Table 3.
Summary of state-of-art DL techniques used for the COVID-19 classification using CXR Abbreviations: Acc.- Accuracy, BP-Bacterial Pneumonia, C-COVID-19, CAM- Class Activation Maps, CAP- Community Acquired Pneumonia, CN- COVID-19 negative, FPN- Feature Pyramid Network, HU- Hounsfield Units, Influ.- Influenza, LT- Lung Tumor, N-Normal, NF- No Findings, P- Pneumonia, Rad.- Radiologist, SARS- Severe Acute Respiratory Syndrome, Seg.- Segmentation, VP- Viral Pneumonia, Sen.- Sensitivity, Spe.- Specificity.
| Ref. | Dataset | Pre-processing | Architecture | Code | Data | K-Fold | Performance reported |
Critical Observations | ||
|---|---|---|---|---|---|---|---|---|---|---|
| Acc. | Sen. | Spe. | ||||||||
| Abbas et al. [70] | Classes:3C/N/SARS 105/88/11 | Augmentation, contrast enhancement | VGG19 with class decomposition and composition | ✓ | ✓ | × | 97.4 | 98.2 | 96.3 | handled the class-imbalance problem using the proposed architecture |
| Abraham and Nair [72] | Classes:2C/CN 453/497 | Resized to different dimensions | Features extracted from multi-CNNs (Squeezenet, Darknet-53, MobilenetV2, Xception, Shufflenet); feature | × | ✓ | × | 91.2 | 98.5 | – | Correlation-based feature selection; bilinear interpolation for resizing; three RGB channels processing with single grayscale image being replicated to all the three channels |
| Classes2: C/CN 71/7 | selection and Bayesnet classifier | 97.4 | 98.6 | – | ||||||
| Afshar et al. [67] | Classes:2C/CN (The number of images are not mentioned) | Resized to 224 × 224 | Custom CNN | ✓ | ✓ | × | 98.3 | 80.0 | 98.6 | 4 convolutional layers and 3 Capsule layers; modified the loss function to handle the class-imbalance problem |
| Agrawal and Choudhary [65] | Classes:2C/N 1143/1 345 | Augmentation; resized to 224 × 224; normalization | Custom CNN | × | ✓ | ✓ | 99.2 | 99.2 | 99.2 | FocusNet [144] inspired CNN architecture having combination of multiple convolutional, 3 residual, and 2 squeeze-excitation blocks in between; evaluation by weighted |
| Classes:3C/N/P 1143/1 345/1345 | 95.2 | 95.2 | 95.6 | F1-score; handled the class-imbalance problem by oversampling technique such as SMOTE; validation done on two separate datasets | ||||||
| Al-Bawi et al. [88] | Classes:3C/N/VP 310/654/864 | None | VGG16 | ✓ | ✓ | × | 95.3 | 98.5 | 98.9 | Replaced last fully connected layer with 3 new convolutional layers |
| Apostol et al. [90] | Classes:3C/N/BP 224/504/700 | Resized to 200 × 266, black background of | VGG19 | × | ✓ | × | 93.5 | 92.8 | 98.7 | Fixed feature extractor with modification only in the last layer |
| Classes:3C/N/P 224/504/714 | 1:1.5 ratio was added to avoid distortion | 96.8 | 98.7 | 96.5 | ||||||
| Brunese et al. [84] | Classes:3C/Pulmonary disease/N 250/2 753/3 520 | Resized to 224 × 224 | VGG16, Grad CAM | × | ✓ | × | 97.0 | 91.0 | 96.0 | Fixed feature extractor with fine tuning of only last layers; added few layers like average pooling, flatten, dense, and dropout layers; two binary classifiers- training one for healthy and pulmonary, and the other for COVID and rest |
| Chowdhury et al. [5] | Classes:3C/N/VP 423/423/423 | Augmentation; resized to 224 × 224; normalization | DenseNet201, activation mapping | × | ✓ | × | 97.9 | 97.9 | 98.8 | Investigation of features of deep layers |
| Das et al. [57] | Classes:2C/CN 538/468 | Resized to 224 × 224, Normalization | Weighted averaging: DenseNet201 Resnet50V2 Inceptionv3 | ✓ | ✓ | ✓ | 91.6 | 95.1 | 91.7 | Development of a Graphical User Interface (GUI)-based application for public use |
| DeGrave et al. [53] | Classes:2C/CN 408/30 805 | Augmentation; resized to 224 × 224 | DenseNet121, interpretation by expected gradient & CycleGAN | ✓ | ✓ | × | – | – | – | Classifier training on 15 classes; comparison of results using AUC |
| Dhiman et al. [85] | Classes:2C/N 50/50 | Resized to 280 × 280 | ResNet101 | × | ✓ | ✓ | 100 | 100 | 98.9 | Analysis of segmented chest area; computational time analysis of multiple architectures; use of J48 decision tree classifier; fine-tuning using a multi-objective spotted hyena optimizer |
| Ezzat et al. [73] | Classes:2C/N 99/207 | Augmentation; resized to 180 × 180; normalization | DenseNet121; Grad-CAM | × | ✓ | × | 98.38 | 98.5 | 98.5 | Hyper-parameters optimization using gravitational search algorithm |
| Gupta et al. [74] | Classes:3C/N/P 361/365/362 | Augmentation, fuzzy color image enhancement and stacking it with original | Integrated stacked multiple CNNs (ResNet101, Xception, InceptionV3, MobileNet, and | × | ✓ | × | 99.1 | – | – | Both image enhancement and denoising |
| Classes:2C/NC 361/727 | Resized 224 × 224 × 3 | NASNet), Grad-CAM | 99.5 | – | – | |||||
| Hammoudi et al. [145] | Classes:4C/N/VP/BP 1493/1 493/1493/1 493 | Resized to 310 × 310 | DenseNet169 | × | ✓ | × | 99.1 | – | – | Measures were presented to associate survival chance with COVID-19 using risk factors like comorbidity, age, and infection rate indicator; Predicted patients' health status. |
| Heidari et al. [75] | Classes:3C/N/P 415/2 880/5 179 | Augmentation, histogram equalization, bilateral low-pass filtering, pseudo-color image generation | VGG16 | × | ✓ | × | 94.5 | 98.4 | 98.0 | handled class-imbalance problem by class weighting; removal of diaphragm regions; three channel processing; addition of 3 fully connected layers in the end |
| Hemdan et al. [78] | Classes:2C/N 25/25 | Resized to 224 × 224 | VGG19 | × | ✓ | × | 90.0 | – | – | One hot encoding on the labels of the dataset i.e. ‘1’ for COVID-19 and ‘0’ for all other images in the dataset |
| Ismael and Sengur [63] | Classes:2C/N 180/200 | Augmentation; resized to 224 × 224, grayscale image copied three times to form RGB image | ResNet50 with SVM | × | ✓ | × | 94.7 | 91.0 | 98.9 | No fine-tuning of ResNet50; analysis of eight well-known local texture descriptors of images |
| Islam et al. [60] | Classes:3C/N/P 1525/1 525/1 525 | Augmentation; resized to 224 × 224 | Custom CNN with LSTM, heatmaps | × | ✓ | ✓ | 99.4 | 99.1 | 98.9 | 12 convolutional layers with 1 fully connected layer and 1 LSTM layer |
| Jain et al. [76] | Classes:2C/CN 440/1 392 | Augmentation, resized to 640 × 640, normalization | ResNet50, ResNet101, Grad-CAM | × | ✓ | ✓ | 97.2 | – | – | Training of 2 two-class classification networks |
| Karthik et al. [26] | Classes:4C/N/BP/VP 558/10 434/2780/1 493 | Augmentation; resized to 256 × 256 | U-Net; custom CNN; interpretation analysis by class saliency maps, guided backpropagation, & Grad-CAM | × | ✓ | ✓ | 97.9 | 99.8 | – | Channel-shuffled dual-branched CNN comprising of three types of convolutions: (1) depth-wise separable convolution, (2) grouped convolution and (3) shuffled grouped convolution; augmentation done with distinctive filters learning paradigm |
| Keles et al. [98] | Classes:3C/N/VP 210/350/350 | Augmentation; resized to 224 × 224 | Custom CNN | × | ✓ | × | 97.6 | 98.7 | 98.7 | One input convolutional layer followed by 2 residual type blocks and 3 fully connected layers |
| Khan et al. [87] | Classes:4C/N/BP/VP 284/310/330/327 | Resized to 224 × 224, resolution of 72 dpi | XceptionNet | ✓ | ✓ | ✓ | 89.6 | 90.0 | 96.4 | handled the class-imbalance problem by undersampling |
| Classes:3C/N/P 284/310/657 | 95.0 | 95.0 | 97.5 | |||||||
| Classes:2C/N 284/310 | 99.0 | 98.3 | 98.6 | |||||||
| Classes:3C/N/P 157/500/500 | 90.2 | – | – | |||||||
| Loey et al. [89] | Classes:4C/N/BP/VP 69/79/79/79 | Augmentation; resized to 512 × 512; normalization | GoogleNet | × | ✓ | × | 80.6 | 80.6 | – | Image generation using Generative Adversarial Network (GAN) |
| Classes:3C/N/BP 69/79/79 | AlexNet | 85.2 | 85.2 | – | ||||||
| Classes: 2C/N 69/79 | AlexNet | 100 | 100 | – | ||||||
| Luz et al. [112] | Classes:3C/N/P 189/8 066/5 521 | Augmentation; normalization | EfficientNet; activation mapping | ✓ | ✓ | × | 93.9 | 96.8 | – | Hierarchical classification; use of swish activation; computational cost analysis by multiply-accumulate (MAC) operations |
| Mahmud et al. [99] | Classes:2C/N 305/305 | Resized to 256 × 256, 128 × 128, 64 × 64, and 32 × 32; normalization | Stacked Custom CNN, Grad-CAM | ✓ | ✓ | ✓ | 97.4 | 96.3 | 94.7 | Multiple residual and shifter units comprising of both depthwise dilated convolutions along with pointwise convolutions; training on multiple resized input images followed by predictions combining using meta learner |
| Classes:2C/VP 305/305 | 87.3 | 88.1 | 85.5 | |||||||
| Classes:2C/BP 305/305 | 94.7 | 93.5 | 93.3 | |||||||
| Classes:3C/VP/BP 305/305/305 | 89.6 | 88.5 | 87.6 | |||||||
| Classes:4C/N/VP/BP 305/305/305/305 | 90.2 | 90.8 | 89.1 | |||||||
| Madaan et al. [77] | Classes:2C/N 196/196 | Augmentation; resized to 224 × 224 | Custom CNN | × | ✓ | × | 98.4 | 98.5 | – | 5 convolutional layers along with a rectified linear unit as an activation function |
| Narayanan et al. [23] | Classes:2C/CN 2504/6 807 | Thresholding; grayscale, resized to 256 × 256; local contrast enhancement | U-Net; ResNet50; CAM | × | ✓ | ✓ | 99.3 | 91.0 | 99.0 | handled the class-imbalance problem by novel transfer-to-transfer learning; replaced last FC layer with two more fully connected layers |
| Nayak et al. [83] | Classes:2C/N 203/203 | Augmentation, normalization | ResNet34 | × | ✓ | × | 98.3 | – | – | Fine tuning of all the layers |
| Oh et al. [27] (VP and C were considered as one class) | Classes:4 N/BP/TB/VP 191/54/57/200 | Data type casting to float 32; histogram equalization; gamma correction; resized to 256 × 256 | FC-DenseNet103 for segmentation; patch-based CNN based on ResNet18; use of Grad-CAM | × | ✓ | × | 88.9 | 83.4 | 96.4 | Morphological analysis of lung area; evaluation of segmentation performance; peculiar pre-processing steps to remove heterogeneity across then dataset |
| Ozturk et al. [66] | Classes:2C/N 127/500 | Resized to 256 × 256 | Modified Darknet-19 | ✓ | ✓ | ✓ | 98.1 | 95.1 | 95.3 | Multiple Darknet layers having one convolutional layer followed |
| Classes:3C/N/P 127/500/500 | 87.0 | 85.4 | 92.2 | by batch normalization and leaky ReLU operations | ||||||
| Panwar et al. [64] | Classes:2C/N 142/142 | Augmentation; resized to 224 × 224 | VGG16 | × | ✓ | × | 88.1 | 97.6 | 78.6 | Utilized first 18 Imagenet pre-trained VGG16 layers and added 5 new different layers (average pooling, flatten, dense, dropout and dense) on the top |
| Pereira et al. [79] | Classes:7 N/C/SARS/MERS/Pnemocystic/Streptococcus/Varicella 1000/90/11/10/11/12/10 | None | Fusion of texture-based features and InceptionV3 features; classification using late fusion of multiple standard classifiers | ✓ | ✓ | × | 95.3 | – | – | handled the class-imbalance problem by re-sampling; multiclass and hierarchical classification |
| Pham et al. [61] | Classes:2C/N 403/721 | Resized to 227 × 227 | SqueezeNet | ✓ | ✓ | × | 99.8 | 100 | 99.8 | Features visualization of different layers |
| Classes:2C/N 438/438 | 99.7 | 99.5 | 99.8 | |||||||
| Rahimzadeh and Attar [80] | Classes:3C/N/P 180/8 851/6 054 | Resized to 300 × 300, augmentation | XceptionNet concatenated with ResNet50V2 | ✓ | ✓ | ✓ | 91.4 | 87.3 | 93.9 | handled the class-imbalance problem by training multiple times on resampled data |
| Sakib et al. [71] | Classes:3C/N/P 209/27 228/5794 | Augmentation using GANs | Custom CNN | × | ✓ | × | 93.9 | – | – | Analysis of different optimization algorithms; 5 convolutional layers along with exponential linear unit as an activation function |
| Sitaula et al. [132] | Classes:5C/N/BP/VP/NF (exact segregation is not given) | Resized to 150 × 150 | VGG16 | ✓ | ✓ | × | 79.6 | 89.0 | 92.0 | Leveraged both attention and convolution modules in the 4th pooling layer of VGG-16 for identifying deteriorating lung regions in both local and global levels of CXR images |
| Tabik et al. [91] | Classes:2 N/C 426/426 | Class-inherent transformation method using GANs | U-Net, ResNet50, Grad-CAM | × | ✓ | ✓ | 76.2 | 72.6 | 79.8 | Quantified COVID-19 in terms of severity levels so to build triage systems; Replaced last layer; fine-tuned all the layers; use of class-inherent transformation network to increase discrimination capacity; fusion of twin CNNs |
| Togacar et al. [51] | Classes:3C/N/P 295/65/98 | Resized to 224 × 224; Data restructured and stacked with the Fuzzy Color technique | Feature extraction using MobileNetV2 and SqueezeNet; processed using the social mimic optimization method; classified using SVM | ✓ | ✓ | ✓ | 98.2 | 97.0 | 99.2 | Image quality improvement using fuzzy technique |
| Toraman et al. [68] | Classes:3C/N/P 1050/1050/1050 | Augmentation; resized to 128 × 128 | Custom CNN | × | ✓ | ✓ | 84.2 | 84.2 | 91.8 | 4 convolutional layers and 1 primary capsule layer |
| Classes:2C/N 1050/1 050 | 97.2 | 97.4 | 97.0 | |||||||
| Ucar et al. [69] | Classes:3C/N/P 66/1 349/3 895 | Augmentation; normalization; resized to 227 × 227 | Bayes-SqueezeNet; activation mapping | × | ✓ | × | 98.3 | – | 99.1 | Handled the class-imbalance problem by multi scale offline augmentation; evaluation of proposed method using multiple metrics such as correctness, completeness and Matthew correlation coefficient; computational time analysis |
| Wang et al. [126] | Classes:3C/N/P | Augmentation; image cropping; resized to 480 × 480 | Custom CNN; interpretation by GSInquire [146] | ✓ | ✓ | × | 93.3 | 91.0 | 98.9 | Multiple projection-expansion-projection-extension blocks; different filter kernel sizes ranging from 7 × 7 to 1 × 1 |
| Wang et al. [28] | Classes:3C/N/CAP 225/1 334/2 024 | Augmentation; resized to 224 × 224 | VGG based Segmentation; ResNet with feature pyramid network | × | ✓ | × | 93.7 | 90.9 | 92.6 | Handled the class-imbalance with multi-focal loss function; residual attention network for localizing infected pulmonary region |