A comprehensive review of COVID-19 detection with machine learning and deep learning techniques

Abstract

Purpose

The first transmission of coronavirus to humans started in Wuhan city of China, took the shape of a pandemic called Corona Virus Disease 2019 (COVID-19), and posed a principal threat to the entire world. The researchers are trying to inculcate artificial intelligence (Machine learning or deep learning models) for the efficient detection of COVID-19. This research explores all the existing machine learning (ML) or deep learning (DL) models, used for COVID-19 detection which may help the researcher to explore in different directions. The main purpose of this review article is to present a compact overview of the application of artificial intelligence to the research experts, helping them to explore the future scopes of improvement.

Methods

The researchers have used various machine learning, deep learning, and a combination of machine and deep learning models for extracting significant features and classifying various health conditions in COVID-19 patients. For this purpose, the researchers have utilized different image modalities such as CT-Scan, X-Ray, etc. This study has collected over 200 research papers from various repositories like Google Scholar, PubMed, Web of Science, etc. These research papers were passed through various levels of scrutiny and finally, 50 research articles were selected.

Results

In those listed articles, the ML / DL models showed an accuracy of 99% and above while performing the classification of COVID-19. This study has also presented various clinical applications of various research. This study specifies the importance of various machine and deep learning models in the field of medical diagnosis and research.

Conclusion

In conclusion, it is evident that ML/DL models have made significant progress in recent years, but there are still limitations that need to be addressed. Overfitting is one such limitation that can lead to incorrect predictions and overburdening of the models. The research community must continue to work towards finding ways to overcome these limitations and make machine and deep learning models even more effective and efficient. Through this ongoing research and development, we can expect even greater advances in the future.

Introduction

A virus is a protein-coated microscopic organism that hijacks the host cells for its development and in turn, makes the host sick. The Severe Acute Respiratory Syndrome Coronavirus Disease 2019 (SARS-CoV-2) is a special type of coronavirus, which inflicts mild symptoms to severe respiratory distress, multi-organ failure even death [1]. In November-December 2019, an outbreak of unknown pneumonia (caused by SARS-CoV-2) began in Wuhan, China, which was later named COVID-19. On 11th March 2020, finally World Health Organization (WHO) declared the COVID19 as a pandemic. Till that time, this deadly virus had already displayed its symptoms in different countries [2]. As of 31st October 2021, the most affected countries are the United States of America, India, Brazil, United Kingdom. India, with 34,246,157 confirmed cases, ranked the second most affected country succeeding the USA [3].

Due to the collective efforts of researchers, doctors, and vaccines; the world is gradually getting out from the havoc. But, still, humanity have a long way to go. So the continuous efforts are still going in the researches to develop various efficient techniques to deal with this pandemic. Efficient Diagnosis is one of those most researched domains. There is a two-step technique for the diagnosis, one is laboratory technique and the other is medical imaging technique.

Laboratory techniques mainly deal with the analysis of sputum and nasopharyngeal lesions. It includes two techniques, one is Rapid Antigen Test (RAT), where the infected patients are revealed when he / she is at the peak of its infection, by detecting the viral protein which is present in bulk during this time. And another technique is Reverse Transcriptase Polymerase Chain Reaction (RT-PCR) which detects the viral RNA in nasopharyngeal swab during the initial stage of the disease, helping in early isolation. Among these two, RT-PCR is the most efficient but it is time-consuming and costly [4]. The second step i.e. Medical Imaging Technique (CT-Scan /X-Ray) imaging is necessary as sometimes RT-PCR produces a false negative. In Radiology Imaging i.e CT-scan / X-Ray are passed through the lungs to detect various characteristics for lung involvement which includes ground-glass opacity, air bronchograms, reverse halo, and perilobular pattern with a specificity of 100% [5].The radiologists play a very important role, while detecting the presence of the above mentioned features. But, if it is replaced by AI, then the process of detection will be faster [6].

In a highly populated and developing country like India, the limitation of performing RT-PCR [7] is more vivid because of its high cost, less availability, less reliability, and more time-consuming nature. Towards the path of the solution to this problem, the researcher came across the application of Artificial Intelligence on CT-Scan / X-Ray. In this paper, 50 research articles of both 2020 and 2021 are reviewed from some popular digital archives, which utilized various ML / DL models, for learning the characteristic feature of COVID from Lung CT-Scans / X-Rays of suspected patients and by using that knowledge the models classifies the COVID-19 from other viral/bacterial pneumonia and healthy ones and some of them showed a very impressive accuracy of about 99% (approx.).

The paper is structured in the following pattern. Section 1 reveals the introduction of the research article. Section 2 provides a brief description of the search mechanism for review. Section 3 contains a brief review of methodology which is used by the research articles as a pre-requisite for classification. Section 4 briefly discusses all AI models utilized for the above requirement. The Future scope and Conclusion are depicted in Sects. 5 and 6 respectively.

Search mechanism

Google Scholar, PubMed, Science Direct (S.D), Web of Science (WoS), arXiv and medRxiv, are some of the popular digital archives which acted as the library for the research articles. Fig. 1 describes the flow of the search and selection mechanism for the current systemetic research article review. These resources are generally considered reliable in the field of medical, bio-engineering, and computer science studies. These digital archives were selected for their originality and adequate scope of studies. Important search keywords were chosen to get hold of the relevant research articles from these archives. “COVID-19”, “SARS-CoV-2”, “Coronovirus”, “Machine learning”, “Deep learning”, “Detection”, “diagnosis”, “prediction”, “classification” etc. are some of the keywords which have been used to search the research articles. Many articles performing COVID-19 detection, using ML and DL models, are retrieved. The article screening is performed where the duplicate articles and articles dealing with data from modalities other than radiology such as protein sequence, laboratory findings, etc. are removed. The remaining articles were sorted on the basis of a complete context study and 50 articles from radiological modalities (X-Ray and CT-Scan) are chosen for the systematic review purpose, from both 2020 and 2021. The data collected are ready for extraction and segregation, based on different techniques used by the authors during feature extraction and classification. The research articles, which are filtered by thorough screening, are studied thoroughly. The common methodology followed by the researchers to perform their research is drafted in Sect. 3.

Fig. 1.

Search and selection mechanism flow for systematic research article review

Methodolody

A number of 50 research articles are selected from various journals, where classifications are performed for X-Rays / CT-Scans images with COVID-19 v/s Non-COVID-19 ( healthy, other viral or bacterial pneumonia), using machine learning (ML) and deep learning (DL) models. All of these follows a sequence of common procedure which are as follows:

Dataset collection

The research articles, throughout this study, used CT-Scan / X-Ray modalities as datasets. They are collected through various public data repositories, hospital databases as well as from closed experimental groups. The database consist of X-rays / CT-Scan images which mainly belongs to COVID-19, other type of Viral/Bacterial Pneumonia and healthy classes. Some of them are presented in Table 1. The various features are extracted, which are further useful for determining, whether a particular input image belongs to a certain class or not.

Table 1.

Descrition of various datasets

Dataset Name	COVID19	Healthy	Other Pneumonia	Ref. no. of the derived dataset	Ref. no of the paper
CXR	371	1341	1345	[58, 60–65]	[8]
Dataset-1	216	837	838	[103, 104]	[10]
Dataset-2	214	670	671	[105]	[10]
SARS-CoV-2 CT scan	1250	615	615	[59, 66–75]	[14]
CXR dataset	220	27	-	[107]	[15]
COVID-CT	2282	4888	4888	[108]	[16]
SARS-CoV-2 CT-scan	1252	615	615	[109]	[17]
X-Ray	296	1341	3875	[109, 110]	[18]
CT images	79	90	90	[57, 76–95]	[21]
X-Ray	247	1341	3875	[59 , 96–106]	[25]

Pre-processing techniques

The scarcity of public datasets leads to a condition that is well known as overfitting. Data scarcity can be caused due to the limited availability of labeled training data or data imbalance. To prevent this, the least abundant datasets are subjected to pre-processing or augmentation through various techniques. Some of the most used preprocessing techniques may include contrast adjustment, intensity adjustment, brightness adjustment, cropping, scaling, resizing, flipping, or rotating at some angle. In some research articles, authors have also used a special type of machine learning framework, Generative adversial network(GAN), to recreate the images from the available datasets with realistic characteristics.

Segmentation

The detection of the region of interest has helped the researchers to obtain an efficient COVID-19 detection system. The U-Net architecture is based on the Convolutional Neural Network(CNN) where informations, from both upsampling and downsampling paths, are combined to perform better segmentation. The researchers also have used a 3D version of U-Net architecture to extract contextual information in CT Scan images. V-Net and VB-Net facilitate volumetric medical image segmentation improving the detection of infected lungs segment.

Feature extraction

In most of the research articles, the feature is generally extracted by deep learning models due to its self-learning capabilities. The feature extractors extract important features from CT-Scan / X-ray images and these features are trained by the models for solving the classification problem. The different techniques used by the authors include VGG, Inception, ResNet, DenseNet, Xception, EfficientNet, and many more. VGG is a CNN which uses 224 x 224 input channels, 3x3 convolution layers with 1x1 convolutional filters followed with ReLU activation function, three fully connected layers, hidden layers with ReLU, and local response normalization (LRN). Inception-v3 (GoogleNet) is a special kind of CNN with inception layer which is a combination of different filters such as 1x1 convolutional layer, 3x3 convolutional layer, 5x5 convolutional layer, 3x3 max-pooling layer with the outputs combined to produce a single output vector for the next stage. ResNet is a popular CNN that supports layers with double or triple skip and batch normalization between the layers to prevent the problem of vanishing gradient. DenseNet is a densely connected CNN with various advantages like strengthening feature propagation, reuse of features, and reducing the possibility of vanishing gradient problems. Xception is an extension of Inception architecture which converts the parallel convolutional layers into a horizontally inseparable convolutional layer. MobileNet is a streamlined architecture with depth-wise separable convolution layers producing a lightweight CNN. EfficientNet uniformly measures all the dimensions (depth, width, and resolutions) through a set of scaling coefficients. Multilayer Perceptron is a layered feed-forward Artificial Neural network with at least 3 layers (input layer, an output layer, and Hidden layer) that utilized backpropagation and non-linear activation functions. The recurrent neural network processes, both inputs at the present state and the output, of the previous state to produce the output of the present state through the different layers.

Classification

Mainly classification is of two types, binary and multiclass classification. In binary classification, the classification is performed between two classes (COVID +ve v/s COVID –ve), whereas in multi-class classification, the classification is performed among many classes (COVID v/s healthy, other viral/bacterial pneumonia, and other pulmonary diseases). The extracted features play an important role in this whole process as they are learned by different models during the training phase. The researchers have used a wide variety of machine and deep learning models during their implementation of the COVID-19 detection system. Some of these models are K-Nearest neighbor (KNN), Logistic regression (LR), Decision tree, XGBoost, AdaBoost, Bagging classifier, Support vector machine (SVM), and many more. KNN classifies a particular object depending on the class of its K (an integer) nearest neighboring object. Logistic Regression is used to classify categorical datasets based on the probability of the presence of an object in a particular class. A Decision Tree is a supervised machine learning algorithm in which data are divided into different decision nodes from the root node based on some conditions the leaf nodes will denote the corresponding outcomes. XGBoost is a machine learning model in which decision trees are added to rectify the previous errors called boosting and here gradient boosting algorithm is implemented. Adaboost is an ensemble machine learning algorithm where the above-mentioned boosting is applied to binary classification. Bagging classifier is another ensemble machine learning algorithm that will help the base classifiers to produce a much powerful classifier in-order to provide more accurate classification. SVM is a supervised machine learning algorithm that is primarily used for binary classification.

Discussion and result

In this survey paper, a few papers are collected on the basis of search mechanisms which are further segregated into the following categories i.e., machine learning, deep learning, and ML + DL. Figure 2a displays the frequency of different models used in the review study. In a number of studies, the researchers have utilized different deep learning models for its self-learning capability of learning features from the CT-Scan / X-Ray image dataset. However, the combination of the deep learning model as a feature extractor and the machine learning model as the classifier is also popularly used by researchers to produce an efficient system. The diagnosis is considered a crucial event for the effective treatment of COVID-19. Therefore, a thorough investigation is required and radiological modalities do achieve this. X-Ray images are the most preferable modality for research as shown in Fig. 2b.

Fig. 2.

a. Rate of using different types of models during the survey, 2. b. Rate of using different modalities during the survey

Table 2 summarizes the selected research articles on the basis of modalities, type of models, techniques used along with accuracy. The reviewed research articles mainly emphasize the study of accuracies of different models (ML / DL / ML+DL) for the classification of COVID-19 v/s non-COVID-19 and other pulmonary diseases. In the field of medical science, machine learning models are found to be useful for the classification of medical images. The features are extracted from the given dataset by algorithms such as hybrid social group optimization algorithm (Singh et.al.) and MrFODE algorithm (Elaziz et.al.) followed by classification using different ML classifiers such as SVM, random forest, KNN, etc. Among the three ML-based research articles, the model proposed by Elaziz et.al. achieved the highest accuracy of 99.65%. In particular, deep learning algorithms can be used for both feature extraction and classification. The most commonly used models may include VGG16, VGG19, ResNet50, Xception, Inception, etc which are either trained from scratch or by transfer learning. The authors have also proposed some new models such as DarkNet, ChexNet, VDSNet, DCovNet, etc. Mahmood et al. implemented the Xception model (X-Ray) as well as the Inception V3 (CT-Scan) model with the following results i.e. accuracy of 99.94%. The hybrid model utilizes the advantages of both machine learning and deep learning models. The CT-Scan / X-Ray images are represented in the form of a matrix of pixels. The deep learning models pick up the important pixels to form different feature matrices and these features matrices are learned by the machine during the training phase using ML models. Wang et al. designed a hybrid model i.e. Xception+SVM which obtained the best accuracy of 99.33% among the other DL+ML models.

Table 2.

Summary of selected research articles

Author	References	Image Modality	Type of AI model	Techniques used	Prediction results
Singh et al.	[8]	X-Ray	ML	Hybrid Social Group Optimization Algorithm- extraction and filtration of features SVM-classification of features	Accuracy obtained by the system is 99.65%
Tang et al.	[9]	CT-Scan	ML	Random forest	Accuracy is 87.5%
Elaziz et.al.	[10]	X-Ray	ML	FrMEMs moment+ MRFODE algorithm+ KNN classifier, operators of MRFO	MRFODE Dataset1 has an accuracy of 96.09% Dataset2 has an accuracy of 98.09%
Abbas et al.	[11]	X-Ray	DL	Three pre-trained model like AlexNet,GoogleNet,SqueexeNet	Two ClassClassification AlexNet has an accuracy of 99.77% GoogleNet has an accuracy of 99.70% SqueexeNet has an accuracy of 99.85% Three Class Classification AlexNet has an accuracy of 97.59% GoogleNet has an accuracy of 96.20% SqueezeNet has an accuracy of 97.47%
Mahmood et al.	[12]	X-Ray + CT-Scan	DL	Xception and Inception V3	Xception on X-Ray dataset and InceptionV3 on CT-Scan obtains an Accuracy 99.94%,
Biswas et.al.	[13]	CT-Scan	DL	Combined VGG16, ResNet50 and Xception using Transfer Learning	Classification accuracy obtained by this system is 98.79%
Shankar et al.	[14]	X-Ray	DL	Barnacle Mating Optimization with cascaded recurrent neural Network(BMO-cRNN)	Accuracy obtained by the system is 97.31%
Mansour et al.	[15]	X-Ray	DL	Inception v4 with Adagrad technique is used for Feature Extraction and Unsupervised DL based Variational Autoencoder(UDL-VAE) for classification	Accuracy obtained by the system is 98.7%
Jangam et al.	[16]	CT-Scan	DL	Stacked ensemble heterogeneous pre-trained Computer Vision model(VGG19,ResNet101,DenseNet169,WideResNet50	On Dataset1 accuracy is 85.71% On Dataset2 Accuracy is 99% On Dataset3 Accuracy is 93.5%
Sarki et al.	[17]	X-Ray	DL	VGG16, Inception V3 and Xception trained using Transfer learning + from scratch	In 1st case, Tertiary Classification obtains an accuracy of 87.50% and in 2nd case the same obtains an accuracy of 93.75%
Elmuogy et al.	[18]	CT-Scan	DL	Worried Deep Neural Network(WDNN) with transfer learning	Accuracy obtained 99.046%
Wang et al.	[19]	CT-Scan	DL	DenseNet121-FPN + COVID-19Net	Accuracy for dataset1 is 78.32% and Accuracy for dataset2 is 80.12%
Zheng et al.	[20]	CT-Scan	DL	UNet, 3D deep neural network ( DeCoVNet)	Accuracy obtained is 90.1%
Wang et al.	[21]	CT-Scan	DL	M-Inception, ROI images extraction and Deep Learning algorithm	Accuracy obtained is 89.5%
Farooq et al.	[22]	X-Ray	DL	Covid ResNet	Accuracy obtained is 96.23%
Maghdid et al.	[23]	X-Ray + CT-Scan	DL	Modified AlexNet model	Accuracy obtained is 94.1%
Apostolopoulos et al.	[24]	X-Ray	DL	VGG19 + MobileNet v2 + Inception + Xception + Inception + ResNet v2	DATASET_1 The best among the CNN models is VGG19 which obtains the accuracy of 98.75% (2-class classification) and 93.48% (3 class classification) DATASET_2 MobileNet v2 shows the highest accuracy 96.78% (2-class) and 94.72% (3-Class)
Alom et al.	[25]	X-Ray + CT Scan	DL	Inception Recurrent Residual Neural Network (IRRCNN)	Accuracy obtained by the proposed model on X-Ray is 84.67% and CT-Scan is 98.78%
Loey et al.	[26]	X-Ray	DL	Alexnet + Googlenet + Restnet18 + GAN	Four Classes GoogleNet has an accuracy of 80.3% Three Classes AlexNet has an accuracy of 85.2% Two Classes GoogleNet has an accuracy of 100%
Butt et al.	[27]	CT-Scan	DL	3D CNN Model (ResNet 18 for feature extraction and ResNet 23 for classification)	Accuracy obtained 86.7%
Rajaraman et al.	[28]	X-Ray	DL	Custom WRN + VGG-16, Inception-V3 + Xception + DenseNet-121 + NasNet-mobile	VGG16 obtains the Best accuracy of 93.02%
Asanaoui et al.	[29]	CT-Scan + X-Ray	DL	VGG19 + DenseNet121 + ResNetV2+ InceptionV3+ InceptionResNetV2 + Xception + MobileNetV2	Inception_Resnet_V2 obtains the best accuracy of 92.18 %
Pathak et al.	[30]	CT scan	DL	ResNet50 + Transfer Learning.	Accuracy obtained is 93.01%
Asif et al.	[31]	X-Ray	DL	DCNN + Inception V3	Accuracy is 98%
Ozturk et al.	[32]	X-Ray	DL	DarkNet	Accuracy obtained by the proposed model for COVID v/s No-Findings=98.08% and COVID v/s No-Findings v/s Pneumonia=87.02%
Brunese et al.	[33]	X-Ray	DL	Three fold model is used (1st and 2nd models is based on VGG16 group used to determine whether anX-ray belongs to COVID19 and 3rd model used to determine the effected region of X-rays if it belongs to COVID19), Gradient-weighted Class Activation Mapping (Grad-CAM) algorithm	Accuracy obtained by the 1st model is 96% and by the 2nd model is 90%
Haque et al.	[34]	X-Ray	DL	CNN Model 1 (4 convolutional layer), CNN Model 2 (3 convolutional layer and CNN Model3 (5 convolutional layers)	Accuracy obtained by Model 1 is 97.56%, by Model 2 = 96.34% and by Model 3 = 96.34%
Haghanifar et al.	[35]	X-Ray	DL	COVID - CXNet (ChexNet)	Accuracy obtained by COVID-CXNet v1 is 99.04% and COVID-CXNet v2 is 98.62%
Ucar et al.	[36]	X-Ray	DL	Deep Bayes - Squeeze Net based CNN(Convolutional neural network)	Overall accuracy obtained is 98.26%
Das et al.	[37]	X-Ray	DL	Extreme Inception (Xception) model (Automated deep transfer-based approach)	Accuracy obtained is 97.41%
Bharati et al.	[38]	X-Ray	DL	Hybrid CNN VGG Data STN (VDSNet)	Overall accuracy is 73%
Sakib et al.	[39]	X-Ray	DL	2D CNN (DL-CRC) model	Accuracy obtained is 93.94%
Misra et al.	[40]	X-Ray	DL	MultiChannel pre-trained ResNet architecture	Accuracy obtained is 95.5%
Mishra et al.	[41]	CT-Scan	DL	A fusion decision model is created by using five DL models VGG16, Inception V3, ResNet50, DenseNet121, DenseNet201	Accuracy obtained is 88.34%
Panwar et al.	[42]	X-Ray + CT-Scan	DL	VGG19 model (a fully connected layer and five extra layers)	Overall accuracy obtained is 89.47%
Abbas et al.	[11]	X-Ray	DL	Deep Convolutional neural network (DeTraC (Decompose, Transfer and Compose))	Accuracy obtained is 93.1%
Zhang et al.	[43]	X-Ray	DL	Multilayer Perceptron, 18-layer ResNet, EfficientNet-B0	The overall accuracy is 95.18%
Hall et al.	[44]	X-Ray	DL	VGG16 and ResNet50	The overall accuracy obtained is 91.24%
Abraham et al.	[45]	X-Ray	DL	Resnet-101, Densenet-201, InceptionResnetV2, NasnetLarge, Xception, Shufflenet, Darknet-53, MobilenetV2, VGG-19 and Squeezenet (pre-trained CNN model) Bayesian Classifier	Accuracy is 97.4%
Sethy et al.	[46]	X-Ray	DL+ML	AlexNet, VGG16, VGG19, GoogleNet, ResNet18, ResNet50, ResNet101, InceptionV3, InceptionResNetV2, DenseNet201 and XceptionNet (Feature extraction) SVM.(Classification)	ResNet50 +SVM obtained the best accuracy of 95.38%
Kumar et al.	[47]	X-Ray	DL+ML	ResNet152 architecture (feature extraction) Classifier: logistic regression, decision trees, k-nearest neighbour, naïve bayes, random forest, adaptive boosting, and XGBoost	Accuracy obtained is 97.7% (ResNet152+XGBoost)
Kassani et al.	[48]	X-Ray + CT Scan	DL+ML	Feature extraction using CNN : MobileNet, Xception, DenseNet ResNet, InceptionRes- NetV2, InceptionV3, VGGNet, NASNet Classifier: decision tree, random forest, XGBoost, AdaBoost, bagging classifier and LightGBM.	The best performance is achieved DenseNet121 feature extractor with Bagging tree classifier with 99% classification accuracy.
Basu et al.	[49]	X-Ray	DL+ML	CNN models: VGGNet (16 layers), AlexNet (8 layers), and ResNet(feature extraction) linear classifier (Classifier)	Accuracy is 85.98%(ResNet+Linear classifier)
Wang et al.	[50]	X-Ray	DL+ML	Feature extractor: CNN models: VGG16, InceptionV3, ResNet50, DenseNet121, and Xception Machine learning classifiers: decision tree, random forest, support vector machine, AdaBoost, Bagging.	Xception + SVM obtains the best Accuracy of 99.33%.
Shi et al.	[51]	CT-Scan	DL+ML	VB-Net (Image Segmentation) combined with V-Net (bottleneck layers), LASSO (feature selection), classification: logistic regression, support vector machine, and neural network, Semantics Aware Random Forest	Accuracy is 87.9%
Sharifrazi et al.	[52]	X-Ray	DL+ML	A fusion of CNN+SVM+Sobel filter with with 10 fold cross validation technique	Accuracy obtained is 99.02%
Das et al.	[53]	X-Ray	DL+ML	VGG19 for feature extraction Bi-stage Classification approach 1st to divide the features into normal and infected(bacterial and COVID) 2nd this infected features are used the classify the images using ML(CART,K-NN,LR,RF,XGB)	Accuracy is 99.26%
Sen et al.	[54]	CT-Scan	DL+ML	1.CNN to extract the features from CT scans 2.A bi-stage FS(Feature Selection)in which in relevant features are filteres using two filter methods(Mutual Information (MI) and Relief-F ) and Dragonfly algorithm (DA) 3.SVM used the features to classify between COVID and non-COVID CT scans	Accuracy from dataset1 is 98.39% And that from 2nd dataset is 90.0%
Qaid et al.	[55]	X-Ray	DL+ML	Hybrid model composed of three blocks 1)CNN block 2)VGG16 or VGG19 using transfer learning 3)CNN with ML algorithm (naiive Bayes, Support Vector Machine, random forest and XGBoost)	Accuracy obtained by this hybrid model for multiclass classification is 97.8%
Saba et al.	[56]	CT-Scan	DL+ML	ML based (KNN and RF) Tranfer Learning based (VGG19 and InceptionV3) Custom designed DL (CNN and iCNN)	Accuracies of KNN, RF, VGG19, InceptionV3, CNN and iCNN are 74.58 ± 2.44%, 96.84 ± 2.6, 94.84 ± 2.85%, 99.53 ± 0.75%, 99.53 ± 1.05%, and 99.69 ± 0.66% respectively.

From the above references, we can observe that ML and DL show the accuracy of more than 99% for the classification of COVID-19 from non-COVID-19 using X-Rays and CT-Scan images which are quite impressive. This review of published research articles shows the importance of different ML and DL algorithm for designing the medical image diagnosis system. The diagnosis system under the supervision of radiologists can be helpful to predict COVID-19 more efficiently. Thus asymptomatic cases can also be detected more accurately further reducing the transmission risk. This diagnosis system can also act as the second step for diagnosis after RT-PCR.

Clinical applications

In this section, the clinical applications for the ML and DL models are summarized. The authors have mentioned various scopes for the application of their proposed mechanisms. Singh et al. concluded that the Hybrid Social Group Optimization Algorithm [8] can be implemented as a mobile application, to make it available at a cheaper price for the early detection of COVID-19. The FrMems algorithm [10] stated by Elaziz et al., can be proposed in various medical domains including early COVID-19 detection. The model presented by Mahmood et al. [12], is capable to inform patients regarding COVID-19-infection through an email notification. Zheng et al. [20] have identified a better-performing model for the detection of lesions using COVID-19-infected CT-Scan data. The suggested methodology by Wang et al. [21] and Farooq et al. [22] have accurately identified whether the patients are healthy or infected by COVID-19. The algorithm proposed by Maghdid et.al [23], can be used to design a mobile application system to identify COVID-19 using the data of the symptoms collected from sensors. Apostopoulos et al. [24] and Alom et al. [25] have concluded that their proposed work can be used to develop an automatic COVID-19 detection system, as it provides higher accuracy while classifying COVID-19 from others. The proposed technique by Loey et al. [26] classifies the stages of COVID-19 utilizing the X-ray medical modality data with higher accuracy and lesser false negative results. Butt et al. presented a DL model in the article [27] which classifies healthy, other viral infections, and COVID-19 patients with an accuracy of 86.7%. Rajaraman et al. [28] have proposed and stated that their methodology can be used for implementation of the future COVID-19 detection system as it showcases lesser false negative results. El Asanoui [29] et al. presented a model that can be utilized, not only in the automatic detection of COVID-19 but also in other medical domains. Other than these applications, the DL and ML models can also be used in other aspects. They can be applied as a clinical diagnostic decision system where they will help medical professionals to detect disease more efficiently and within less time. They can help in predicting the survival rate after getting infected by a disease (namely COVID-19) in the form of a survival prediction system. In turn, they can also assist doctors in designing a better-personalized treatment plan to ensure quick recovery of the patient. Thus, the ML and DL algorithms can be utilized in various forms as efficient clinical tools. But, there are various limitations in its realistic implementation which are mentioned in the following section along with its possible scope of research.

Challenges and future research direction

The implementation of different ML, DL, and combinational models is accompanied by several challenges. The researcher’s hard work has provided us with various types of diagnosis and detection strategies for COVID-19. Although, researchers have found out some of the suitable models for detection of the virus, the knowledge available is still insufficient. The unavailability of a considerable amount of clinical image data (generally CT scan and X-Ray images), could decrease the degree of reliabilty of these models.. The misclassification of the given image can be considered as another challenge, that signifies the presence of unclear, noisy, inaccurate, missing, and redundant data, which makes the datasets unfit to be used as the training dataset. Thus the machine trained using these images are neither feasible nor they are reliable.With the passing time, more and more COVID-19 CT-Scan / X-Ray datasets will be available. And with these the ML / DL models will be able to get trained more effectively, increasing their efficiency and reliability for the detection of the disease.

Conclusion

All researchers are in search of a convenient solution to hinder this pandemic. Till now, AI proved to be one of the best diagnostic tool for many medical avenues, so researchers are trying to inculcate AI in the field of diagnosis of COVID19. Amidst the requirement of the quick detection of COVID19, in this review paper, an ensemble of some of the most popular ML and DL models, are created which are used for the detection of COVID-19. Through this paper, it is being hoped that all the researchers will get a decent idea about the COVID-19 detection models using ML / DL and medical images. Thus, more efficient models could be built, which would help in relaxing the overburdened healthcare systems and saving lives from this deadly virus. In this paper, some of the ML / DL models are discussed and and analyzed. After all such review and analysis, it is concluded that their combination can result in achieving an efficient and effective system for COVID-19 diagnosis with an efficiency of 99.33%. The limitation of these models is the result of the less availability of image data of patients suffering from COVID-19. As this disease is only one and half years old, not much COVID19 data are available in the databases. The efficient prediction of COVID-19 may not be possible due to the scarcity of the dataset. For the limited images, the model would work well for the training data but is not generalized enough to give the correct result for new input data due to overfitting conditions. If somehow someone is wrongly predicted then it may lead to more transmission and overburdening of the healthcare system. Thus, these models could only be used if they are efficient and accurate for detecting this disease. Among the different models, some of them are giving a satisfying performance with a small dataset, which is evident that AI models have a huge potential to fill the gap in COVID-19 detection. Therefore, the researchers are hopeful and working towards making the automatic detection of COVID-19 to make it a reality sooner. Until then, we all the citizens have to maintain the social distancing and get vaccinated to curb the spread of the havoc named COVID-19.

Abbreviations

ML

Machine Learning.

DL

Deep Learning.

AI

Artificial Intelligence.

WHO

World Health Organization.

SARS-CoV-2

Severe Acute Respiratory Syndrome Coronavirus 2.

COVID-19

Corona Virus Disease 2019.

CNN

Convolutional Neural Network.

KNN

K-Nearest Neighbour.

SVM

Support Vector Machine.

LR

Logistic Regression

RF

Random Forest

XGBoost

eXtreme Gradient Boosting

RAT

Rapid Antigen Test

RTPCR

Reverse Transcriptase Polymerase Chain Reaction

S.D

Science Direct

WoS

Web of Science

Availability of data and materials

Not applicable.

Declarations

Ethical approval

Not applicable.

Consent to participate

Not applicable.

Consent to publication

Not applicable.

Conflict of interests

The authors declare that they have no conflict of interest.