COVID-19-The Role of Artificial Intelligence, Machine Learning, and Deep Learning: A Newfangled

Abstract

The absolute previously infected novel coronavirus (COVID-19) was found in Wuhan, China, in December 2019. The COVID-19 epidemic has spread to more than 220 nations and territories globally and has altogether influenced each part of our day-to-day lives. As of 9th March 2022, a total aggregate of 44,78,82,185 (60,07,317) contaminated (dead) COVID-19 cases were accounted for all over the world. The quantities of contaminated cases passing despite everything increment essentially and do not indicate a controlled circumstance. The scope of this paper is to address this issue by presenting a comprehensive and comparative analysis of the existing Machine Learning (ML), Deep Learning (DL) and Artificial Intelligence (AI) based approaches used in significance in reacting to the COVID-19 epidemic and diagnosing the severe impacts. The paper provides, firstly, an overview of COVID-19 infection and highlights of this article; Secondly, an overview of exploring various executive innovations by utilizing different resources to stop the spread of COVID-19; Thirdly, a comparison of existing predicting methods of COVID-19 in the literature, with focus on ML, DL and AI-driven techniques with performance metrics; and finally, a discussion on the results of the work as well as future scope.

Introduction

The new Coronavirus or COVID-19, by WHO, could be an open-world emergency. These days, COVID-19 is swiftly transmitting from its source in Wuhan City, China, to everywhere throughout the world [1–4]. The complications of fighting the novel Coronavirus epidemic have put an end to a few financial and socio-social activities in numerous social methods worldwide. On the contrary hand, it has set off a torrential slide of research, both inside and outside of entryways to the clinical space, to help networks conquer this test by limiting its unfavorable effects. The immensity of those logical endeavors and the speed at which the information regarding this matter has been produced present noteworthy troubles for everyone to stay up on these turns of events [5]. Through the reaction to the COVID-19 emergency, numerous healthcare organizations have expanded their utilization of broadcast communications. From clinics to residency programs, from understanding cooperation at home to those in isolation, virtual correspondence gives a sheltered way to proceed with our obligations due to this pandemic [6–8]. Let us perceive how it spreads. The contamination that causes COVID-19 is imparted through droplets made when a tainted individual coughs, sneezes or breathes out. These droplets are too generous even to consider hanging observable all around and quickly fall on floors or surfaces [9]. An individual can be contaminated by taking in the disease if the inside closeness of someone who has COVID-19, or by reaching a polluted surface and a short time later individual eyes, nose, or mouth. COVID-19 affects various people in different habits. Most polluted people will make it delicate to coordinate symptoms. Primary side effects are fever, tiredness, and dry cough. A few people may encounter a throbbing pain, nasal clog, runny nose, sore throat, and diarrhea [10]. An ordinary course takes 5–6 days from being defiled with the contamination for reactions to show up. In any case, it can take up to 14 to 21 days. People with mild symptoms who are healthy should self-isolate in their homes. Search for clinical thought in case individuals have a fever, a cough, and breathing problems.

The world is in utmost danger owing to COVID-19 spread by an infection of the novel corona family. Most of the world nations have seen countless COVID-19 cases since December 2019 onwards, as shown in Fig. 1 and the deaths graph exhibited in Fig. 2.

Fig. 1.

Dissemination of COVID-19 cases overall globe (9th March, 2022) [68]

Fig. 2.

Dissemination of COVID-19 deaths overall globe (9th March, 2022) [68]

To combat this epidemic, researchers from various research groups are looking for computer-aided systems like the Internet of Things [113], Machine Learning (ML) or Deep Learning (DL) methodologies [21], Big Data [26], as well as Blockchain [21] that can help overcome the difficulties posed by COVID-19. These techniques can be employed to manage the transmission of the virus, diagnose it, or indeed develop and produce a vaccine or drug to remedy it.

There have been two coronavirus pandemics in history, which include severe acute respiratory syndrome (SARS-CoV) [56] as well as Middle Eastern Respiratory Syndrome (MERS) [25]. SARS-CoV is a respiratory viral infection that can be transmitted through the air. It was discovered in 2003. According to the WHO website [16], the virus had over 8,000 reported cases worldwide during its course, affecting over 26 countries. MERS is a respiratory viral infection with symptoms are similar to SARS-CoV.

Machine Learning as a subset of Artificial Intelligence, has demonstrated significant potential in a variety of industries, including retail [114], investing [115], medicine [42, 50], pharmaceutical drugs [116], data security [117], solar cells [118], non-volatile memories [119], modeling of electronic devices [120, 121], supercapacitors [122], agriculture [123–125] and many others [65]. Machine learning methods can be programmed to mimic human intelligence. In the healthcare field, for instance, various ML techniques can be trained and applied to clinical diagnosis [126]. To identify anomalies, ML frameworks have been vastly trained on a dataset of medical images such as Computed Tomography (CT) Scans, Magnetic Resonance Imaging (MRI), or X-rays [25]. Cancer [11], diabetes [12], fatty liver [22], as well as other diseases can all benefit from ML models. For an instance, Prostate cancer can be identified by utilizing ML techniques with a predictive performance of 97.13% [127].

In general, machine learning and deep learning methods have been extensively employed in the health sector, and they can also be employed to quantify patient data as well as diagnose COVID-19. The following are the main contributions of this survey:

• Deep learning strategy empowers fast identification of COVID-19 on chest CT tests.

• It is fundamental to approve that the deep learning technique is a vast-scope dataset.

• The innovation move happened in the public arena during the COVID-19 epidemic and its impact.

• Innovation-based systems in controlling the pandemic and helping society.

• Key innovations incorporate Machine Learning, Deep Learning, Artificial Intelligence correspondence, and surveillance.

• Impact of existing advances challenges, what’re more, future effects on society.

The rest of the paper is structured as follows. Section 2 discusses the related work on COVID-19. Section 3 investigates Data Science enabled by the classification and foreseeing of COVID-19 using radiography images. This section is split into three segments that go over preventing and diagnosing COVID-19 infection using ML and AI mechanisms; DL mechanisms; and various DL mechanisms. Lastly, in Sect. 4, we discuss, outline future work, and conclude.

Related Work

We choose to help the general public handle COVID-19 by exploring different executed innovations just as the potential advances that would assist in regulating the epidemic, introducing our examination as a composition. In request to compose the manuscript on this novel subject, one has to know what’s going around and requires monotonous studying of online data from checked sources and writing reviews to support the achievability of the methodology revealed in the web sources. The related work is done by utilizing sources, for example, Google researcher, Elsevier, PubMed, Springer, and IEEE. Firstly, the underlying screening is finished by examining the work’s title, and afterward, concurring with the areas in our examination, the recognized work is arranged. Neurological symptomatology, neurological hazard factors for poor visualization, pathophysiology for micro invasion, and activities taken by neurological or neurosurgical administrations to deal with the current COVID-19 emergency are checked on Long-term neurodegenerative impacts of COVID-19 presently can’t seem to be explained, however, hypothesized dependent on past involvement in other beta-coronaviruses [11]. Nervous system science and neurosurgical practice designs are being changed significantly and continually refreshed with distributions illustrating early encounters and proposals dependent on facilities furthermore, areas generally influenced. [12].

CT confidently plays a significant role in this pandemic circumstance for early distinguishing proof of COVID-19 pneumonia. Typical CT highlights incorporate fringe GGOs with multifocal conveyance and a dynamic development toward sorting out pneumonia designs. Centrilobular knobs, mucoid impactions, and one-sided segmental or lobar solidifications propose a bacterial starting point of pneumonia or superinfection. RT-PCR (real-time polymerase chain reaction) stays required for definite affirmation; however, its energy can be deferred, requiring the test to rehash the CT highlights is interesting [13, 14]. The most well-known cutaneous appearance of COVID-19 was seen as maculopapular exanthema (morbilliform), introduced in 36.1% (26/72) patients. The different cutaneous appearances included: a papulovesicular rash (34.7%, 25/72), urticaria (9.7%, 7/72), difficult acral red–purple abscesses (15.3%, 11/72) of patients, livedo reticularis injuries (2.8%, 2/72), and petechiae (1.4%, 1/72). Most of the lesions were restricted to the storage compartment (66.7%, 50/72). Nonetheless, 19.4% (14/72) of patients experienced cutaneous indications in the hands and feet. Skin injury improvement happened before the beginning of respiratory manifestations or COVID-19 determination in 12.5% (9/72) of the patients, and injuries were suddenly mended in all patients within ten days. The more significant part considers detailed no connection between COVID-19 seriousness and skin sores [15, 16]. Expanding the aerobic limit is suggested to improve safe and respiratory capacities which would assist in countering COVID-19 [17].

The remarkable rate at which logical distributions identified with the COVID-19 pandemic are developing, and the enormous cultural concerns identified with numerous perspectives identified with the effects of the pandemic, make incorporating analytical information more significant than any other [18]. A continuous topic of the COVID-19 pandemic is the requirement for boundless accessibility of precise, what’s more, effective analytic testing for the identification of COVID-19 and antiviral antibodies in contaminated people. Most tests for early identification of COVID-19 RNA depend on the opposite interpretation of polymerase chain response. Yet, isothermal nucleic corrosive intensification tests, including interpretation, intervened enhancement, and CRISPR-based procedures, are promising other options. Recognizable proof of people who have created antibodies to the COVID-19 contamination needs serological tests, counting enzyme-linked immunosorbent assay (ELISA), and parallel stream immunoassay [19].

Data Science Enabled by Classification and Foreseeing of COVID-19 Using Radiography Images

As individuals, we know how critical a job the advances play in our everyday lives except the viewpoint of innovation in helping the relief of contamination and in controlling circumstances like the COVID-19 epidemic is something we neglect to take note of. Hence, we authors conceptualized the parts of innovation used as different methodologies to assist in a pandemic situation [20]. For surviving the epidemic conditions, the commitment of various advancements can be recognized as straightforwardly affecting and impacting in a roundabout way. Straightforwardly affecting innovations are the ones that help in confining procedures to diminish contamination, aid medical services, and support the general public to work as one. In the general public, it is discovered that the following areas: Machine Learning, Deep Learning, Artificial Intelligence, and Blockchain [21] are the sound regions where advances are widely utilized and investigated and whose application has been effectively aiding the general public in an epidemic circumstances [22–24].

Artificial Intelligence is seen as an insurgency in this data period happening all around the globe, and it has enormous applications. The presentation of data innovation in well-being has progressed the field from numerous angles [25]. To begin with, we will characterize the specialized terms associated with the field of artificial intelligence.

Owning to the rapid spread of the Coronavirus, specialists are facing over-the-top complications in the conclusion of Coronavirus. Regardless of how the Reverse Transcription Polymerase Chain Reaction (RT-PCR) procedure is the fundamental methodology utilized in Coronavirus assurance [26], it encounters limits like low pandemic affectability, needs extra time, and is short in supply. Presently, an assessment of clinical pictures is maybe the most promising rising assessment locale in the clinical consideration region. In this manner, the examination of clinical pictures or instance, Chest X-ray, Computed Tomography, Ultrasound, and scanners can endure the obstacles of RT-PCR. As a high-level development,moreover playing a crucial viewpoint in the evasion of disease, the overall well-being crisis is likewise looking for the help of advanced innovation such as the utilization of Data Science (Machine Learning (ML), Artificial Intelligence (AI), and Deep Learning (DL) models) for efficacy prognosis of clinical pictures to stop the spread of the infection. The connection between AI, ML, and DL as exhibited in Fig. 3.

Fig. 3.

Elucidation of Data Science from the connection among AI, ML, and DL

Machine Learning and AI Mechanism

The focal capacity of AI is the derivation of insightful strategies without having any data on the fundamental segment that is unknown or inadequately described. These procedures are prepared for making complex examples from huge, disorderly, or complex data. The amalgamation of proactive strategies with a master system lessens the subjectivity issue. These techniques offer great assistance for the clinical prognosis. AI methodology can be used to find the plague designs as these estimations are used in the procedure and expecting clinical pictures [27]. Thus, in the current pandemic, a couple of examines have utilized various methodologies of machine learning hitherto Support Vector Machine, Logistic Regression, Random Forest, Decision Tree, etc., in the prescient and prognosis of COVID-19.

Computed Tomography (CT) can be an effective supplement for Coronavirus prognosis in the current pandemic setting. CT, without a doubt, assumes a significant part in the early recognition of Coronavirus pneumonia. Typical CT highlights incorporate fringe GGOs with multifocal circulation and a reformist development towards getting pneumonia designs sorted out [13]. Adjacent to clinical methodology and treatments, since AI guarantees another world view for medical care, a few distinctive computer-based intelligence apparatuses based upon ML techniques are utilized to examine the information and decision-making mechanisms. Nonetheless, unlike other medical care issues, to identify Coronavirus, AI-driven devices are required to have dynamic learning-based cross-populace train/test techniques that utilize multitudinal and multimodal information [28]. Investigate the Coronavirus general conclusion list with AI to improve the analysis exactness for a clinical reason. Used four kinds of simulated intelligence innovation to screen all patients, accomplish 18 files related to critical Coronavirus prognosis and the main trait are WBC, Eosinophil tally, Eosinophil rate, 2019 novel Covid RNA (2019n-CoV), and Amyloid-An in a lab, likewise coordinated with 2019 China infection determination clinical aide [29]. The most widely recognized appearances and examples of lung anomaly on CXR in Coronavirus to prepare the local clinical area in its endeavors to battle this pandemic [30].

With prescient examination abilities, the AI system applied genuine patient information to give fast clinical dynamic help. Coronavirus has introduced a squeezing required as a) clinicians are as yet creating clinical insight into this novel sickness, and b) asset limits in a flooding epidemic need troublesome asset assignment choices. The prescient models gained from patients’ verifiable information from these two emergency clinics accomplished 70% to 80% exactness in foreseeing serious cases [31]. The utilization of AI for the location of subpleural pulmonary lesions (SPLs) in Ultrasound (US) images, Yupeng Xu et al. [32] proposed a novel boundary-restored network (BRN) for computerized SPL division to keep away from issues related to manual SPL division. Dual-stage boundary restored network (BRN) beat existing division techniques (U-Net and a fully convolutional network (FCN)) for the division exactness boundaries including DSC (83.45 ± 16.60%), MCC (0.8330 ± 0.1626), Jaccard (0.7391 ± 0.1770), ASSD (5.68 ± 2.70 mm), and MSSD (15.61 ± 6.07 mm). Additionally beat the first BRN as far as the DSC by practically 5%.

Vinod et al. [33, 34] expounded a sound system that distinguishes Coronavirus tainted individuals among ordinary people using AI with CT scans and chest x-ray pictures. The picture determination device uses a decision tree classifier for discovering novel COVID contaminated individuals. The suggested method accomplishes an accuracy of 87% in Chest X-rays while it uncovers 82% exactness in the CT picture data set. Lesion dissemination and multi-projection lesions in the two lungs were available in many patients (80 cases; 72.7%). Lesions most frequently elaborate both the fringe and focal zones (62 cases; 56.4%). Concerning morphology, 56 patients (50.1%) showed patchy lesions that were incompletely melded into enormous regions [34]. The normal and abnormal CT highlights of Coronavirus pneumonia and talked about the fundamental differential analysis of Coronavirus pneumonia. The imaging discoveries in this viral pneumonia exhibited a wide range, and there are no pathognomonic imaging discoveries for Coronavirus pneumonia [35]. Chest CT is suggested as the primary line imaging test for recognizing Coronavirus pneumonia, which could permit early identification of the average chest signs and an ideal assessment of the illness's seriousness and beneficial impacts. Because of the high false-negative pace of nucleic acid analysis, it is significant to make the last determination precisely by alluding to both clinical appearances and imaging discoveries. Distinctive radiological innovations have their particular benefits and importance [36].

Similar investigation of machine learning and soft computing methods to foresee the Coronavirus epidemic as an option to susceptible infected recovered (SIR) and susceptible, exposed infectious removed (SEIR) techniques. It gives an underlying benchmarking to show the capability of AI for future exploration [37]. André Filipe de Moraes Batista et al. [38] suggested five machine learning techniques (neural networks, random forests, gradient boosting trees, logistic regression, as well as support vector machines) were prepared on a rare example of 70% of the patients, and execution was tried on new inconspicuous information (30%). The support vector machines calculation (AUC: 0.85; Affectability: 0.68; Explicitness: 0.85; Brier Score: 0.16).

Utilizing CT pictures, Akbari et al. [39] recommended dynamic form models to assess and decide pneumonia contamination brought about by the Covid illness (Coronavirus). A foundation of active contour models (ACM) containing form portrayal and object boundary depiction strategies were introduced. Authors accomplished execution with FRAGL-Dice: 96.4%, Jaccard: 93.2%, and LSACM-96.8% exactness. Bo Wang et al. [40] suggested an AI framework that consequently dissects CT pictures and gives the likelihood of contamination to identify Coronavirus pneumonia quickly. The proposed model achieved an affectability of 0.974 and an explicitness of 0.922 on the test dataset, which incorporated an assortment of aspiratory illnesses.

The broad computer reproductions show better proficiency and adaptability of this end-to-end learning mechanism on CT picture division with picture improvement contrasting with the best-in-class division draws near, particularly GraphCut, Medical Image Segmentation (MIS), and Watershed. The methods for factual measures got utilizing the exactness, affectability, F-measure, accuracy, MCC, Dice, Jacquard, and particularity are 0.98, 0.73, 0.71, 0.73, 0.71, 0.71, 0.57, 0.99 respectively [41]. Prognostic demonstrating of endurance was executed utilizing radiomic highlights and clinical information, independently or in the blend. Maximum relevance minimum redundancy (MRMR) and XGBoost were utilized for variable selection and classification [42]. The consolidated model (Lung + Lesion + Clinical) had the most noteworthy prognostic capacity with AUC = 0.95 ± 0.02, ACC = 0.88 ± 0.04, SEN = 0.88 ± 0.06 and SPE = 0.89 ± 0.07. The 95% CI for these boundaries were 0.95–0.96, 0.88–0.89, 0.87–0.90, and 0.87–0.90, respectively.

Andreia S. Gaudenico et al. [43] suggested that texture-based investigation into the extent of other aspiratory sicknesses has been utilized to monitor, screen, and give important data to a few sorts of findings. A complexity index (CI) in light of the amount of entropy esteems is utilized to characterize solid subjects and Coronavirus patients, showing an exactness of 89.6% and affectability of 96.1%, and explicitness of 76.9%. Three best-performing classical AI mechanisms during the preparation stage—1) fine Gaussian support vector machine (SVM), 2) fine k-closest neighbor (KNN), and 3) ensemble bagged model (EBM) trees were picked for additional assessment on an independent test informational collection [44]. Ensemble Bagging Model Trees (EBM) with the preferred radionics highlights is the most appropriate to recognize Coronavirus and other lung contaminations with a general affectability of 87.8% and explicitness of 97% (95.2% precision and 0.9228 area under the curve) and is strong across seriousness levels. AI-driven instruments are restricted to one information type, either CT scan or CXR, to identify Coronavirus positive cases. Coordinating different information types might give more data in identifying oddity designs because of Coronavirus [45]. The proposed model accomplished generally an accuracy of 96.28% (AUC = 0.9808 and false-negative rate = 0.0208). Vinod et al. [56, 57] suggested a deep covix-net model identify COVID-19 individuals with 96.8% multiple classification accuracies among COVID-19, normal, and Pneumonia chest X-ray image database. Moreover, it obtained a 97% accuracy among COVID-19 and normal CT scan images. The schematic workflow of ML, and DL-based techniques for the prognosis of COVID-19 as illustrated in Fig. 4.

Fig. 4.

Schematic workflow of ML and DL-based approaches for prognosis of COVID-19

Deep Learning Mechanism

Mizuho Nishio et al. [46] recommended the CADx framework use VGG16 as a pre-trained method and a mix of regular techniques and data expansion strategies. The proposed model accomplished performance with three-class precision and was assessed for a test set with 125 CXR pictures. The three-class accuracy of the computer-aided design framework was 83.6% between Coronavirus pneumonia, non-Coronavirus pneumonia, and the healthy. Affectability for Coronavirus pneumonia was over 90%. Xiangjun Wu et al. [47] proposed a multi-view combination technique utilizing a deep learning model to screen patients with Coronavirus using CT pictures with the most extreme lung districts in pivotal, coronal, and sagittal perspectives. The suggested model has accomplished AUC, precision, affectability, and explicitness of 0.819, 0.760, 0.811, and 0.615, respectively. For picture tissue grouping, two realized deep learning methods, SegNet and U-net recommended by Adnan Saood et al. [48]. SegNet is described as a scene division organization and U-net as a clinical division instrument. The proposed strategy outcomes exhibited the overall capacity of SegNet in grouping contaminated/non-tainted tissues contrasted with different strategies (with 0.95 mean exactness). In contrast, the U-net showed better outcomes as a multiclass segment (with 0.91 mean precision).

Federated learning mechanisms to foster a protection-saving AI model for Coronavirus clinical picture analysis with great speculation capacity on concealed worldwide datasets. Edward H. Lee et al. [49] proposed Deep COVID DeteCT (DCD), a deep learning convolutional neural network (CNN) that utilizes the whole chest CT volume to consequently anticipate Coronavirus (COVID +) from non-Coronavirus (COVID −) pneumonia and typical controls. The consideration of non-China destinations in preparing fundamentally improved characterization execution with an area under the curve (AUCs) and exactness above 0.8 on most test locales. ROC curves for lesion detection on the interior testing set and three outer validation sets. The consequences of six settings are examined, including individual model-1, individual model-2, individual model-3, their model outfit, their joint preparing model, and the combined learning model [50]. The consequences of AUC are given 95%. An open-source system, Covid CT Net, suggested by Tahereh Javaheri et al. [51], is made out of many deep learning frameworks that precisely separate Coronavirus from local area gained pneumonia (CAP) and other lung sicknesses. Covid CT Net builds the exactness of CT imaging recognition to 95% contrasted with radiologists (70%).

A deep learning system recognizes Coronavirus from clinical pictures as an assistant testing strategy to improve indicative affectability [52]. High scores for numerous factual files (F1 scores > 96.72% (0.9307, 0.9890) and explicitness > 99.33% (0.9792, 1.0000)). The theory is that utilizing deep learning (DL) to 3D CT pictures could assist in recognizing Coronavirus diseases. Utilizing information from 920 Coronavirus and 1073 non-Coronavirus pneumonia patients, Bohan Liu et al. [53] fostered a modified DenseNet-264 model, COVIDNet, to order CT pictures to one or the other class. Tested on an autonomous arrangement of 233 Coronavirus and 289 non-Coronavirus pneumonia patients, COVIDNet accomplished a precision score of 94.3% and an area under the curve of 0.98. A deep learning system that coordinates a convolutional neural network and a capsule network. DenseCapsNet, another deep learning structure, is shaped by the combination of a dense convolutional network (DenseNet) and the capsule neural network (CapsNet) [54], utilizing their benefits and decreasing the reliance on convolutional neural networks on a large amount of data. The strategy can acquire an exactness of 90.7% and an F1 score of 90.9%, and the affectability for distinguishing Coronavirus can reach 96%.

Xinggang Wang et al. [55] fostered a deep learning-based model for automatic Coronavirus prognosis on chest CT is useful to counter the episode of SARS-CoV-2. A weakly-supervised deep learning system was created utilizing 3D CT volumes for Coronavirus characterization and lesion localization. When using a likelihood edge of 0.5 to order Coronavirus positive and Coronavirus negative, the calculation acquired a precision of 0.901, a positive prescient value of 0.840, and an extremely high adverse prescient value of 0.982.

The coronavirus detection neural network (COVNet) [57] was created to extricate visual highlights from volumetric chest CT checks to recognize Coronavirus. CT scans of community-acquired pneumonia (CAP) and other non-pneumonia anomalies were incorporated to test the vigor of the model. The per-scan affectability and explicitness for distinguishing Coronavirus in the autonomous test set was 90% (95% confidence interval [CI]: 83%, 94%; 114 of 127 outputs) and 96% (95% CI: 93%, 98%; 294 of 307 scans), respectively, with an AUC of 0.96. A visual understanding mechanism for clarifying the deep learning techniques and applying the tool in Coronavirus prognosis. Yuchai Wan et al. [58] plan a detailed understanding of the deep model from alternate points of view, including preparation patterns, analytic execution, learned highlights, extractors, hidden layers, and support locales for a demonstrative decision forth. This model accomplishes the symptomatic result of 94.75%, 93.22%, 96.69%, 97.27%, and 91.88% in the measures of precision, affectability, explicitness, positive prescient value, and prescient negative value, which are 8.30%, 4.32%, 13.33%, 10.25%, and 6.19% higher than that of the correlated traditional strategies. The pictured highlights in 2-D and 3-D spaces give the motivations for the prevalence of the model.

One significant obstacle in regulating the spread of this infection is the shortcoming and deficiency of clinical trials. There have been expanding endeavors on growing deep learning techniques to analyze Coronavirus dependent on CT examines. Xuehai He et al. [59] proposed a Self-Trans methodology contrasted and a few best-in-class baselines. This methodology accomplishes an F1 of 0.85 and an AUC of 0.94 in diagnosing Coronavirus from CT scans. A viable system that can give a promising arrangement by transferring knowledge from nonexclusive object recognition tasks to explicit space assignments. Asmaa Abbas et al. [60] suggested a deep CNN, called Decompose, Transfer, and Compose (DeTraC), to characterize Coronavirus chest X-ray pictures. The proposed strategy accomplished a precision of 93.1% (with an affectability of 100%) and was performed by DeTraC in recognition of Coronavirus X-ray pictures from typical and extremely intense respiratory condition cases. A deep learning model, Coronavirus identification neural network (COVNet), was created to remove visual highlights from volumetric chest CT tests to recognize Coronavirus. Community-acquired pneumonia (CAP) and other non-pneumonia CT tests were incorporated to test the strength of the technique. A deep learning strategy could identify Coronavirus on chest CT tests (AUC of 0.96). A deep learning strategy to recognize community-acquired pneumonia on chest CT tests (AUC of 0.95) [61].

Quickly created AI-based mechanized CT picture examination apparatuses can accomplish high exactness in recognizing Covid positive patients just as an evaluation of illness burden. Accomplished classification outcomes for Covid versus Non-Covid cases per thoracic CT investigations of 0.996 AUC (95%CI: 0.989–1.00) on Chinese control and contaminated patients. Conceivable working point: 98.2% affectability, 92.2% particularity [62]. A deep convolutional neural network (CNN) based design, so-called as CovXNet, is recommended [63] that uses depth-wise convolution with fluctuating expansion rates for productively removing expanded features from chest X-rays. Execution with the exactness of 97.4% for Coronavirus/Typical, 96.9% for Coronavirus/Viral pneumonia, 94.7% for Coronavirus/Bacterial pneumonia, and 90.2% for multiclass Coronavirus/ordinary/Viral/Bacterial pneumonia. Contamination areas were portioned out from the pneumonic CT picture set utilizing a 3D deep learning technique [64]. The simulation outcome of the benchmark dataset showed that the general precision rate was 86.7% as far as all the CT cases are taken together.

AI framework for quick Coronavirus discovery and performed broad factual investigation of CTs of Coronavirus dependent on the computer-based intelligence framework. A deep convolutional neural network-based framework can accomplish an RoC of 97.17%, affectability of 90.19%, and particularity of 95.76% for Coronavirus on interior test companion of 3,203 scans and AUC of 97.77% [65]. Jannis et al. [66] utilized an interpretability strategy for the spatiotemporal confinement of pneumonic biomarkers, which are considered helpful for human-insider-savvy situations in a dazed report with clinical specialists. An edge-based model that effectively recognizes Coronavirus LUS recordings from sound and bacterial pneumonia information with an affectability of 0.90 ± 0.08 and an explicitness of 0.96 ± 0.04. A convolutional neural network (CNN) was prepared on LUS pictures with B lines of various aetiologies [67]. As approved utilizing a 10% information holdback set, CNN analytic execution was contrasted and studied LUS-equipped doctors. The prepared CNN execution on the free dataset showed a capacity to separate between Coronavirus (AUC 1.0), NCOVID (AUC 0.934), and HPE (AUC 1.0) pathologies. This was essentially better compared to doctor capacity (AUCs of 0.697, 0.704, 0.967 for the Coronavirus, NCOVID, and HPE classes, respectively), p < 0.01.

DL-based elucidations for the helped prognosis of lung infections, the use of DL strategies to investigate lung ultrasonography (LUS) pictures [69]. Ensemble of U-Net, U-Net +  + , and Deeplab v3 + with Online-expansion achieved an exactness of 96% and Dice of 0.75. Handled explicitly for deep learning techniques and is expected to fill in as a beginning stage for an open-access drive. A deep convolutional neural network (POCOVID-Net) on this 3-class dataset accomplishes an exactness of 89% and a video precision of 92% by a more significant part vote. For distinguishing Coronavirus specifically, the technique attained an affectability of 0.96, explicitness of 0.79, and F1-score of 0.92 in a 5-overlap cross-approval [70]. The programmed division of lung and lung lesions in chest CT outputs of affirmed or suspected Coronavirus patients. Ensembling various strategies can support the general test set execution for lung division, binary lesion division, and multiclass lesion division, bringing about mean Dice scores of 0.982, 0.724, and 0.469, respectively [71]. The coronavirus aspiratory contamination division profundity network is alluded to as the Attention Gate Dense Network Improved Dilation Convolution-UNET (ADID-UNET) [72]. The test outcome exhibits that the ADID-UNET technique can precisely portion Coronavirus lung-contaminated regions, with execution measures more prominent than 80% for measurements like Exactness, Particularity, and Dice Coefficient (DC).

The most recent data mining and AI methods like Convolutional Neural Networks (CNN) can be enforced alongside X-ray and CT scan pictures of the lungs for the precise and quick identification of the illness, helping alleviate the shortage of testing kits. A characterization exactness of 99.1% for 2 class grouping, 94.2% for 3 class arrangement, and 91.2% for 4 class classification was originated [73]. The utilization of a profound network requires an adequately huge preparing set, which isn’t accessible practically. On the other hand, the utilization of a shallow CNN may not give prevalent outcomes since it cannot rich component extraction because it lacks enough convolutional layers. The precision and running time perspective utilizing restricted training samples. Over 76% and 94% overall classification precision are acquired by the recommended strategy [74] in CT scan and X-ray pictures datasets, respectively. An area classifier executes contaminated region transformation from source space to target space in an Adversarial Learning way and learns area invariant region proposal network (RPN) in the Faster R-CNN model. Wei Li et al. [75] recommended the NIA mechanism (Network-in-Network, Case Standardization, and Adversarial Learning) and led broad trials on two Coronavirus datasets to approve the methodology. NIA-Network with an arbitrary introduction on the equivalent dataset, and the outcomes are 94% for affectability, 90.2% for explicitness, and 92.1% for precision, respectively. Those outcomes are under 94.2% for affectability, 99.5% for explicitness, and 96.82% for precision in which a pre-trained replica introduces ResNet-50.

A deep learning-based improved Depiction Group method for productive Coronavirus chest X-ray arrangement [76], a multiclass miniature of 97% explicitness, 95% f1-score, and 95% characterization precision. The test study incorporates two arrangements of benchmark capacities, specifically standard capacities and CEC2013 capacities, having a place with various classes, for example, unimodal, multimodal, and unconstrained improvement capacities [77]. The recommended technique attained an exactness of 64% and an F1-Score of 62%. Corona-Nidaan, a lightweight deep convolutional neural network, is proposed to recognize Coronavirus, Pneumonia, and Normal cases from chest X-ray picture investigation, with no human mediation. Mainak Chakraborty et al. [78] presented an introductory minority class oversampling strategy for managing imbalanced dataset issues. The model accomplished 95% exactness for a three-class arrangement with 94% accuracy and review for Coronavirus cases. While examining the attainment of different pre-trained models, it is additionally discovered that VGG19 exceeds other pre-trained CNN models by accomplishing 93% exactness with 87% sensitivity and 93% accuracy for Coronavirus contamination identification.

Li et al. [79] proposed a stacked autoencoder detector model to significantly improve the exhibition of the discovery models such as precision rate and recall rate. The model achieves the average accuracy, precision, recall, and F1-score rate of 94.7%, 96.54%, 94.1%, and 94.8%, respectively. The following Table 1 depicts the performance analysis of some deep learning approaches that may enable the researchers to select an appropriate deep learning approach and architecture for resolving conflicts in the COVID-19 pandemic as well as further scope to improve the overall performance of different approaches.

Table 1.

Attainment examination of different deep learning mechanisms in the prognosis of COVID-19

Author	Month & Year	Data	Method	Accuracy	F1- Score	Specificity	Sensitivity
Kuchana et.al [80]	Nov, 2020	CT Scan	2D deep learning architecture with U-Net	Not Reported	97.30%	Not Reported	Not Reported
Mei et.al [81]	May, 2020	CT Scan	Convolutional neural network (CNN), Multi-Layer Perceptron (MLP), and Joint Model	CNN model: 79.6%, MLP: 74.2% and Joint model: 83.5%	Not Reported	CNN model: 75.9%, MLP: 68.3% and Joint model: 82.8%	CNN model: 83.6%, MLP: 80.6% and Joint model: 84.3%
Wang et al. [82]	Nov, 2020	X-Ray	COVID-Net, a deep CNN design model	COVID-Net: 93.3% comparing this model with VGG-19: 83% and ResNet-50: 90.6%	Not Reported	Not Reported	COVID-Net: 91% comparing this model with VGG-19: 58.7% and ResNet-50: 83%
Che Azemin et.al [83]	Aug, 2020	X-Ray	ResNet-101 CNN architecture was used	71.90%	Not Reported	71.80%	77.30%
Harmon et.al [84]	Aug, 2020	CT Scan	3D Classification model (resampled the cropped lung region of CT to a fixed size (192 × 192x64 voxels) for input to algorithm and Hybrid 3D model (192 × 192x32)	3D: 89.6% and Hybrid 3D: 89.5%	Not Reported	3D: 91.6% and Hybrid 3D: 95.1%	3D: 84.5% and Hybrid 3D: 75.1%
Hassantabar et.al [85]	July, 2020	X-Ray	Deep neural network and Convolutional Neural Network	CNN architecture: 93.2% and DNN: 83.4%	CNN architecture: 97.9% and DNN: 83.4%	CNN architecture: 99.7% and DNN: 82.3%	CNN architecture: 96% and DNN: 86%
Panwar et.al [86]	May, 2020	X-Ray	Deep learning neural network-based method nCoVnet	88%	Not Reported	78.57%	97.62%
Ghulam Gilanie et al. [87]	Feb, 2021	X-Ray and CT Scan	CNN	96.60%	Not Reported	95.65%	96.24%
Tao Yan et al. [88]	July, 2020	CT Scan	Multi-scale convolutional neural network	96.30%	Not Reported	95.60%	89.10%
Rodolfo M. Pereira et al. [89]	May, 2020	X-Ray	i) a multi-class classification; ii) hierarchical classification, since pneumonia can be structured as a hierarchy	Not Reported	Multi class: 65% and Hierarchical classification: 89%	Not Reported	Not Reported
Song Ying et al. [90]	March, 2021	CT Scan	DRE-Net	93%	93%	93%	93%

Various Deep Learning Mechanisms

An interpretable computer-based intelligence structure evaluated by master radiologists on the premise of how well the consideration maps center around the indicatively functional picture areas. The pre-trained Inception-v3 technique accomplished an AUC execution of 100% in pneumonia versus Coronavirus, 96% in normal versus pneumonia versus Coronavirus, and 93% for quaternary classification. It merits referencing that the exchange learning strategy gave a solid gauge to the analyzed lesion, notwithstanding the information increase moderating the restricted arrangement of Coronavirus X-rays [91]. Patients with Coronavirus indications have been identified utilizing eight particular profound learning strategies, which are VGG16, InceptionResNetV2, ResNet50, DenseNet201, VGG19, MobilenetV2, NasNetMobile, and ResNet15V2, using two datasets: one dataset incorporates 400 CT examine and another 400 chest X-ray pictures. NasNetMobile performed well compared to all other techniques by accomplishing a precision of 82.94% in CT scans and 93.94% in chest X-ray datasets [92].

An advisable Convolutional Neural Network (CNN) illustration begins by correlating a few famous CNN techniques. Horry et al. [93] optimized the preferred VGG19 technique for the picture modalities to exhibit how the methods can be utilized for exceptional scares and testing Coronavirus datasets. The prescribed model accomplished up to 86% accuracy for X-rays, 100% for Ultrasound, and 84% for CT scans. Computer-aided design conspires first applies two picture preprocessing steps to eliminate most diaphragm regions, measure the first picture utilizing a histogram equalization innovation, and a reciprocal low-pass channel. Then, the original image and two separated pictures are used to frame a pseudo-color picture. The CNN-based computer-aided design conspires to yield a general exactness of 94.5% (2404/2544) with a 95% certainty period in grouping three classes. The computer-aided method additionally yields 98.4% affectability (124/126) and 98.0% explicitness (2371/2418) in arranging cases with and without Coronavirus contamination [94]. Tulin Ozturk et al. [95] suggested a DarkNet technique as a classifier for you only look once (YOLO) continuous article recognition framework. The authors carried out 17 convolutional layers and presented various shifts on each layer. The model created a characterization precision of 98.08% for binary class and 87.02% for multiclass cases. Table 2 represents the analysis of performance metrics of various transfer learning algorithms in the diagnosis of COVID-19 infection.

Table 2.

Diagnosis of COVID-19 enabled by various transfer learning algorithms

Author	Month & Year	Data	Method	Accuracy	F1- Score	Specificity	Sensitivity
Jain et.al [96]	Oct, 2020	X-Ray	InceptionV3, Xception and ResNeXt	InceptionV3 model: 95%, ResNeXt: 93% and Xception: 96%	InceptionV3 model: 95%, ResNeXt: 96% and Xception: 86%	Not Reported	InceptionV3 model: 92%, ResNeXt: 96% and Xception: 78%
Prabira Kumar Sethy et al. [97]	March, 2020	X-Ray	Resnet50 + SVM	95.38%	91.41%	93.47%	97.29%
Wang.s et al. [98]	Jan, 2021	CT Scan	Modified Inception transfer learning followed by internal and external validation	Internal validation: 89.5% and external validation: 79.3%	Internal validation: 77% and external validation: 63%	Internal validation: 88% and external validation: 83%	Internal validation: 87% and external validation: 67%
Iason Katsamenis et al. [99]	Dec, 2020	X-Ray	Transfer Learning	Model-1:92% Model-2:95% Model-3:96%	Model-1:92% Model-2:95% Model-3:97%	Not Reported	Model-1:92% Model-2:95% Model-3:97%
Varalakshmi Perumal et al. [100]	Aug, 2020	X-Ray and CT Scan	Transfer Learning and Haralick features	93%	Not Reported	Not Reported	90%
Adi Alhudhaif et al. [101]	May, 2021	X-Ray	DenseNet-201	94.96%	92.11%	Not Reported	94.59%

Ten notable convolutional neural networks were utilized to recognize contamination of Coronavirus from non-Coronavirus gatherings: AlexNet, VGG-16, VGG-19, SqueezeNet, GoogleNet, MobileNet-V2, ResNet-18, ResNet-50, ResNet-101 furthermore Xception [102]. Amid all mechanisms, the finest exhibition was accomplished by ResNet-101 and Xception. ResNet-101 could recognize Coronavirus from non-Coronavirus cases with an AUC of 0.994 (affectability, 100%; explicitness, 99.02%; precision, 99.51%). Xception accomplished an AUC of 0.994 (affectability, 98.04%; explicitness, 100%; precision, 99.02%). Nonetheless, the presentation of the radiologist was moderate with an AUC of 0.873 (affectability, 89.21%; particularity, 83.33%; precision, 86.27%). CoroNet, a Profound Convolutional Neural Network model to consequently identify Coronavirus contamination from chest X-ray pictures. The recommended technique depends on the Xception mechanism pre-trained on the ImageNet dataset and trained end to end on a dataset arranged by gathering Coronavirus and other chest pneumonia X-ray pictures from two distinctive publicly accessible data sets. CoroNet has been trained and tested on the pre-arranged dataset, and the trial outcomes exhibited that the suggested technique [103] attained a general exactness of 89.6%, accuracy and review rate for Coronavirus cases are 93% and 98.2% for 4-class cases (Coronavirus vs. Pneumonia bacterial vs. pneumonia viral vs. normal).

With transfer learning, identifying various anomalies in small clinical picture datasets is a reachable objective, frequently yielding noteworthy results. The datasets utilized in this test are two. The outcomes recommend that deep learning with X-ray imaging may separate significant biomarkers identified with the Coronavirus sickness, while the best precision, affectability, and specificity attained are 96.78%, 98.66%, and 96.46%, respectively [104]. The COVIDX-Net incorporates seven unique designs of profound convolutional neural network models, for example, the adjusted Visual Geometry Group Network (VGG19) and the second form of Google MobileNet. The VGG19 and Dense Convolutional Network (DenseNet) models [105] exhibited a decent and comparable exhibition of mechanized Coronavirus arrangement with f1-scores of 0.89 and 0.91 for normal and Coronavirus, respectively. Mohamed Loey et al. [106] suggested three deep transfer models are chosen for examination. The techniques are Alexnet, Googlenet, and Restnet18. Those techniques are selected for analysis through this exploration as it encloses a few layers on their designs. This will bring about lessening the intricacy, the consumed memory, and the execution time. The Googlenet is the primary deep transfer model as it accomplishes 80.6% in testing exactness. The Alexnet is chosen to be the fundamental deep transfer model in the following situation as it achieves 85.2% in testing exactness.

Hu et al. [107] recommended a multimodal channel and receptive feld consideration network joined with ResNeXt (MCRFNet) to arrange ultrasound images. The network can naturally combine shallow features and decide the significance of various channels and particular felds. Utilizing multicentre and multimodal ultrasound information from 104 patients, the indicative model accomplished 94.39% exactness, 82.28% accuracy, 76.27% affectability, and 96.44% explicitness. A multi-layer fusion functionality of each block is preferred by Ghulam Muhammad et al. [108] to improve the effectiveness of Coronavirus screening. The recommended fusion technique has 92.5% exactness, 91.8% precision, and 93.2% recovery utilizing the information assortment. These metric effectiveness levels are significantly higher than those used in any of the cutting-edge CNN adaptations. A profound learning-based mechanism for programmed evaluation of lung parenchymal irregularities in chest CT of Coronavirus patients and to relate quantitative outcomes with clinical and lab boundaries [109]. In the cohort with 93% of patients giving trademark Coronavirus chest CT discoveries, the mechanism gave quick results (middle time for programmed assessment 120 s) with just 7% of cases requiring manual amendment. Another deep learning calculation for Coronavirus's mechanized conclusion involves a couple of tests for training. The created model utilized a self-administered methodology dependent on unaided learning. The way that it can beat ResNet-50 is astounding [110].

The contrastive performs multiple tasks convolutional neural network (CMT-CNN), which is made out of two assignments. The fundamental assignment is to analyze Coronavirus from other pneumonia and healthy control. The method undertaking is to empower nearby accumulation; however, a contrastive misfortune: first, each picture is changed by a progression of increases (Poisson noise, rotation, and so forth). Contrastive learning (as a module) brings significant precision improvement for profound learning techniques on both CT (5.49%-6.45%) and X-ray (0.96%-2.42%) without requiring different explanations [111]. Zhou Tao et al. [112] prescribed transfer learning to install model boundaries and pre-train three deep convolutional neural network models: AlexNet, GoogleNet, and ResNet. The model accomplished an affectability, particularity, F_Score, and Matthews connection coefficient (MCC) are 98.56%, 99.4%, 97.87%, and 96.81%, respectively.

Conclusion

In this paper, we have started reviewing the AI, ML, and DL arrangements in the fight against the COVID-19 pandemic. Initially, we have given a presentation on the COVID-19 infection, the essentials and inspirations of artificial intelligence, and essential information for discovering quick and compelling methodologies that can adequately battle the COVID-19 epidemic, and the review highlights have also been presented. At that point, we reviewed the utilization of artificial intelligence for recognition and analysis, foreseeing the outbreak, biomedicine, and pharmacotherapy. The utilization of extensive information for the COVID-19 outbreak has been likewise introduced, counting episode forecasts, the infection spread, and diagnosis with the help of a data science platform.

Several obstacles must be resolved in the future to fully utilize the advantages of AI, ML, and Transfer Learning techniques. A lack of reliable massive data, incorrect data, noisy data, a lack of the intersection of AI and medical fields, and data privacy are some of the major challenges. Furthermore, in the near future, high-quality medical datasets with large numbers of samples will be critical. Furthermore, the datasets should include COVID-19 patients at various stages, with the presence of borderline patients being important in determining the effectiveness of a classifier and also identifying the asperity levels with higher efficacy. The consistency of COVID-19 prognosis can be enhanced by using numerous imaging approaches for each suspected patient, such as Magnetic Resonance (MRI), X-ray, Ultrasound, as well as CT scan.

Declarations

Conflict of interest

None.