A Systematic Review on Deep Structured Learning for COVID-19 Screening Using Chest CT from 2020 to 2022

The emergence of the COVID-19 pandemic in Wuhan in 2019 led to the discovery of a novel coronavirus. The World Health Organization (WHO) designated it as a global pandemic on 11 March 2020 due to its rapid and widespread transmission. Its impact has had profound implications, particularly in the realm of public health. Extensive scientific endeavors have been directed towards devising effective treatment strategies and vaccines. Within the healthcare and medical imaging domain, the application of artificial intelligence (AI) has brought significant advantages. This study delves into peer-reviewed research articles spanning the years 2020 to 2022, focusing on AI-driven methodologies for the analysis and screening of COVID-19 through chest CT scan data. We assess the efficacy of deep learning algorithms in facilitating decision making processes. Our exploration encompasses various facets, including data collection, systematic contributions, emerging techniques, and encountered challenges. However, the comparison of outcomes between 2020 and 2022 proves intricate due to shifts in dataset magnitudes over time. The initiatives aimed at developing AI-powered tools for the detection, localization, and segmentation of COVID-19 cases are primarily centered on educational and training contexts. We deliberate on their merits and constraints, particularly in the context of necessitating cross-population train/test models. Our analysis encompassed a review of 231 research publications, bolstered by a meta-analysis employing search keywords (COVID-19 OR Coronavirus) AND chest CT AND (deep learning OR artificial intelligence OR medical imaging) on both the PubMed Central Repository and Web of Science platforms.


Introduction
As a result of severe acute respiratory syndrome coronavirus 2, the novel coronavirus (nCov), or simply COVID-19, emerged from Wuhan province, China (SARS-CoV-2) [1].At the beginning of March 2020, the World Health Organization (WHO) classified it as a pandemic, and since then, the extent of the threat has been documented through confirmed cases and fatalities [2,3].Following previous work [4], the spread of the COVID-19 virus can be expressed using the following transmission model, The ω allows the rate of COVID-19 increase to be less than proportionate.This is only possible if ω < 1.The average serial interval is predicted to be around 4.5 days, given that interaction with infected people has remained constant throughout the interval [5,6].Cough, headache, fever, muscle aches, shortness of breath, and dizziness are the most common symptoms of COVID-19 [7,8].Symptoms may or may not be evident in some cases.The virus quickly weakens the subject's immune system, resulting in death [9].The WHO recommends vailing inclination among authors toward CT scans.CT scans offer heightened sensitivity and visualization capabilities, albeit with the tradeoff of increased radiation exposure.The incorporation of deep learning (DL) algorithms has significantly streamlined clinical assessment and expert interpretation, rendering computer-aided diagnosis (CADx) models pivotal as supportive diagnostic tools in COVID-19 detection.Consequently, CADx imaging tools have garnered trust and now play a crucial role in COVID-19 screening.Within this study, our primary focus centered on CT scans, and the noteworthy contributions can be succinctly outlined as follows: • Providing an account of accessible CT datasets and their utilization in deep learning (DL) for the classification of COVID-19; • Conducting a performance evaluation that contrasts existing DL models through dataset utilization and methodological approaches; • Implementing transfer learning (TF) and data augmentation (DA) techniques in the development of DL models; • Proposing prospective directives for DL investigations within this particularly sensitive domain.The remainder of the paper is organized as follows: Section 2 provides an overview of significance of the COVID-19 virus.The research scope and criteria for selecting articles Throughout the pandemic, numerous researchers have dedicated their endeavors towards the classification and detection of COVID-19 using CT scans and X-rays.While both modalities possess their respective advantages and disadvantages, there has been a prevailing inclination among authors toward CT scans.CT scans offer heightened sensitivity and visualization capabilities, albeit with the tradeoff of increased radiation exposure.The incorporation of deep learning (DL) algorithms has significantly streamlined clinical assessment and expert interpretation, rendering computer-aided diagnosis (CADx) models pivotal as supportive diagnostic tools in COVID-19 detection.Consequently, CADx imaging tools have garnered trust and now play a crucial role in COVID-19 screening.Within this study, our primary focus centered on CT scans, and the noteworthy contributions can be succinctly outlined as follows: • Providing an account of accessible CT datasets and their utilization in deep learning (DL) for the classification of COVID-19;

•
Conducting a performance evaluation that contrasts existing DL models through dataset utilization and methodological approaches; • Implementing transfer learning (TF) and data augmentation (DA) techniques in the development of DL models; • Proposing prospective directives for DL investigations within this particularly sensitive domain.
The remainder of the paper is organized as follows: Section 2 provides an overview of significance of the COVID-19 virus.The research scope and criteria for selecting articles are discussed in Section 3. Section 4 focuses on AI-guided medical imaging and presents a summary of COVID-19 screening investigations.This includes discussions on data collections and their sources (Section 4.1); the most used CT imaging tools (Section 4.2); the methodological contributions of DL-based methods/models in the years 2020, 2021, and 2022 (Section 4.3); performance comparisons (Section 4.4); dataset sizes (Section 4.5); transfer learning (Section 4.6); and data augmentation (Section 4.7).Finally, in Section 5, we conclude the paper and provide future guidelines.

COVID-19: Background and Its Relevance
Schalk and Hawn [21] identified an ostensibly new respiratory illness in chicks in 1931 (between 2 days and 3 weeks old).We refer to a few papers, such as Frabricant (1998) [22] and Cook et al., for more specific progress on infectious bronchitis investigations (2021) [23].SARS-CoV [24] was discovered in China between 2002 and 2003.Approximately 8000 people were infected during these years, with a 9.5% fatality rate.Bats or civet cats were suspected of being the cause of the disease [25].MERS-CoV-a version of the coronavirus-was discovered in 2012 in Saudi Arabia [26].In 2019, we had approximately 2500 MERS-CoV infections, with a 30% death rate [27].The transmission agents were thought to be camels this time (dromedaries) [28].SARS-CoV-2 is one of the oldest known viruses, infecting humans through a common cold.The virus is transmitted via inhalation or ingestion of droplets produced by coughing and sneezing.The viral structure comprises roughly 30,000 nucleotides and contains four structural proteins: spike, membrane, envelope, nucleocapsid [29][30][31], and various nonstructural proteins.N-protein (viral positive strand RNA) is also found in the protein shell or capsid.This strand acts as a parasite in human cells that proceeds to replication and transcription.Lung screening for nCoV looks like influenza-associated pneumonia in terms of analysis [32,33].

Study Scope and Selection Criteria
Before commencing our review, let us adhere to a systematic workflow delineating multiple stages, including identification, screening, eligibility, and inclusion criteria, as illustrated in Figure 2. To identify relevant studies, we employed search keywords (COVID-19 OR Coronavirus) AND chest CT AND (deep learning OR artificial intelligence OR medical imaging) on both the PubMed Central Repository and Web of Science platforms.Following this, duplicate entries were eliminated.Specifically, we focused on experiment-based research publications utilizing deep learning (DL) models/algorithms, limited to the year 2020 for publication.To ensure rigor, we excluded preprint articles from arXiv, medRxiv, and TechRxiv due to their non-peer-reviewed status.Our assessment encompassed diverse aspects, such as dataset characteristics (size and source), technical intricacies (DL models), and corresponding performance metrics, enhancing the potential for comprehensive metaanalysis.Our primary objective is not solely to delineate performance rankings among research articles, but rather to gauge the progress achieved since the onset of the pandemic.Furthermore, our evaluation extends to critical considerations like dataset scale, data augmentation techniques, and the applicability of transfer learning methodologies.

AI for Medical Imaging for COVID-19
By extracting distinctive features, AI-guided technologies can streamline complex data representations, making decision making achievable.They lead to a wide range of applications, including drug discovery [35], innovative healthcare [36], biomedicine [37], and medical image analysis [38].Deep learning algorithms are prevalent in all cases [39,40].However, we still have challenges in developing a clinical screening tool that considers various variables.Recent studies have discussed the value of artificial intelligence (AI) in the prognostication and diagnosis of medical images [41,42].According to the WHO, COVID-19 is a global public health emergency and the most significant test we currently face [43][44][45].Research on supervised algorithms for COVID-19 identification (classification) and segmentation has been the primary focus since the first quarter of 2020 [46][47][48].Few promising investigations concentrate on dual-sampling attention networks [49].Interestingly, approaches that fall under the purview of unsupervised learning outperform supervised ones, even though most rely on supervised techniques [50][51][52].In what follows, we first provide CT scan datasets and their respective sources.The research articles that used CT scans are then reviewed for their methodological contributions.Although we acknowledge thousands of research articles published in the year 2020 and 2021, our study is limited only to experiment-based (with DL models), peer-reviewed articles other than preprints: medRxiv, TechRiv, and arxiv by using exact search keywords in PubMed Central Repository and Web of Science (Section 3).

Dataset and Availability
Plenty of CT scan-based datasets are available in the literature to identify COVID-19.Almost all the previously stated peer-reviewed studies employed various datasets in their setup and system designs.Before training/validating their systems, the general concept is to acquire datasets (private or public access) from internet sources or prepare with their method.Therefore, understanding and determining the best approaches from the available reports is physically challenging.Data unavailability is a widespread issue for computational scientists as many datasets are required for their machine/deep learning model.

AI for Medical Imaging for COVID-19
By extracting distinctive features, AI-guided technologies can streamline complex data representations, making decision making achievable.They lead to a wide range of applications, including drug discovery [35], innovative healthcare [36], biomedicine [37], and medical image analysis [38].Deep learning algorithms are prevalent in all cases [39,40].However, we still have challenges in developing a clinical screening tool that considers various variables.Recent studies have discussed the value of artificial intelligence (AI) in the prognostication and diagnosis of medical images [41,42].According to the WHO, COVID-19 is a global public health emergency and the most significant test we currently face [43][44][45].Research on supervised algorithms for COVID-19 identification (classification) and segmentation has been the primary focus since the first quarter of 2020 [46][47][48].Few promising investigations concentrate on dual-sampling attention networks [49].Interestingly, approaches that fall under the purview of unsupervised learning outperform supervised ones, even though most rely on supervised techniques [50][51][52].In what follows, we first provide CT scan datasets and their respective sources.The research articles that used CT scans are then reviewed for their methodological contributions.Although we acknowledge thousands of research articles published in the year 2020 and 2021, our study is limited only to experiment-based (with DL models), peer-reviewed articles other than preprints: medRxiv, TechRiv, and arxiv by using exact search keywords in PubMed Central Repository and Web of Science (Section 3).

Dataset and Availability
Plenty of CT scan-based datasets are available in the literature to identify COVID-19.Almost all the previously stated peer-reviewed studies employed various datasets in their setup and system designs.Before training/validating their systems, the general concept is to acquire datasets (private or public access) from internet sources or prepare with their method.Therefore, understanding and determining the best approaches from the available reports is physically challenging.Data unavailability is a widespread issue for computational scientists as many datasets are required for their machine/deep learning

CT Imaging Tools
DenseNet is a new convolution network architecture proposed by G. Huang, Z. Liu, and K. Weinberger in the paper "Densely Connected Convolutional Networks" [53].They reported that the proposed architecture achieved high performance with four benchmark tasks: CIFAR-10, CIFAR-100, SVHN, and ImageNet.
The VGG network is a convolution neural network model implemented by K. Simonyan and A. Zisserman in the paper "very deep convolutional networks for large-scale image recognition" [54].This architecture was beneficial for classification accuracy, and the highest performance accuracy was archived with 14 million images belonging to 1000 classes.
InceptionNet is a deep convolution neural network architecture proposed by C. Szegedy and others in the paper "Going deeper with convolutions" [55].This proposed approach was experimentally verified on the ILSVRC 2014 classification and detection challenges, significantly outperforming the current state of the art.
A.G.Howard proposed MobileNets architecture in the paper "MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications" [56].This architecture based on depth-wise separable convolutions, and it provided an effective result for a wide variety of tasks.
ResNet is a residual learning framework proposed by K.He and others in the paper "Deep Residual Learning for Image Recognition" [57].This architecture is easy to train as compared to the deep neural network.The result came from this architecture winning 1st place on the ILSVRC 2015 classification task.
Y. Le Cel proposed a CNN architecture in the paper "Backpropagation Applied to Handwritten Zip Code Recognition" [58].This architecture is flexible to network design and is used in image classification.
UNet: O. Ronneberger introduced UNet architecture in the paper "U-Net: Convolution Networks for Biomedical Image Segmentation" [59].This architecture achieved excellent performance on different biomedical segmentation applications.

Identification of COVID-19 Using CT Imaging Tools (2020-2022)
In this particular context, our focus has been on investigating the practical application of CT imaging as a valuable diagnostic tool for accurately detecting instances of COVID-19.This investigation has been concentrated on the timeframe spanning from 2020 to 2022 (See Table 2).[66] used the DL algorithm to discriminate between negative mild and severe cases of COVID-19.They reported the AUCs of 0.944, 0.860, and 0.884 for each class.Jaiswal et al. [67] implemented the DenseNet201 model and compared the performance with three different DL models.They reported that the maximum accuracy of 99.82% was achieved using DenseNet201.Babukarthik et al. [68] implemented a model based on the GDCNN algorithm to classify normal vs. COVID-19.The proposed model's performance was compared with five different DL models.They reported that the proposed model outperforms other models with an accuracy of 98.84%.Mohammed et al. [69] developed the RestNet+ model to classify COVID-19 vs. other pneumonia.They reported proposed model accuracy of 77.6%.Han et al. [70] proposed attention-based deep 3D multi-instance learning (AD3D-MIL), compared it with traditional multi-instance learning, and reported the proposed model accuracy of 97.9%.Jiang et al. [71] used AI models for the diagnosis of COVID-19 by use of a cGAN structure image that can generate realistic city images with two types of infections: ground-glass opacity and consolidation.They reported that the model achieved an accuracy of 98.37% for COVID19+ vs. COVID-19−.Gunraj et al. [72] introduced the COVIDx-CT model to identify COVID-19 vs. normal vs. pneumonia.The proposed model performance was compared with the ResNet50, NASNet-A-Mobile, and EfficientNetB0 models.They reported that the model achieved an accuracy of 99.1%.
Fan et al. [47] developed the Inf-Net (covid lung CT infection segmentation) and semi-Inf-Net model (cutting-edge segmentation).These models could detect objects with low-intensity contrast between infected and normal tissues.They reported specificity and sensitivity of 97.55% and 86.75%, respectively.Mishra et al. [73] applied a deep CNNbased approach including five models: VGG16, InceptionV3, ResNet50, DenseNet121, and DenseNet201 decision and developed a new model.They reported a model AUC value of 88.3%.Javor et al. [74] devised a DL model based on the ResNet50 model architecture to classify covid patients with 6868 CT images.They reported an AUC value of 95.6%.Silva et al. [75] devised a model, Efficient Covid Net, along with a voting-based approach and cross-dataset analysis.The proposed model achieved an accuracy of 87.68%.Pathak et al. [76] used ResNet50 architecture to identify COVID-19 and reported an accuracy of 93.01%.Wu et al. [77] conducted a multicenter study with 294 COVID-19-positive cases.They designed a model using a DL network trained using multi-view images, and 76% accuracy was reported.Peng et al. [78] [102] proposed a model to assess the disease severity using VGG16 network architecture.Benbrahim et al. [103] used InceptionV3 and ResNet50 architecture to identify normal vs. COVID-19 patients.The model has attained an accuracy of 99.01%.Sharma et al. [104] proposed a model based on ResNet architecture and Grad-cam, which achieved an accuracy of 87.6%.
In 2020, we found 47 articles for identifying COVID-19 using different types of datasets, where 23 worked on the private dataset.Most authors used CNN, ResNet, DenseNet, Aleand xNet, and DCNN architectures and compared results with available DL methodology.A few of them modified the existing structure of DL architecture and proposed a new model with different names like Covid19Net, CoVNet, DecoVNet, and so on.In addition to these, we found a few articles where authors also concentrated on classifying and segmenting COVID-19 patients from influenza, nonviral community-acquired pneumonia, and nonpneumonia diseases.

2021
Ibrahim et al. [105] used four deep learning architectures, namely VGG19-CNN, ResNet152V2, ResNet152V2 + GRU, and ResNet + Bi-GRU, to classify COVID-19 vs. normal.The maximum accuracy of 98.05% was achieved by the VGG19-CNN model.The authors claimed this model could also identify lung cancer and pneumonia, the first deep learning model in the literature.Goncharov et al. [106] implemented a multitask spatial-1 model to identify COVID-19 vs. normal class using severity score.The model outperforms other approaches and achieved an AUC score of 0.97 ± 0.01 between COVID-19 and healthy control.Additionally, the Spearman correlation method was used to find severity quantification.Zhang et al. [107] devised a new five-layer DCNN model with 3CB + 2FCBs for COVID-19 diagnosis.The implemented method was compared with six deep learning algorithms: RBFNN, K-ELM, ELM-BA, 6L-CNN-F, GoogleNet, and ResNet18.A model accuracy of 93.64% was obtained with stochastic pooling, providing better performance than average and max pooling.Song et al. [108] implemented the DRE-Net model to identify COVID-19 vs. healthy people among 274 patients.They compared their model performance with three DL models, namely VGG16, DenseNet, and ResNet.The maximum accuracy of 86% was achieved by DRE-Net.Additionally, this model could also identify bacterial pneumonia patients due to covid with 93% accuracy.Yao et al. [109] conducted a retrospective multicenter study to identify mild COVID-19 pneumonia by implementing CNN-based DL model.They also compared the model performance with the radiologist.The overall sensitivity and specificity were 91.5% and 90.5%.Acar et al. [110] used nine DL models, namely VGG16, VGG19, Xception, ResNet50, ResNet50V2, InceptionV3, Inception-ResNetV2, DenseNet121, and DenseNet169.They used internal and external datasets to access each model's performance with normal and augmented datasets.Finally, accuracy was improved from 3% to 9% for each DL model.Ravi et al. [111] used a stacked ensemble meta-classifier and deep learning-based feature fusion approach in CXR and CT images to classify COVID-19 vs. non-COVID-19 samples.They performed a comparison study with existing available pertained CNN models.Finally, a maximum accuracy of 99% was reported using CT data.
Chen et al. [112] used different ResNet architectures to classify normal vs. COVID-19 vs. other pneumonia.Of all, ResNet50 provided the best classification accuracy of 91.21%.They also compared their result with the radiologist, and the proposed model achieved an overall accuracy of 89.01%.Huang et al. [113] implemented a FaNet network to classify normal vs. COVID-19 with 416 samples.They compared the result with six different models: AlexNet, ResNet, MobileNet, VGG, SENet, and DenseNet.Finally, 98.28% accuracy was reported for diagnosis assessment via FANet.The authors also claimed that their proposed model could assess the severity of COVID-19 with an accuracy of 94.83%.Jangam et al. [114] utilized stack ensemble techniques to develop an automatic COVID-19 detection system and compared the performance with four pretrained DL models.They reported 84.73%, 99%, and 90.75% accuracy for three different datasets.Singh et al. [115] implemented a MobileNet model that takes lesser time for covid classification.They compared their model performance with three DL model architectures: the proposed model reported an accuracy of 96.40%.Alirr et al. [116] devised FCN using Unet architecture for COVID-19 infection vs. lung segmentation that was evaluated qualitatively and quantitatively with a diverse dataset.They reported that the proposed model has a sensitivity and specificity of 82.2% and 95.1%, respectively.Kundu et al. [117] established a fully automated DL model for differentiating COVID-19 vs. non-COVID-19 patients.The proposed model performance was compared with three DL models: InceptionV3, ResNet34, and DenseNet201.They reported a proposed model accuracy of 97.81%.Saad et al. [118] implemented DFC model to identify COVID-19 vs. non-COVID-19 samples.The model performance was compared by 14 other different methods.They reported a model accuracy of 98.9%.Fung et al. [119] implemented an SSInfNet model that utilized DL to support rapid COVID-19 diagnosis and reported an AUC of 98.66%.Tan et al. [120] implemented the VGG16 model to classify COVID-19 with an accuracy of 98%.Lascu et al. [121] utilized ResNet101 architecture to classify COVID-19 among four class labels.They reported an accuracy of 94.9%.Lassau et al. [122] built an AI model to determine the severity score to diagnose severe evolution for COVID-19.They compared their proposed approach to an existing severity score of 11, and performance improvement was reported.Pan et al. [123] determined the correlation between the conventional CT scoring system and the proposed DL-based quantification.They reported that the proposed DL quantification correlated with conventional CT scoring and demonstrated a potential benefit in estimating COVID-19 severity.Yan et al. [124] developed a Fast.AI ResNet framework to differentiate COVID-19 vs. pneumonia vs. normal.The authors compared the model performance with three DL models: VGG16, DenseNet121, and ResNet152.Finally, the maximum accuracy was achieved by ResNet50 with Fast.AI.Shalbaf et al. [125] used 15 pretrained CNN architectures and developed an ensemble model using majority voting criteria.Rahimzadeh et al. [126] proposed a new feature pyramid network with the ResNet50V2 model to classify COVID-19.The model performance was compared with two DL models: Xception and Resnet50V2.The proposed model's accuracy of 98.49% was reported.Lee et al. [127] devised the DeteCT model to automatically predict COVID-19+ from COVID-19-, pneumonia, and normal controls.They reported that the proposed model AUC value is more than 80% on most test sides.Mishra et al. [128] used a deep learning algorithm to diagnose COVID-19.The authors also worked on finding ANN's severity index of covid infection.They reported an accuracy of 99%.Zhang et al. [129] proposed an improved segmentation model called the residual attention U-shaped Network.The model was evaluated using 100 scan datasets resulting in mIoU and dice coefficient values of 84.5% and 73.4%, respectively.
Barbosa et al. [130] conducted a retrospective study to differentiate between COVID-19 vs. non-COVID-19 patients.They concluded that the CNN-trained model achieved an expert level of accuracy in quantifying COVID-19 airspace disease.Zhao et al. [131] developed a new approach, an image deformation-based segmentation model, SP-V-Netbased.They reported that the model achieved an accuracy of 94.60% for COVID-19 classification.Jadhav et al. [132] proposed a COVID-19-view by incorporating a novel DL method to classify the patients into positive and negative COVID-19 cases.This model can also be used for lung segmentation, lesion localization, and detection.They reported an accuracy of 95.2%.Guiot et al. [133] developed a detection model on 181 COVID-19+ cases using VGG16 architecture.They reported the proposed model accuracy of 85.18%.Yao et al. [134] devised a model named as CSGBBNet for the classification of COVID-19 and reported an accuracy of 98.49%.Singh et al. [135] designed a DL-based model for detecting COVID-19.The model performance was compared with four DL models: Gen-ProtoPNet, NP-ProtoPNet, ProtoPNet, and VGG16.They reported that the proposed model outperformed with an accuracy of 99.29%.Zhu et al. [136] used ResNet50 to classify normal vs. COVID-19 by 1357 confirmed positive cases.The model performance was compared with VGG19 +GoogleNet architecture-based DL model.They reported that the model achieved an accuracy of 93%, which is better than other DL models.Kuchana et al. [137] developed a model based on UNet architecture for two segmentation tasks: lung spaces and COVID-19 anomalies.The model performance was compared with standard UNet and attention UNet.They reported that the proposed model obtained a F1 score of 97.31%.Khalifa et al. [138] used DL semantic segmentation architecture for COVID-19 lesion detection.The model consists of an encoder and decoder component.They reported that the model achieved 99.3% accuracy.Bhuyan et al. [139] developed a model to detect COVID-19, classification, and segmentation.The model performance was compared with and without mass segmentation via different kth validation techniques.They reported that the proposed model (FrCN) accuracy was optimal with mass segmentation and fourth-fold validation techniques.Heidarian et al. [140] [148] used five CNN models, namely DenseNet169, InceptionResNetV2, InceptionV3, ResNet50, and VGG16, to identify the effectiveness of the dataset.They reported that a maximum accuracy of 96% was obtained using synthetic data.Hu et al. [149] proposed DSN-SAAL model, and performance was compared with seven models: VGG16, ResNet50, DenseNet169, Self-Trans, contrastive COVIDNet, transfer CheXNet, and cross-dataset analysis.They reported that the proposed model outperforms all used datasets.The achieved average accuracy of the proposed model is 95.43%.Jingxin et al. [150] used the DL approach based on ResNet50 and compared it with Mark R-CNN, UNet.They reported an accuracy of 97.83% via Ours-SP.Balaha et al. [151] developed a covid detection model named CovH2SD based on VGG16 architecture.A total of nine experiments were performed (ResNet50, ResNet101, VGG16, VGG19, Xception, MobileNetV1, MobileNetV2, DenseNet121, and DenseNet169) on CT images.Of all, the best result was achieved by VGG16.Turkoglu et al. [152] proposed a model named as MKs-ELM-DNN based on DenseNet201 architecture.They compared the performance of six models (AlexNet, GoogleNet, VGG16, MobileNetV2, ResNet18, and InceptionV3).The maximum accuracy of 98.36% was achieved by DenseNet201.Ahamed et al. [153] proposed a model based on a modified ResNet50V2 architecture to differentiate between COVID-19, normal controls, and viral and bacterial pneumonia.The model performance was compared with nine pre-trained CNN models and reported an accuracy of 99.99% for two-class cases (COVID-19/normal).Pathan et al. [154] devised a COVID-19 classification model that deployed an ensemble of five CNNs architecture for feature extraction, and extracted features were again selected by a binary grey wolf optimizer.Model performance was compared with four existing studies, and 96% accuracy was reported.Cruz et al. [155] implemented a model based on an ensemble method using six pretrained DL models: VGG16, ResNet50, wideResNet50-2, DenseNet161, DenseNet169, and InceptionV3.They compared the model performance with eight different models.The maximum accuracy of 86.70% was achieved by the proposed ensemble method.Hasan et al. [156] designed a model based on two fundamental deep learning models, VGG16 and VGG19, for the classification of COVID-19.The model performance was compared between original vs. modified images.They reported that 87.37% accuracy was achieved using original images, whereas 90.14% accuracy was reported for modified images.Basset et al. [157] devised a model to classify COVID-19 based on lung area infection segmentation.They compared their model performance against other studies: R2UNet, CE-Net, and CPFNet.The proposed model outperforms with 96.80% accuracy.
Fu et al. [158] designed and compared a classification model named DenseAnet with seven models.They reported that the maximum accuracy of 90.27% was achieved using DenseAnet.Aslan et al. [159] proposed a hybrid model based on mAlexNet+ BiLSTM architecture, and 98.70% accuracy was reported.Kundu et al. [160] used the Sugeno fuzzy integral ensemble of four pretrained deep learning models, namely VGG11, GoogleNet, SqueezeNet v1.1, and wideResNet50-2.The proposed model achieved an accuracy of 98.93%.Müller et al. [161] used 3D UNet architecture to classify COVID-19+ and normal slices, and the performance was estimated using the DSC score.They reported that the proposed model performed well compared with existing studies.Li et al. [162] developed a deep learning model called CheXNet and evaluated their proposed method with other existing methods.Finally, maximum accuracy was achieved by the proposed DL model with an accuracy of 87%.Zhang et al. [163] created an end-to-end multiple-input deep convolutional attention network based on a convolution attention module.The model provided better outcomes than eight state-of-the-art approaches.They reported that the model obtained an accuracy of 98.02%.Xu et al. [164] proposed two models: CARes-UNet and semi-CARes-UNet.They compared the model's performance with nine existing models, and the semi-CARes-UNet model provided the best outcome close to ground truth.Mondal et al. [165] proposed a DL model, namely CO-IRv2, to classify COVID-19.They used three optimizers and achieved 96.18% accuracy for binary classification.Chen et al. [166] developed an ensemble CNN (covid-CNN) model based on five pretrained DL architectures: VGG19, ResNet101, DenseNet201, InceptionV3, and InceptionResNetV2.They compared the proposed model's performance with the existing CNN model.Covid-CNN obtained the maximum accuracy of 96.7%.Alshazly et al. [167] considered seven CNN networks: SqueezeNet, Inception, ResNet, ResNeXt, Xception, ShuffleNet, and DenseNet, and their performance was compared.ResNet101 and DenseNet201 performed best, with an accuracy of 99.4% and 92.9%, respectively.
Voulodimos et al. [168] proposed a few-shot UNet model and compared it with the conventional UNet model.They observed that the proposed model F1 score was improved by 5.394 ± 3.015% and the increment of precision and recall value by 1.162 ±2.137% and 4.409 ± 4.790%, respectively.Khan et al. [169] proposed a model MC-SVM along with optimal deep model features.They reported that the proposed model achieved an accuracy of 98%.Rajasekar et al. [170] designed a hybrid learning model to identify COVID-19.They used CNN for feature extraction, and MLP was employed for classification.The model showed an accuracy of 94.89% compared with conventional MLP and CNN, where 86.95% and 80.77% accuracy were noted, respectively.Xie et al. [171] designed a CNNbased DL model to identify COVID-19 from other suspected ones.They used UNet and COVIDNet architecture for segmentation, whereas the ResNet50 network was deployed for classification.Sethy et al. [172] devised three approaches: VGG19 + SVM, VGG19, and LBP feature + quadratic SVM to identify COVID-19 patients.They used 13 pretrained DL models to compare the proposed approach.The average accuracy was 77.28%, whereas the maximum accuracy of 85.7% was achieved by LBP feature + quadratic SVM approach.Özyurt et al. [173] used shuffleNet CNN architecture to classify COVID-19 patients and reported an accuracy of 98.99%.Garain et al. [174] designed a three-layer DCSNN to screen for COVID-19.They developed two variants: spike train-based, and potential-based, and the performance was compared with three DL architectures.Of all, the potential-based model provided the optimal outcome with an accuracy of 99.51%.
Elghamrawy et al. [175] implemented a COVID-19 classification model, AIMDP based on CNN architecture.They compared the model performance with five existing DL models, and the designed model achieved an accuracy of 98%.The authors also used WOA optimization techniques to select the most relevant patient sign.Sen et al. [176] used CNN to extract the features and the bi-stage feature selection method to identify the most relevant feature for identifying COVID-19.Finally, the SVM classification algorithm reported 90% and 98.39% accuracy for two datasets.Teodoro et al. [177] applied pretrained CNNs with three classification algorithms: KNN, SVM, and DNN.Among all, CNN EfficientNetB0 performed best along with the SVM-RBF kernel.They reported that the proposed approach achieved an average performance of 98.56%.Yasar et al. [178] used 24-layer CNN architecture with and without local binary pattern CT images for COVID-19 vs. normal.They reported that the maximum efficiency of 94.56% was obtained using no pipeline approaches instead of pipeline approaches.Brahim et al. [179] proposed a DL model named COV-CAF, and the performance was compared with four preexisting COVID-19 classification models and reported an accuracy of 97.59%.Afshar et al. [180] introduced a new COVID-19 dataset named COVID-CT-MD and applied DL and ML algorithms to check the effectiveness of the dataset.They reported that 93% accuracy was obtained by introducing the dataset and underlying studies.Liu et al. [181] specially developed an automated classification model, COVIDNet, to distinguish between COVID-19 and seven other types of pneumonia.They reported that the model achieved an accuracy of 94.3%.Kundu et al. [182] devised an ensemble model based on three CNN architectures: VGG11, wideResNet50-2, and InceptionV3.They reported an average accuracy of 98.86%, which was better than other DL architecture.Pal et al. [183] used two CNN architectures: VGG16 and InceptionV3, to classify COVID-19.They reported an accuracy of 84%, which was achieved using individual CNN models.
Biswas et al. [184] initially used three CNN architectures: VGG16, ResNet50, and Xception.The authors introduced a stacked model (VGG16 + ResNet50+ Xception) via an ensemble learning technique, and 98.79% accuracy was reported.Helwan et al. [185] used three DL models, namely ResNet18, ResNet50, and DenseNet201.Of all, DenseNet201 performed best, with an accuracy of 98.7%.Castiglione et al. [186] proposed the ADECO-CNN approach and compared it with pretrained CNN models, namely VGG19, GoogleNet, and ResNet.They reported that the proposed approach-based model achieved an accuracy of 99.99%.Yan et al. [187] performed a quantitative analysis and designed a DL model named CovidSegNet to segment COVID-19 infections.The model performance was compared with the preexisting FCN, UNet, VNet, and UNet++ networks.Finally, the CovidSegNet model provided the best performance with a dice coefficient of 72.6% for COVID-19 segmentation.Suri et al. [188] presented a COVLIAS 1.0 system that consists of two hybrid DLs for COVID-19 segmentation.They compared the model performance with the conventional NIH model.
Nair et al. [189] proposed the CoRNet DL model and compared it with five existing DL models: AlexNet, VGG16, SqueezeNet, VGG19, and ResNet50.CoRNet achieved high performance with an AUC value of 95%.Wan et al. [190] designed a modified AlexNet architecture and compared it with LBP + SVM, and deep feature + SVM, where AlexNet performed well with an accuracy of 94.75%.Guo et al. [191] proposed a model based on a modified version of ResNet18 to diagnose COVID-19.They reported 98.88% and 99.80% model accuracy for two-and fivefold cv.Xia et al. [192] proposed a rapid screening classifier to diagnose COVID-19.The classifier provided the best outcome with CXR and clinical features, whereas CT-based diagnosis outperformed severe cases of COVID-19.Polat et al. [193] used a CNN to identify COVID-19 and all balanced datasets.The accuracy of the proposed model was 93.26%.Li et al. [194] developed VGG16 deep learning model to classify COVID-19 vs. CP vs. NC.They reported an accuracy of 93.57%, which was achieved using a newly designed model.Owais et al. [195] proposed the DAL-Net model and compared the model performance with seven DL models: VGG16, VGG19, UNet, FCN, DeepLabV3+, MobileNetV2, and ResNet.They reported an AUC of 97.80%, which was better than others.Jia et al. [196] proposed a modified ResNet to classify COVID-19 vs. non-COVID-19 infections vs. normal control.Five CNN architectures (VGG, Inception, DenseNet, SqueezeNet, MobileNet) and two specific detection models (COVID net and CovidNet-CT) were used for comparative studies.They reported that the proposed model achieved an accuracy of 99.3%.He et al. [197] proposed multitasking multi-instance UNet to identify the severity assessment of COVID-19 and the segment of the lung lobe.They reported an accuracy of 98.5% for the assessment of COVID-19 severity.Murugan et al. [198] applied a whale optimization algorithm to ResNet50 to optimize DL architecture and built a WOANet model to classify COVID-19.The proposed architecture achieved an accuracy of 98.78%, providing a better outcome than the nonoptimized ResNet50.Naeem et al. [199] introduced a new DL model named CNN-LSTM and compared it with conventional DL models such as VGG16 and VGG19.The proposed model achieved an average accuracy of 90.98.Kalane et al. [200] proposed a UNet architecture to classify COVID-19; overall, 94.10% accuracy was reported.Fouladi et al. [201] used ResNet50, VGG16, CNN, CAENN, and machine learning approaches (NN, SVM, RF, SGD LR, and MLP) to classify COVID-19 where NN achieved high performance with an accuracy of 94%.On the other hand, the classification accuracies of ResNet50, VGG16, CNN, and CAENN were obtained as 92.24%, 94.07%, 93.84%, and 93.04%, respectively.Wang et al. [202] developed a new approach based on a deep feature fusion combination of an improved CNN model.This model performed better than the other 15 DL models, with an average accuracy of 96.66% reported.Yu et al. [203] proposed three models, ResNet101-C, NNet-C, and ResGNet-C, to classify pneumonia caused by COVID-19 vs. normal.The ResGNet-C model provided better performance with an accuracy of 96.62%.Gao et al. [204] proposed DCN for COVID-19 diagnosis that can be achieved from an individual classification level.They used internal and external datasets to evaluate this proposed model by comparing five DL models.The proposed model outperforms by attaining 96.74% and 92.87% accuracy for internal and external datasets, respectively.Sahoo et al. [205] implemented the COVIDCon model and compared it with other state-of-the-art algorithms.They reported that the proposed model attained an accuracy of 99.06% with the CT scan dataset.Lacerda et al. [206] built an AI model based on optimized VGG16 and compared it with the baseline model of VGG16.They reported that the optimized model attained an accuracy of 88%, whereas 87% accuracy was reported for the baseline model.
Siddiqui et al. [207] introduced the ID2S-COVID19-DL system to classify COVID-19, and 98.11% system accuracy was reported.Haikel et al. [208] produced a DL model named EfficienNet-B3-GAP-ensemble and applied it to two datasets.They reported that the proposed model achieved an accuracy of 99.72% and 88.18%, respectively.Bekhet et al. [209] proposed a fully automated hybrid CNN model to classify COVID-19.They reported that the proposed model attained accuracy of 92.02%.Kaushik et al. [210] developed the VGG16 model and compared the proposed model performance with three DL models: CNN, DenseNet, and XceptionNet.The authors reported that the VGG16 model outperformed with an accuracy of 95.26%.El-Shafai et al. [211] built an automated COVID-19 detection model named SR-GAN and compared the model accuracy of 99.05% with 13 DL models.Masud et al. [212] proposed a CNN model and compared it with three DL architectures: MobileNetV2, InceptionV3, and Xception.They reported that the proposed model achieved an accuracy of 96%.El-Shafai et al. [213] used CNN architecture and studied optimizers with different batch sizes and constant learning rates.Finally, a comparative study was presented using optimizer and activation functions.They reported that the proposed model achieved 100% accuracy.
Kassania et al. [214] proposed a method based on DenseNet121 + Bagging, and 99% accuracy was achieved by this method for the detection of COVID-19.Wang et al. [215] proposed a model based on f 3D UNet++-ResNet50 architecture for the classification and segmentation of COVID-19.They reported that the model attained an AUC score of 99.1%.Ahuja et al. [216] implemented a DL model based on ResNet18 architecture that attained 99.4% accuracy.Pu et al. [217] used the UNet network and BER algorithm to identify COVID-19 severity and progression.They reported that the proposed model performed well, with a sensitivity of 95%.Maghdid et al. [218] developed a model based on AlexNet architecture with an accuracy of 94.1%.Kumar et al. [219] used a deep neural network to detect COVID-19, and 98.4% accuracy was reported.Wang et al. [220] applied a modified Inception transfer learning model with 1065 positive COVID-19 cases.The model attained an accuracy of 79.03% on the external testing dataset.
In 2021, we found 116 articles and 36 that worked based on the private datasets to identify COVID-19.Most authors utilized CNN, ResNet-XX, VGG-XX, DenseNet-XX, and UNet architecture, and some authors introduced new models such as FewShot, Fractal-CovNet, CCSHNet, COVIDCon, and CovH2SD.Later, the authors applied an optimization approach to their proposed model and compared their model performance with existing models.All details are documented in Table 3.

2022
Khurana and Soni [221] used four DL architectures, namely ResNet50, efficient netB0, VGG16, and CNN, to detect the presence of COVID-19.Of all, ResNet50 obtained the highest accuracy of 98.9%.Canayaz et al. [222] proposed two new methods to diagnose COVID-19 using DL and ML algorithms.Two DL models, ResNet50 and MobileNetV2, are used for feature extraction along with two classification algorithms, SVM and KNN.The total experiment was performed in three steps using individual and mixed datasets.The reported accuracies are 95.79%,99.06%, and 99.37% for MobileNet, ResNet50 + SVM, and ResNet50 + KNN, respectively.Subhalakshmi et al. [223] proposed a DLMMF model to identify COVID-19.The proposed architecture is based on InceptionV4 and VGGNet16, which are used to extract features from the dataset.The Gaussian naïve Bayes classifier was deployed as a final classifier for disease detection.Zouch et al. [224] used two DL architectures, ResNet50 and VGG19, to detect COVID-19.Both models obtained an accuracy of 99.35% and 96.77%.Balaha et al. [225] introduced a DL framework for early detection and prognosis of COVID-19.Seven different CNNs architectures are used, and for classification, maximum accuracy of 99.61% was obtained using EfficientNetB7.The authors also reported an accuracies of 98.70% and 97.40% obtained by ensemble bagged trees and trees (fine, medium, and coarse) for the early prognostic phase.
Habib et al. [226] proposed a classification system for COVID-19 with a hybrid feature extraction approach.Three different architectures, ResNet101, DenseNet201, and weber local descriptor, were used to classify COVID-19, lung opacity, healthy, and viral pneumonia.They reported that the proposed model achieved an accuracy of 99.3%.Montalbo et al. [227] used six DL architectures, InceptionV3, Xception, ResNet50V2, DenseNet121, and EfficientNetB0, to classify COVID-19.They compare the performance between truncated models and general models.Of all, the maximum accuracy of 97.41% was obtained using InceptionResNetV2 with truncated models.Ali et al. [228] devised a model to identify COVID-19 severity using CNN and KNN.They compared the result with the existing classification model, and 95.65% accuracy was reported.In the next experiment, modified CNN achieved an accuracy of 92.80% for detecting pneumonia on mixed data.Pandey et al. [229] proposed an efficient model to diagnose COVID-19 using three DL architectures (ResNet50, MobileNet, VGG16).The authors used image segmentation and compared the model performance.The maximum accuracies of 99.28% and 83.18% were achieved via VGG16 along with OTSU segmentation and without segmentation.
Liu et al. [230] introduced a new framework named DCNN + IMPA (internet protocol marine predator) to diagnose COVID-19.They reported that the model achieved an accuracy of 97.57%.Luo et al. [231] developed a model to detect COVID-19 vs. normal vs. CAP using Resenet-50 and UNet.They reported a maximum efficacy of 93.84%, and 92.86% was achieved in testing and validation set via UNet.Saheb et al. [232] proposed an ADL-CDF architecture to detect COVID-19.A maximum accuracy of 98.49% was reported.Batra et al. [233] proposed a model based on the architecture of InceptionV3, and the reported accuracy was 93%.The authors also worked on X-ray images where the same model performed best.The model performance was compared with two other models, VGG16 and ResNet50V2.Cao et al. [234] introduced a CNN model to detect COVID-19, and 82.7% accuracy was achieved.They compared the model performance with three other CNNs: Goolenet-RI, ResNet50-RI, and GoogleNet-TL.Of all, the top F1 score of 79.1% was obtained via the proposed model.Yazdani et al. [235] developed a model based on CNN and NN to detect COVID-19 using low-level and deep features.Local neighborhood difference pattern was performed to extract handcrafted features, and MobileNetV2 was used to extract deep features.The optimal accuracy of 99.61% was obtained by combining texture and deep features using CNN architecture.Bhuyan et al. [139] experimented with classifying COVID-19 with CNN architecture.The authors compared the model performance with mass segmentation and without mass segmentation with a fourfold validation technique.They reported that average accuracy of 99% and 97.75% was achieved with mass segmentation and without mass segmentation, respectively.Ibrahim et al. [236] used hybrid deep learning techniques to identify COVID-19.They used three DL architectures, namely, VGGNet, CNN, high-resolution network with segmented images, and 95% accuracy was reported.Akinyelu et al. [237] performed a comparative study with 12 DL architectures: VGG16, VGG19, ResNet50, InceptionV3, Xception, MobileNetV2, ResNet101V2, DenseNet169, DenseNet121, InceptionResNetV2, NASNetLarge, and densenet201.NASNetLarge, InceptionResNetV2, and DenseNet169 provided good accuracies of 99.86%, 99.78%, and 99.71%.The authors also reported that VGG16 and densenet121 produced the highest sensitivity of 99.94%.Florescu et al. [238] proposed a model based on VGG16 with a federated learning approach to detect COVID-19.They reported that the model performed well in the training and validation phase with categorical accuracy of 83.82% and 79.32%, respectively.
Jingxin et al. [150] introduced the DL model for COVID-19 lesion detection and segmentation.They used ResNet50 architecture, and 98.39% accuracy was reported.Baghdadi et al. [239] devised a model for COVID-19 detection on both two and three classes.Maximum accuracies of 99.74% and 98% were attained via MobileNetV3Large (two-class) and SENet154 (three-class), respectively.They also compared the model output with other CNN models like LeNet5 CNN, covid faster R-CNN, lightCNN, fuzzy + CNN, dynamic CNN, and optimized CNN.Shaik et al. [240] used various pretrained models such as VGG16, VGG19, InceptionV3, ResNet50, ResNet50V2, InceptionResNetV2, Xception, and MobileNet.Further, the authors created a strong ensemble approach using these trained models to detect COVID-19 infection.The maximum average accuracy of 93.33% was reported with 5-and 8-clf, respectively.Reis et al. [241] devised a new COVID-DSNet model to detect COVID-19 along multiclass target labels.The maximum accuracy of 97.60% was achieved via CT scans where the target labels are COVID-19 vs. normal.The author further used mixed datasets (X-ray and CT) and proposed three models: COVID-DSNet + LSTM, COVID-DSNet +FCC, and COVID-DSNet.The reported average accuracy was 95.64%.Garg et al. [242] devised a DL model based on efficient net-B5 to detect COVID-19.The model attained an accuracy of 98.45%, and 97.69% accuracy was reported for multiclass datasets.Fan et al. [243] developed a COVID-19 detection model based on CNN (ResNet152) architecture and transformer network (Deit-B).The proposed model attained a maximum accuracy of 96.7%, better than a typical CNN (95.2%).
A 3D CNN is interesting.Karthik et al. [244] developed a DL framework based on 3D CNN, and the model performed best.Verma et al. [245] used NNs to train the model to make a CovCT application for the detection of COVID-19.The developed model attained an accuracy of 99.58%.Smadi et al. [246] developed a model named SEL-COVIDNET for the diagnosis of COVID-19, which was tuned with DenseNet121, InceptionResNetV2, and MobileNetV3Large.The authors experimented with multiclass and binary classification.They reported that their model obtained an accuracy of 98.79% (COVID-19 vs. normal).Further, the model achieved an accuracy of 98.52% for X-ray and CT mixed data.Fallahpoor et al. [247] [249] proposed the LightEfficientNetV2 model and compared the performance with the two best other models, namely MobileNetV2 (without tuning) and Xception (with tuning).They reported that 97.48% accuracy was attained via LightEfficientNetV2, which was best compared to two other models.Li et al. [250] proposed a MultiR-Net, a 3D deep learning model to classify COVID-19 and lesion segmentation.The proposed model performance was compared with four different models: DenseNet, Res2Net, Zhou's, and JCS.They reported that the highest classification accuracy of 92.647% was achieved using MultiR-Net.Hemalatha et al. [251] used a hybrid random forest deep learning classifier to detect COVID-19, and 99% accuracy was reported.The authors also claimed that their proposed methodology is fitted for edge computing with higher detection accuracy.Wang et al. [252] built an SSA-Net segmentation model, which helps to diagnose COVID-19, and 70.31% DSC was reported.Qi et al. [253] used a capsule network with ResNet50 for slice-level prediction, and 93.4% accuracy was reported.Also, the authors claimed their method achieved 100% accuracy for patient-level prediction.
O guz and Ya gano glu [254] proposed a hybrid method combining in-depth features extracted from ResNet50, and SVM was used as a final classifier to detect COVID-19.They used AlexNet, ResNet50, ResNet101, VGG16, VGG19, GoogleNet, SqueezeNet, and Xception architecture to extract deep features.Five classification algorithms, SVM, RF, KNN, DT, and NB, were deployed on extracted features.Finally, maximum accuracy of 96.296% was obtained via the SVM classifier with ResNet50.Ravi et al. [111] used EfficientNet architecture to predict COVID-19, and 99% accuracy was reported.The model performance was compared with others pretrained.The authors also used the t-SNE method to visualize CT test data.Yang et al. [255] proposed the F-EDNC model to recognize COVID-19 and compared the performance with FC-EDNC, O-EDNC, and CANet.The maximum efficacy of 97.55% was attained via F-EDNC.Mijares et al. [256] used CNN to diagnose COVID-19, and 94.89% classification accuracy was reached.Heidari et al. [257] utilized blockchainbased CNNs to detect COVID-19.They reported that 99.34% and 99.76% accuracy was attained in the testing phase for four-and two-class classifications.Singh and Kolekar [115] developed a fine tune model based on MobileNetV2 architecture to diagnose COVID-19.The model performance was compared with three deep learning models such as VGG19, DenseNet201, and VGG16.They reported that the proposed model attained an efficacy of 96.40% with ten-times shorter response time.Ortiz et al. [258] devised a prognosis model for COVID-19, and 91% accuracy was reported.The authors noted that the accuracy label was improved in CT features, patient demographics, and image segmentation.Sangeetha et al. [259] deployed two DL architectures, VGG19 and ResNet152V2, to diagnose COVID-19.They reported that both models achieved an accuracy of 98%.Mohammed et al. [260] proposed an optimal deep learning model based on ResNet50 architecture to diagnose COVID-19.A total of 15 DL models were used to compare the performance, and a maximum of 91.46% accuracy was reached via ResNet50.In contrast, InceptionV3 provided the lowest performance.
Joshi et al. [261] introduced the MFL-Net model to recognize COVID-19 using individual and combined datasets.They reported an average accuracy of 96.12%, whereas 96.13% was attained via the mixed form.Zhang et al. [262] introduced a DL model based on VGG19 with globalmaxpool2D to detect COVID-19.They reported that the proposed model achieved an accuracy of 94.12%, which was best compared to others.Mouhafid et al. [263] utilized two ensemble learning methods, stacking and weighted average ensemble (WAE), to combine the performance of three fine-tuned-based learners such as VGG19, ResNet50, DenseNet201.The result showed that the maximum accuracy of 98.59% and 95.05% was achieved via WAE.Dara et al. [264] applied ResNet architecture to implement the classification algorithms to identify COVID-19.The authors used ResNet18, ResNet50, ResNet101 and compared their performance.Finally, ResNet39 was chosen and used the parameters obtained from ResNet18, ResNet50, ResNet101.The global and local models achieved an accuracy of 97.53 [267] used VGG19 architecture to identify COVID-19, and model performance was compared with efficient net-B4 + CLAHE.They reported that the proposed model achieved an accuracy of 86%.The authors also worked on X-ray and combined datasets where the model attained accuracies of 97% and 90%, respectively.
As before, Manconi et al. [268] proposed a 3D inception CNN architecture to detect COVID-19.Especially, 3D InceptionV1 and InceptionV3 models were built and compared their performance.Further, an ensemble classifier is deployed on CNN models.The maximum accuracy of 98.21% was reached via InceptionV1 with a voting strategy.Cheng et al. [269] designed a COVID-19 detection model based on VBNet + LSTM, and 89% accuracy was reported.Lu et al. [270] introduced a new COVID-19 detection system named CGNet.The model architecture is based on the combination of ResNet18 and k nearest neighbors.The proposed model achieved an accuracy of 97.78%, and Grad-CAM provided visual explanation.Owais et al. [271] proposed DSS-Net to identify COVID-19 and 96.58% accuracy was reported.Yoo et al. [272] developed the 2D UNet model to classify the COVID-19 disease.The authors compared their model performance along with internal and external validation.The Pearson correlation coefficient suggested that the model performs well between UNet outputs and visual CT scores.Suri et al. [273] proposed a model combining DL and hybrid DL to track lesion location and segmentation.They used VGG-SegNet, ResNet-SegNet, VGG-UNet, and ResNet-UNet.The best AI model was ResNet-UNet, with a 92% correlation coefficient with a prediction time of less than 1 s.Ghose et al. [274] proposed a DL model based on densenet-169 to recognize COVID-19.They reported that the proposed model achieved an accuracy of 99.95%.The authors also worked on X-ray images, and 99.59% accuracy was reached.Gunraj et al. [275] proposed Covid-Net CT architecture to detect COVID-19.A maximum accuracy of 99% is reported.Yousefzadeh et al. [276] used UNet architecture for lobe segmentation, and the KNN classifier was applied to predict the severity of infection due to COVID-19.They reported that the proposed model has a 71% to 74% dice score.Choudhary et al. [277] experimented on two DL models, VGG16 and ResNet34, to detect COVID-19.The optimal accuracy of 95.47% was obtained via resnet34.Chouat et al. [278] used four DL architectures, namely VGGNet19, ResNet50, InceptionV3, and Xception, to detect COVID-19.Of all, VGGNet19 and Xception models outperformed with an accuracy of 90.5% and 89.5%, respectively.Dialameh et al. [279] proposed a DL model called Deep CT-Net on DenseNet121 to detect COVID-19.They reported that the proposed model attained an AUC of 88.6%.Venkatachalam et al. [280] proposed a CNN model with BBO that helps the layers selection process.The proposed model performance was compared with existing models, namely VGG16, InceptionV3, ResNet50, MobileNet.The result showed that the proposed model outperformed InceptionV3 and ResNet50.The reported accuracy was 98.5% and 97.6% in the test and train phases, respectively.Latif et al. [281] introduced a hybrid method using ResNet18 and GoogleNet2000 and the extracted features via SVM.They reported that the proposed model achieved an accuracy of 99.91%.El-Shafai et al. [282] devised a CNN framework to detect COVID-19, and 98.49% accuracy was reported.Xue et al. [283] used CNN architecture to identify COVID-19.The proposed model achieved an accuracy of 97.67%.El-Shafai et al. [284] proposed a CNN structure to detect COVID-19, and 100% accuracy was reported.
In 2022, we found 68 articles, and 14 articles worked on a private dataset to identify COVID-19.Most authors used CNN, ResNet-XX, VGG-XX, DenseNet-XX, and EfficientNet-XX architecture.A few authors invented hybrid frameworks like SpaSA and CNN, MOMHTS optimized hybrid random forest deep learning, and ResNet18 + GoogleNet2000 features with SVM.Further, a few authors also worked on X-ray images and combined their CT data with them.The details are documented in Table 4.

Performance Comparison
This section compares the performance of different DL architectures used from 2020 to 2022.There has been substantial research on COVID-19 screening utilizing chest CT scans from the beginning of 2020.A fair comparison among the authors can only be made when they have used exact data collection, evaluation protocol, and performance metrics.Unlike other healthcare issues, in COVID-19, there has been a growing tendency in dataset size resulting in incremental studies over time.As high-end machine learning methods, such as deep learning models, require a massive quantity of data, writers have explored the usage of data collection size.In our study, convolutional neural networks (CNNs) are the most popular, followed by residual neural network (ResNet).In what follows, we categorized the authors' work using the same architecture (see Tables 5-12).
Convolutional neural network (CNN): Eighty-seven authors utilized CNN architecture to detect COVID-19 from 2020 to 2022.Among them, 22, 48, and 17 articles were identified in 2020, 2021, and 2022, respectively (see Table 5).Finally, El-Shafai et al. [284] proposed a model in 2022 that achieved the highest accuracy of 100%.Residual neural network (ResNet): Fifty-two authors used ResNet architecture and 12, 21, and 19 papers were documented in 2020, 2021, and 2022, respectively (see Table 6).In this architecture, fifteen authors preferred private datasets to build their model.Finally, Ahamed et al. [153] proposed a model that achieved the highest accuracy of 99.99% using the C15 dataset.Visual geometry group (VGG): Twenty studies were reported using this network, where 3, 15, and 9 articles were found in 2020, 2021, and 2022, respectively (see Table 7).Balaha et al.'s [16] proposed model achieved the highest accuracy of 99.33% using C1, C3, and C24 datasets.Also, three authors used private datasets.Densely connected convolutional networks (DenseNet): Fourteen studies were identified where authors used this framework.The highest accuracy of 99% was achieved by Kassania et al. [214].Most authors preferred publicly available datasets; only three worked on private datasets (See Table 8).Inception: Nine authors used the Inception network for their study, and three worked on private datasets.Finally, El-Bana et al. [93] achieved the highest accuracy of 99.5% with 99.8% sensitivity.UNet: Twenty-seven articles were identified, and 14 researchers utilized private datasets.The highest accuracy of 99% was achieved by Munusamy et al. [146].EfficientNet: Five authors preferred this network in their research, where the highest accuracy (99.61%) was obtained by Balaha et al. [225].The greater the data, the higher the performance we state in machine learning, particularly for DL-based models.As far as machine learning algorithms are concerned, this is not true, as it necessitates all possible manifestations associated with a specific disease (COVID-19, in our case).However, the size of the dataset allows for the likelihood of having new topics (i.e., manifestations).We refer to earlier works for additional information [285,286].Underfitting and overfitting were not discussed in most of the published papers due to the lack of data.Furthermore, the authors used the holdout method to train/test the model instead of k-fold cross-validation.Another critical issue is an unbalanced dataset, where possible bias was not well argued.

Data Augmentation
Data augmentation plays a significant role in reducing overfitting, which can improve the performance of DL models.The principal objective of the augmentation technique is to uplift the available raw data by adding slightly modified copies of the source or, in some cases, the synthetic image generated from the existing data.For COVID-19 imaging tools, authors often used augmentation methods for aiming more features from the limited available data.Let us discuss the use of augmentation techniques on CT imaging tools over the last two years (see Table 13).We found a total of 10 articles; among them, the researchers of 9 articles used conventional augmentation techniques such as resizing, zooming, gaussian noise, blur, spatial transformation, contrast adjustment, flipping, scaling, cropping, rotation, intensity transformation [69,71,75,84,85,92,93,96,145].Similarly, one author used GAN [62], considered under the classical augmentation approach.

2022
A total of 12 articles are listed where authors used data augmentation methods.The most popular approaches are horizontal flip, anticlockwise rotation, scaling, brightness change, and contrast enhancement [224,236,237,239,244,261,284].In addition, GANs, cycle GAN, CCGAN [225], and RAIOSS [139] were also opted for in a few studies.

Conclusions and Future Scope
In this paper, we studied peer-reviewed research findings/articles on AI-guided tools for COVID-19 analysis/screening using chest CT scans images in the years 2020, 2021, and 2022.Our research was confined to deep learning methods for detecting COVID-19 in CT scans, and we identified data collections, methodological procedures, and discussion of prospective methodologies and challenges.Using the search terms (COVID-19 OR Coronavirus) AND chest CT AND (deep learning OR artificial intelligence OR medical imaging), we systematically reviewed 231 research papers and meta-analyses on the PubMed Central Repository and Web of Science.Unlike standard articles, we did not analyze pre-print publications like those in ArXiv, TechRxiv, and medRxiv.
Future guidelines for effectively using CT imaging and deep learning (DL) in COVID-19 screening should prioritize data augmentation (DA) and diversity to enhance DL model generalizability.Recommendations include augmenting data with varying noise levels and incorporating scans from diverse populations, disease stages, and comorbidities.
Finetuning existing models on COVID-19 data helps to learn disease-specific features while leveraging knowledge from pretraining on other medical images.Emphasizing explainability and interpretability is essential, using techniques like attention maps, saliency maps, or Grad-CAM to highlight influential regions in CT scans, increasing trust in DL model decisions.
zi, b = Ai, b Ki, b (Zi, b − 2 − Zi, b − 8)ˆω, where zi, b represents the total number of infections in a country i, for a date b.Ki, b gives the population ratio still unaffected by COVID-19.Ai, b shows the transmitting rate and Zi, b is the cumulative number of subjects who have shown symptoms by the date b.

Table 1 lists
25CT scan datasets utilized for COVID-19 screening.Out of these datasets, 19 are employed for classification purposes, 3 for segmentation tasks, and 2 for identification and detection purposes, respectively.

Table 2 .
Chest CT imaging tools, dataset, and their performance are measured in terms of accuracy (ACC), specificity (SPEC), sensitivity (SEN), and area under the curve (AUC) for 2020.based on per-lobe lung-and per-patient-level.They reported that the sensitivity was superior to the AI model for per-lobe lung and per-patient levels, respectively. residents , DenseNet121 was pretrained on ImageNet to create a classification model.They reported that the model achieved the highest performance of 89.1 in AUC.Qian et al. [79] used 2D-CNN architecture to design a DL model to classify COVID-19 patients.They reported that the model performed well.Li et al. [80] used the proposed a COVID-Fast model based on CNN to detect COVID-19 and non-COVID-19 cases.They reported the proposed model accuracy of 90.82% for COVID-19 identification.Ahsan et al. [141] implemented six deep CNN models: VGG16, MobileNetV2, InceptionResNetV2, ResNet50, ResNet101, and VGG19 with 400 CT images.They reported that MobileNetV2 outperforms with an accuracy of 98.5%.Zhang et al. [142] implemented a GARCD model to classify COVID-19+ and normal.The performance was compared with four models: ResNet, GADCD, VGG19, and DenseNet.They reported GARCD model achieved an optimal AUC value of 98.7%.Chaddad et al. [143] used deep CNN architecture (AlexNet, DenseNet, GoogleNet, NASNet-Mobile, ResNet18, and Dark-Net) to classify COVID-19 vs. normal.The proposed model has achieved an accuracy of 82%.They also claimed that the proposed model could classify COVID-19+ or COVID-19from X-ray images.Yousefzadeh et al. [144] implemented a deep learning-based covid classification model named ai-corona.They reported proposed model performed well, and the average AUC was 98%.Chen et al. [145] proposed a model based on few-shot learning in ResNet50 architecture to classify COVID-19 vs. non-COVID-19 with few samples.The performance of the new algorithm-based model was compared with three different methods: ResNet152, DenseNet161, and VGG16.They reported accuracy and AUC value of 86.8% and 93.1%, respectively.Munusamy et al. [146] developed a FractalCovNet model consist of UNet architecture to classify COVID-19.They compared the model performance with ResNet50, Xception, InceptionResNetV2, VGG16, and DenseNet.They reported the proposed model accuracy of 99%.Wang et al. [147] developed a CCSHNet model based on a DCFDCA algorithm to classify COVID-19.The proposed model performance was compared with 12 existing models.They reported that the model outperformed.Jiang et al.

Table 3 .
Chest CT imaging tools, dataset, and their performance are measured in terms of accuracy (ACC), specificity (SPEC), sensitivity (SEN), and area under the curve (AUC) for 2021.
* Other data collections were also used.
and 90.66, respectively.Ozdemir et al. [265] used ResNet50 architecture and extended it with a feature-wise attention layer to classify COVID-19.They reported that 95.57% accuracy was obtained via the proposed model.Ahuja et al. [266] devised a COVID-19 classification model named McS-Net based on ResNet18, and 98.07% accuracy was reported.Messaoud et al.

Table 4 .
Chest CT imaging tools, dataset, and their performance are measured in terms of accuracy (ACC), specificity (SPEC), sensitivity (SEN), and area under the curve (AUC) for 2022.

Table 5 .
Comparison: COVID-19 detection using CNN architecture according to the year of publication.
* Other data collections were also used.

Table 6 .
Comparison: COVID-19 detection using ResNet architecture according to the year of publication.
* Other data collections were also used.

Table 7 .
Comparison: COVID-19 detection using VGG architecture according to the year of publication.
* Other data collections were also used.

Table 8 .
Comparison: COVID-19 detection using DenseNet architecture according to the year of publication.
* Other data collections were also used.

Table 9 .
Comparison: COVID-19 detection using Inception architecture according to the year of publication.
* Other data collections were also used.

Table 10 .
Comparison: COVID-19 detection using UNet architecture according to the year of publication.
[222]eNet: Three authors used this architecture, and Canayaz et al.'s[222]proposed model achieved the highest accuracy of 99.06% using C1 and C4 datasets.

Table 11 .
Comparison: COVID-19 detection using MobileNet architecture according to the year of publication.

Table 12 .
Comparison: COVID-19 detection using EfficientNet architecture according to the year of publication.

Table 13 .
Data augmentation techniques according to year of publication.