model based on machine learning techniques for prediction of the stock market Hybrid intelligent techniques for text categorization, International Conference on Advanced Computer Science Applications and Technologies Manufacturing intelligent Corvus corone module for a secured two way image transmission under WSN Intelligent secured two-way image transmission using corvus corone module over WSN, Wireless Pers Efficient cooperative image transmission in one-way multi-hop sensor network Visual sensor intelligent module based image transmission in industrial manufacturing for monitoring and manipulation problems Hyper spectral image classification using dimensionality reduction techniques Image transmission over decode and forward based cooperative wireless multimedia sensor networks for Rayleigh fading channels in medical Internet of Things (MIoT) for remote health-care and health communication monitoring Effectiveness of artificial intelligence techniques against cyber security risks apply of IT industry Fluid-structure interaction during water hammer in a pipeline with different performance mechanisms of viscoelastic supports Coronavirus disease (COVID-19) cases analysis using machine-learning applications Novel unilateral dental expander appliance (udex): a compound innovative materials The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.