. 2021 May 28;69:102814. doi: 10.1016/j.bspc.2021.102814

Table 1.

Deep CNN approaches with their methods/models, data type used and performance metrics for COVID-19.

Method	Data Type	3 class (**)	2 class (*)	Ac%	Pr%	Sn%	Sp%	Architectural properties	Dataset	Ref.
COVID-Net	Chest X-Ray	$✓$	$✓$	93.30	92.4	91.0	NA	Tailored deep CNN have been used. Heterogeneous mix of convolution layers with diversity of kernel sizes (7x7 to 1x1) and grouping configurations are used. 1x1 convolutions are used for either input features to lower dimension or for expanding features to higher dimension. 3x3 are adopted for reducing computational complexity.	COVIDx dataset has been used which contain 13,975 CXR images across 13,870 patients.	[174]
Faster- RCNN	Chest X-Ray	$\times$	$✓$	97.36	99.29	97.65	95.48	VGG-16 network based Faster Regional CNN have been used. 16 weight layers are used with stacked 3x3 convolutional layers. Max. pooling is used. 2 FC layers with 4096 units are used with Softmax classifier.	183 COVID-19 +ve X-ray images and 13,617 Non-COVID X-ray images of chest were used.	[16]
VGG-19	Chest X-Ray	$✓$	$✓$	98.0* 93.0**	NA	92.0	98.0	Adopted transfer learning with CNN. Stacked 3x3 convolutional layers are used. Max. pooling is used for dimensional reduction. Two fully connected layers each with 4096 nodes are used with Softmax classifier. 19 weight layers are used.	Total 1427 X-ray images including 224 COVID-19 +ve case images were used.	[176]
Deep CNN ResNet-50	Chest X-Ray	$✓$ ***	$✓$ ***	98.0	NA	96.0	100	Adopted 3 variants of CNN models (InceptionV3, ResNet50, InceptionResnetV2). ResNet-50 has 48 convolution layers with 1 Max. pool and 1 Avg. pool layer. It over 23 million trainable parameters. Moreover, 50 weight layers are used.	Data used include 50 open access COVID-19 X-ray images from Joseph Cohen and 50 typical images from kaggle repository.	[170]
COVIDX-Net	Chest X-Ray	$\times$	$✓$	90.0	100	100	NA	COVIDx-Net adopts 7 CNN models. VGG9, DenseNet201, ResNetV2, InceptionV3, InceptionResNetV2, Xception, and MobileNetV2	Data used contain 50 chest X-ray images (25 covid +ve, 25 healthy)	[181]
DRE-Net	Chest CT scan	$\times$	$✓$	86.0	80.0	96.0	NA	Adopted ResNet50. ResNet-50 has 48 convolution layers with 1 Max. pool and 1 Avg. pool layer. It over 23 million trainable parameters. Moreover, 50 weight layers are used.	CT images of chest (777 covid +ve, 708 healthy).	[162]
M-Inception	Chest CT scan	$\times$	$✓$	82.90	NA	81.0	84.0	Modified Inception is used. Use of 7x7 convolution is adopted. 48 layers of deep CNN has been utilized.	Total 453 images (including 195 COVID-19 +ve and 258 Non-COVID-19 images.) were used.	[163]
ResNet	Chest CT scan	$\times$	$✓$	86.7	81.03	86.7	NA	3D CNN model with location-attention mechanism has been adopted. 2 convolutional layers interspersed with 2 max pooling layers followed by 2 fully connected layers are used.	Total 618 CT samlpes containing 219 images with COVID-19, 175 normal CT images, and 224 images containing viral pneumonia were used.	[164]
Dark CovidNet	Chest X-Ray	$✓$	$✓$	98.08* 87.02**	98.03* 89.96**	95.13* 858.35**	95.3* 92.18**	Based on DarkNet model. DarkNet is open source neural network framework using C and CUDA. YOLO is adopted for better detection. YOLO is 106 layer fully convolutional architecture with consecutive 3x3 and 1x1 convolutions.	Totally 1125 images containing 500 pneumonia, 125 COVID-19, and 500 with no findings were used.	[13]
COVNet	Chest CT scan	$✓$	$\times$	96.0	NA	90.0	96.0	ResNet50 is used as backbone. ResNet-50 has 48 convolution layers with 1 Max. pool and 1 Avg. pool layer. It over 23 million trainable parameters. Moreover, 50 weight layers are used.	4352 chest CT scans from 3322 patients are collected containing (1292 covid +ve, 1735 CAP, and 1325 Non-pneumonia abnormalities)	[168]
Insta CovNet-19	Chest X-Ray	$✓$	$✓$	99.53* 99.08**	100* 99.0**	99.0* 99.0**	NA	Integrated stacked deep CNN utilizing pre-trained models like Xception, ResNet101, MobileNet, InceptionV3, and NasNet. Xception is 71 layers deep network and uses 36 (3x3) convolutional layers for feature extraction with 16 input and 32 output channels. It has 4608 parameters. MobileNet has 28 layers and 4.2M parameters.	Combined dataset [179], [166] is used which contain 361 covid images, 1341 pneumonia and 1345 normal images.	[171]
COVI Diagnosis - Net	Chest X-Ray	$✓$	$\times$	98.26	98.26	98.26	99.13	Based on deep Bayes-SqueezeNet. A squeeze convolutional layer has only 1 1 filters. Max. pooling, RELU activation and Dropout techniques are adopted in SqueezeNet. FC layer is absent.	COVIDx dataset [174] containing 5949 chest radiography images from 2839 patients (1583 normal, 4290 pneumonia, 76 covid infected.	[172]
CVDNet	Chest X-Ray	$✓$	$\times$	96.69	96.72	96.84	NA	Adopts residual Neural Network. ResNets can have variable sizes, depending on how big each of the layers of the model are, and how many layers it has e.g. 34, 50, 101, etc. $3 \times 3$ convolutions are normally used with kernel size of 7, and a feature map size of 64. Max. pooling is adopted.	Dataset [179], [166] containing 219 COVID-19, 1345 with viral pneumonia and 1341 healthy images.	[163]
CoroNet	Chest X-Ray	$✓$	$✓$	99.0* 95.0 89.6*	98.3* 95.0 90.0*	99.3* 96.9 89.92*	98.6* 97.5 96.4*	Adopted Xception architecture. Xception is 71 layers deep network and uses 36 (3x3) convolutional layers for feature extraction with 16 input and 32 output channels. It has 4608 parameters.	1251 chest X-ray images containing 310 normal, 327 pneumonia viral, 330 pneumonia becterial, and 284 covid-19)	[173]
COVID-CAPS	Chest X-Ray	$\times$	$✓$	95.7	na	90.0	95.8	Based on Capsule Networks. Encoders and decoders constitute 6 layers in capsule network.	COVIDx dataset [174] has been used.	[175]
mALexNet + BiLSTM	Chest X-Ray	$\times$	$✓$	98.70	98.77	98.76	99.33	Adopts residual Neural Network. ResNets can have variable sizes, depending on how big each of the layers of the model are, and how many layers it has e.g. 34, 50, 101, etc. $3 \times 3$ convolutions are normally used with kernel size of 7, and a feature map size of 64. Max. pooling is adopted.	Dataset [179], [166] containing 219 COVID-19, 1345 with viral pneumonia and 1341 healthy images.	[10]
COVID - ResNet	Chest X-Ray	$✓$ ***	$✓$ ***	96.23	100	100	100	Based on ResNet50 model. ResNet-50 has 48 convolution layers with 1 Max. pool and 1 Avg. pool layer. It over 23 million trainable parameters. Moreover, 50 weight layers are used.	COVIDx dataset [174] has been used.	[180]