Computers in Biology and Medicine 143 (2022) 105233

Contents lists available at ScienceDirect

Computers in Biology and Medicine

journal homepage: www.elsevier.com/locate/compbiomed

A review of deep learning-based detection methods for COVID-19

Nandhini Subramanian *, Omar Elharrouss, Somaya Al-Maadeed, Muhammed Chowdhury
Qatar University College of Engineering, Computer Science and Engineering, Qatar

A R T I C L E I N F O A B S T R A C T

Keywords: COVID-19 is a fast-spreading pandemic, and early detection is crucial for stopping the spread of infection. Lung
COVID-19 detection images are used in the detection of coronavirus infection. Chest X-ray (CXR) and computed tomography (CT)
DL-Based COVID-19 detection images are available for the detection of COVID-19. Deep learning methods have been proven efficient and better
Lung image classification
performing in many computer vision and medical imaging applications. In the rise of the COVID pandemic,
Coronavirus pandemic
Medical image processing
researchers are using deep learning methods to detect coronavirus infection in lung images. In this paper, the
currently available deep learning methods that are used to detect coronavirus infection in lung images are
surveyed. The available methodologies, public datasets, datasets that are used by each method and evaluation
metrics are summarized in this paper to help future researchers. The evaluation metrics that are used by the
methods are comprehensively compared.

1. Introduction challenges, vaccine hesitancy, and vaccine complacency. A vaccine is a

prevention measure rather than a cure [4]. Even with the availability of
The World Health Organization (WHO) declared the spread of the the vaccine, early detection of the coronavirus is important, as it can
coronavirus infection a pandemic in March 2020, which is called the facilitate tracing of the people who were in contact directly and indi­
coronavirus pandemic or COVID-19 pandemic. The coronavirus rectly. By tracing these people, further spread of the pandemic can be
pandemic is caused by severe acute respiratory syndrome coronavirus 2 avoided. COVID-19 infection manifests as lung infection, and computed
(SARS CoV 2). The outbreak originally started in Wuhan, China, and tomography (CT) and chest X-ray (CXR) images are primarily used in the
later spread to every country in the world [1]. The coronavirus spreads detection of lung infection of any type [5].
through respiratory droplets of the infected person that are produced Along with doctors and clinical personnel, researchers and technol­
through cough or sneeze. These droplets can further contaminate the ogists are focusing their efforts on early detection of coronavirus in­
surfaces increasing the spread. Coronavirus-infected persons may suffer fections. According to PubMed [6], 755 academic articles were
from mild to severe respiratory illness and may require ventilation published with the search term “coronavirus” in 2019, and this number
support [2]. Older people and people with chronological disorders are rose to 1245 in the first 80 days of 2020. Artificial intelligence and deep
easily prone to coronavirus infection. Thus, many governments have learning methods are the most commonly used methods by researchers
closed their borders and locked down people to break the cycle and for the detection of coronavirus infection from CT and CXR images. Deep
prevent the spread of the pandemic [3]. learning methods have shown significant performance in many research
With the sequencing of ribonucleic acid (RNA) from the coronavirus, applications, such as computer vision [7], object tracking [8], gesture
many vaccines are being developed worldwide. The developed vaccines recognition [9], face recognition [10], and steganography [11–13].
use both traditional and next-generation technology with six vaccine Deep learning methods are widely used because of their improved per­
platforms, namely, live attenuated virus, inactivated virus, protein or formance compared to traditional methods. In contrast to traditional
subunit, viral vector-based, messenger RNA (mRNA), and deoxy­ methods and machine learning methods, the features need not be
ribonucleic acid (DNA). Although vaccines can reduce the rapid spread hand-picked. By changing the parameters and configurations of the deep
and facilitate the development of immunity via the production of suit­ learning convolutional neural network (CNN) architecture, a model can
able antibodies, the efficacy of the vaccines is still 95%. Many issues are be trained to learn the best possible features for the dataset in use. Re­
encountered in administering the vaccine, such as supply chain logistical searchers have used deep learning methods to explore the field of

* Corresponding author.
E-mail addresses: [email protected] (N. Subramanian), [email protected] (O. Elharrouss), [email protected] (S. Al-Maadeed), mchowdhury@qu.
edu.qa (M. Chowdhury).

https://doi.org/10.1016/j.compbiomed.2022.105233
Received 6 June 2021; Received in revised form 10 January 2022; Accepted 10 January 2022
Available online 29 January 2022
0010-4825/© 2022 Qatar University. Published by Elsevier Ltd. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/by/4.0/).
N. Subramanian et al. Computers in Biology and Medicine 143 (2022) 105233

medical imaging even before the coronavirus pandemic. With the recent 2.2. Transfer learning and fine-tuning
pandemic, the use of deep learning methods for the detection of coro­
navirus infection from images has increased tremendously. After designing, creating and building a deep learning model, the
A detailed survey of the available deep learning approaches for the number of epochs is set to start training. During training, random
detection of coronavirus infection from images such as CT scans or CXR weights are initialized, which will be refined during each epoch to make
images is conducted in this paper. Although other surveys are available the result closer to the classification score. However, in transfer
in the literature, most of them cover a wider scope. For example, Ulhaq learning, instead of using random weight values, the model can be
et al. [14] surveyed all methods that address coronaviruses, such as initialized with weight values from pretrained models. Transfer learning
medical image processing, data science methods for pandemic performs best when there is a limited availability of training data. When
modeling, AI and the Internet of things (IoT), AI for text mining and performing transfer learning, the last layer of the pretrained model ar­
natural language processing (NLP), and AI in computational biology and chitecture is replaced with a fully connected layer with the same number
medicine. This provides an overall view of what is happening in the of classes as the new dataset. The architecture is retrained to use the
research world. A survey on the application of computer vision methods model for the new dataset [24].
for COVID-19 [15] described the segmentation of lung images. This Another method, namely, fine-tuning, is also used when the dataset
paper aims to exclusively describe coronavirus detection methods using is small. Similar to transfer learning, the last layer of the architecture is
deep learning methods. In the hope of helping researchers develop replaced and redefined. The only difference is that in transfer learning,
better coronavirus detection methods, this paper summarizes all the all the layers are retrained, while in fine-tuning, some layers can be
methods that have been reported in the literature. Along with the redefined and retrained according to the application [25]. One major
methods, the used datasets, commonly used metrics for evaluation and disadvantage of these methods is that the size of the input image cannot
comparison are discussed and future direction are elaborated in this be changed. Therefore, if the pretrained model uses a smaller image
paper. dimension and transfer learning has to be conducted on a dataset with a
larger image dimension, resizing the image is compulsory. Resizing a
2. Background large image to a smaller image can affect the performance of the model
in some cases. Careful consideration must be taken when transfer
Before discussing the details of the available methods for coronavirus learning and fine-tuning are implemented.
infection detection, it is essential to have a working knowledge of deep
convolutional neural networks and popular CNN architectures. In this
section, a brief overview of CNN architectures and main points on 2.3. Available architectural families
available CNN architectures are presented.
Several available architectures generalize well irrespective of the
2.1. Convolutional neural networks dataset or application. Various popular architectures, such as AlexNet,
VGG, Inception, ResNet, DenseNet, MobileNet, and Xception, are sum­
Convolutional neural networks, specifically artificial neural net­ marized in this section.
works, are a branch of deep learning methods that are inspired by the AlexNet is a simple five-layer convolutional neural network. There
natural visual perception mechanism of living organisms [16]. CNNs are are two variants of the VGG network – VGG16 and VGG19 [26]. The
nothing but stacked multilayered neural networks. There are three VGG architecture was originally proposed for image recognition appli­
major categories of layers, namely, convolutional layers, pooling layers cations. In VGG16 and VGG19, 16 and 19 wt layers are used with a
and fully connected layers. The first layer of any CNN model is an input smaller convolutional filter size of 3 × 3. The network won first and
layer, where the width, height and depth of the input image are specified second places in the ILSVR (ImageNet) competition [27] in 2014. The
as the input parameters. Immediately after the input layer, convolu­ size of the input image is fixed to 224 × 224. The model is trained on the
tional layers are defined with the number of filters, filter window size, ImageNet dataset, which contains millions of images [28].
stride, padding and activation as the parameters. Convolutional layers In contrast to CNN architectures, in which the layers are stacked, a
are used to extract meaningful feature maps for the input location by new architecture with an inception block is introduced in InceptionNet
calculating the weighted sum [17,18]. Then, each feature map is passed [29]. Several variants are available in the inception family. The incep­
through an activation function, and bias is added to form the output. tion network is also used for image classification and localization and
Usually, rectilinear unit (ReLU) activation is used as the activation participated in the ILSVR (ImageNet) competition [27] in 2014. Instead
function [19]. of increasing the depth of the model by adding additional layers, the
Pooling layers are used to reduce the size of the output from the authors apply various filter sizes to the input image simultaneously in
convolutional layers. As the model increases in size with an increasing the inception block. This leads to the growth of the model width. All the
number of filters in the convolutional layer, the output dimensionality outputs of the inception block are concatenated and fed to the next
also increases exponentially, which makes it hard for computers to inception block. Available versions include InceptionV1 (GoogLeNet)
handle. Pooling layers are added to reduce the dimensions for easy [29], InceptionV2 and InceptionV3 [18], InceptionV4 and Inception­
computation and sometimes to suppress noise. The pooling layer can be ResNet [30]. The input image size that is accepted by the model is 224 ×
a max pooling, average pooling, global average pooling, or spatial 224.
pooling layer. The most commonly used pooling layer is a max pooling ResNet [31] is also used in image classification methods and was the
layer [20]. The output is flattened to form a single-array feature vector, winner of the ILSVRC 2015 [27]. The ResNet family uses the residual
which is fed to a fully connected layer. Finally, a classification layer is block, which is a network-in-network in their architecture. Five steps
defined with activation functions such as sigmoid, softmax and tanh with convolutional and identity blocks are used to define the network.
functions [21]. The number of classes is specified in this layer, and the Similar to the VGG family, the input image size is 224 × 224. Many
extracted features are aggregated into class scores. variations are available. Inception-ResNet [30] is a hybrid architecture
Batch normalization layers are applied after the input layer or after that combines the inception and residual blocks. The input image size
the activation layers to standardize the learning process and reduce the for InceptionResNet is 229 × 229.
training time [22]. Another important parameter is the loss function, The DenseNet architecture [32] is a variation of the ResNet archi­
which summarizes the error in the predictions during training and tecture. Similar to the ResNet family, a residual identity block is used to
validation. The loss is backpropagated to the CNN model after each build the architecture, except concatenation is conducted in place of
epoch to enhance the learning process [23]. summation. Traditional CNN models have L connections for L layers,

2
N. Subramanian et al. Computers in Biology and Medicine 143 (2022) 105233

whereas the DenseNet model has L(L+1) direct connections. Each layer is SqueezeNet is performed by Ref. [46]. Two-class and three-class clas­
2
connected into every other layer in a feed-forward fashion. The feature sification with and without data augmentation is performed with five­
maps of all the previous layers are used as input to the current layer, and fold cross-validation and stochastic gradient descent (SGD)
the feature map of the current layer is fed to all the other layers. The size optimization. Fig. 2 illustrates the working principle of [46] stepwise.
of the accepted input image is 224 × 224. Similar to Ref. [46], binary and multiclass classification on
MobileNets are compact architectures with depthwise separable NASNet-Large, DenseNet169, InceptionV3, ResNet18, and Inception
convolutional layers that can be used in mobile phones and embedded ResNet V2 are implemented by Punn et al. [47]. However [47], uses a
systems [33]. Usually, 2D convolutional layers are used, but in depth­ weighted class loss function and random oversampling methods to
wise separable convnets, two 1D convolutional layers are used. Doing so overcome the disproportionate rates in the classes. The class with the
has helped reduce the number of parameters and, hence, decrease the ”COVID” label is given higher weight, since it is of higher significance
computation and training times and memory usage. There are 54 layers, than other classes, using the weighted class loss function. In the random
and the input image size is 224 × 224. oversampling method, the classes are balanced by increasing the num­
Xception [34] architectures are similar to the Inception family, ber of samples in the minority class by data augmentation. For denois­
where inception blocks with depthwise separable convolutional layers ing, an image mask is created using binary thresholding and subtracted
are used. The input image size is 229 × 229, and the number of layers is from the original image. Fine-tuning is performed by keeping non­
71. trainable layers as the base model and adding four trainable convolu­
tional layers, one fully connected layer and one classification layer.
Transfer learning is also used by Wang et al. [48], but instead of the
3. Summary of the research methods
whole image, region of interests (RoIs)/image patches are provided as
input. A total of 195 COVID-positive and 258 COVID-negative image
Since COVID-19 is a novel pandemic, only a few datasets with a
patches are used for training. These image patches are input into a
limited number of samples are publicly available. The best strategy that
pretrained network for feature extraction, followed by a fully connected
can be followed with the limited availability of data is either transfer
classification layer for classification.
learning or fine-tuning (Section 2.2). Although new CNN architectures
Generative adversarial networks (GANs) are used extensively for
can be constructed, to improve the performance, a wider range of images
image reconstruction [49]. Data augmentation is one application of
under each class is required. According to this study, the majority of the
GANs [50]. Since the dataset is small, more data are obtained using a
papers use transfer learning methods, a few rely on fine-tuning, and only
GAN for data augmentation, and the augmented data are split into
a handful propose a novel CNN architecture with comparable perfor­
training and testing sets to train a deep CNN model for binary classifi­
mance to transfer learning-based methods. The majority of the works
cation [51]. Three phases are used. First, in the preprocessing phase, the
use transfer learning from models that are pretrained on the ImageNet
GAN is used for data augmentation. Second, transfer learning on Alex­
dataset. Additionally, the input image size to the architecture is either
Net, SqueezeNet, GoogleNet, and ResNet18 is performed to train the
224 × 224 or 229 × 229, but the dataset that is used to train and test the
model. Finally, in the testing phase, the trained model is evaluated.
model contains images of various sizes. A simple preprocessing step is
Along with fine-tuning on the top layers of the CNN, VGG16, VGG19,
used to resize the images in the dataset to fit into the shape of the input
DenseNet201, Inception_ResNet_V2, Inception_V3, Xception, Resnet50,
layer of the network. In this section, first, transfer learning and fine-
and MobileNet_V2 architectures, a comparative study is conducted [52].
tuning-based methods and the CNN architectures that are used will be
Three convolutional layers with a filter size of 3 × 3, two max-pooling
specified. Then, methods with novel CNN architectures will be
layers with a filter size of 2 × 2, a fully connected layer and, finally, a
described. Finally, methods that do not belong to these categories will be
classification layer with a sigmoid classifier are proposed. Intensity
described in detail. Fig. 1 presents an overall summary of all the methods
normalization [53] and contrast limited adaptive histogram equaliza­
that are reviewed in this paper.
tion (CLAHE) [54] are performed on the images during preprocessing.
First, a dataset is synthesized using a fuzzy color technique. Then,
3.1. Transfer learning and fine-tuning approaches another dataset is created by combining the original and fuzzy color
images using the stacking technique. Transfer learning and fine-tuning
Transfer learning is the go-to method for most of the papers. Pre­ are performed on the created dataset [55]. Transfer learning on a
trained models that are trained on the ImageNet database are used to combination of chest X-ray and CT scan images using the VGG19-CNN,
perform transfer learning. Although the method is the same, different ResNet152 V2, ResNet152 V2 + gated recurrent unit (GRU), and
architectures are used in the works [35]. Even if the architectural family ResNet152 V2 + bidirectional GRU (Bi-GRU) architectures for multi­
is the same, different variants are used. Cross-validation is another class classification is performed by Ibrahim et al. [56]. Transfer learning
technique that is used in some of the methods. In addition, methods with on 3D CT scans using ResNet architectures is also conducted [57]. A
new CNN models are considered, which also utilize the benefits of machine-learning algorithm-based method is also designed and evalu­
transfer learning when the dataset is very small. ated for coronavirus detection [58,59].
A comparative study of the available deep learning architectures,
namely, MobileNet V2, Inception, Xception, Inception ResNet V2 and 3.2. Novel architectures
VGG19, that use the transfer learning method is performed by Aposto­
lopoulos et al. [35]. Three models - ResNet50, InceptionV3 and COVID-Net [60] utilizes a new CNN architecture for detecting COVID
Inception-ResNetV2 - are also utilized [36]. Transfer learning on from CXR images, and an open-source COVID dataset, namely, COVIDx,1
InceptionV3 [37] and AlexNet [38] along with data augmentation is is introduced. COVID-Net can classify CXR images into one of three
another variation. ResNet18, ResNet50, SqueezeNet [39], and classes. The architecture is based on lightweight residual
DenseNet-121 [40] are used for transfer learning along with data projection-expansion projection extension (PEPX) design patterns with
augmentation methods [41]. Transfer learning using VGG19, Dense­ two stages of projections, expansions, a depthwise representation and an
Net201, InceptionV3, ResNet152, InceptionResNetV2, Xception, and extension. The authors perform transfer learning by training the CNN
MobileNetV2 is conducted by Ref. [42]. Transfer learning on ImageNet architecture initially on the ImageNet dataset and subsequently on the
with VGG16 and ResNet50 [43] by replacing the last layer of both COVIDx dataset.
VGG16 and ResNet50 with one global average pooling layer [44] and
two fully connected layers is used in Ref. [45].
Transfer learning on AlexNet, ResNet18, DenseNet201 and 1
https://github.com/lindawangg/COVID-Net.

3
N. Subramanian et al. Computers in Biology and Medicine 143 (2022) 105233

Fig. 1. Overall workflow summary of all the methods. The first step is the acquisition of the data, and the imaging format can be chest X-ray (CXR) or CT scan. The
second step is preprocessing, such as image resizing and data augmentation. Then, the preprocessed data are trained using one of the three methods. The trained
model is used for classification and evaluation.

Fig. 2. Stepwise diagrammatic representation of transfer learning by Chowdhury et al. [46]. The first step is the acquisition of the patients’ data from an X-ray
imaging machine. Both two-class classification and three-class classification are performed. Second, in the image resizing (preprocessing) step, the input layer of the
deep learning method is fit. Data augmentation is performed in one of the experiments. Then, transfer learning is performed on various deep learning architectures.
Finally, the trained model is saved, and classification is performed.

A model with three parts, namely, a backbone, a classification head DeTraC, which consists of three phases: feature extraction, decomposi­
and an anomaly detection head, is proposed by Zhang et al. [61]. The tion and class composition. Using the backbone architecture, features
pretrained backbone architecture on ImageNet is used to extract from images are obtained. Then, training using the SGD optimizer is
high-level features from X-ray images, and these features are fed to the performed, followed by class composition for classification.
classification and anomaly detection heads to produce a score. A cu­ COVIDLite is a novel architecture that uses the depthwise separable
mulative score for every ’l’ predictions is also used. convolutional neural network (DSCNN) to classify CXR images for
COVID-CAPS is a capsule network-based framework for detecting the coronavirus detection [64]. A preprocessing step (CLAHE) is used to
presence of COVID infection from CXR and CT scan images [62]. One improve the visibility and enhance the white balance. White balancing is
advantage of using a capsule network is that it can perform well even performed to enhance the color fidelity of the images. Fast COVID-19
when data are scarce. Transfer learning is also used in this framework. detector (FCOD) is another variant of the depthwise separable con­
However, this is contrary to other methods of transfer learning on a volutional neural network, which is based on the inception architecture
model that is pretrained with X-ray images from a publicly available [65]. Using depthwise separable convolutional layers in place of the
dataset 2. This has advantages over the other methods for transfer normal convolutional layer decreases the computational complexity and
learning on the ImageNet dataset. computation time. Similar to Ref. [65], depthwise separable convolu­
A novel CNN model is proposed by Abbas et al. [63], namely, tional layers are used in the XceptionNet architecture by Singh et al.

4
N. Subramanian et al. Computers in Biology and Medicine 143 (2022) 105233

[66]. A hierarchical classification method in place of flat classification is

A novel CNN with one convolutional block with a 16-filter con­ another proposed variation [81]. Hierarchical classification considers
volutional layer, batch normalization and ReLU activation and two fully the relationships between classes, conducts local classification and
connected layers with softmax classification is proposed by Maghdid trains models to perform the classification. Since the dataset is small
et al. [67]. The pretrained Alexnet on the ImageNet dataset is compared even after customization, to avoid underfitting or overfitting of the
with the proposed model. A set of tailored CNN models that are based on model, the available data are expanded using data augmentation tech­
established architectures is proposed by Ref. [68]. Each detected image niques. The EfficentNet [82] architecture family is used as the base
can belong to one of three classes, namely, normal, viral pneumonia and model for the classification, which is extended by adding batch
bacterial pneumonia. Additionally, an estimator for the infection rate is normalization and dropout, followed by three fully connected layers and
provided from the predictions. classification using softmax. Additionally, instead of training from
A custom CNN model that accepts concatenated features from two scratch, transfer learning on ImageNet dataset weights is carried out.
models (Xception and ResNet50V2) and passes them through a con­ ResNet50 is used as the base model for classifying the image into
volutional layer and a classification layer is proposed by Ref. [69]. three classes: normal, bacterial pneumonia and viral pneumonia [83]. If
Similarly, deep features are extracted from MobileNet as the base model, the prediction is viral, the image is input into DenseNet169 to further
and they are input into a global pooling layer and a fully connected classify it as COVID or not. This is similar to hierarchical classification
layer. Then, the feature vector is input into the classifier for classifica­ except that a single model is used for the full overflow, in contrast to
tion by Ref. [70]. Three types of techniques are tested, namely, Ref. [81]. Global average pooling (GAP) and SE-Structure are used to
fine-tuning, transfer learning and training from scratch. As in Refs. [69, increase the performance of the model. Contrast limited adaptive his­
70], a deep convolutional neural network architecture, namely, CoroNet togram equalization (CLAHE) and the MoEx structure that is formed
[71], is used to classify X-rays into four classes: normal, bacterial from normalization are used for image enhancement to help increase the
pneumonia, viral pneumonia and COVID-19 positive. The architecture is accuracy. A gradient class activation map (Grad-CAM) is used for visu­
based on Xception as the base; however, a dropout and two fully con­ alization to help doctors [84]. U-Net is used to segment the lung in the
nected layers are used. The Darknet-19 [72] based architecture, which is image, which is also provided as input to the DenseNet model. A
used for general object detection, is called the DarkCovid net [73]. It workflow that is similar to Ref. [83] is proposed by Gozes et al. [85].
uses fewer layers than Darknet-19 with average pooling and softmax for First, lung segmentation using U-Net is performed to extract the ROIs.
classification, and transfer learning on the ImageNet dataset is The ROIs are provided as input for classification, and Grad-CAM is used
performed. for visualization.
A four-phase method for COVID-19 detection is implemented by A preprocessing step, which includes contrast and edge enhancement
Ozyurt [74]. The feature extraction technique is emphasized by using using histogram equalization (HGE), application of the Perona–Malik
techniques such as exemplar-based pyramid feature generation, ReliefF, filter (PMF), and elimination of noise by unsharp masking edge
and iterative principal component analysis (PCA) analysis. The final enhancement, is conducted before the detection of coronavirus infection
stage is classification using a deep neural network (DNN) and an arti­ [86]. This preprocessing can help the model learn and generalize better.
ficial neural network (ANN). CovXNet is a novel CNN architecture with An ensemble-based method is employed for detection by training the
depthwise convolutional layers [75]. Not only is this novel architecture VGG, ResNet, and DenseNet architectures. An ensemble of the best
trained from scratch but also different modifications, such as transfer model predictions is used to obtain the final prediction.
learning and fine-tuning, are designed to compare the performances of COVID-MobileXpert is a deep learning-based hardware-friendly
various methods. Both binary classification and multiclass classification model with a knowledge transfer and distillation framework [87].
are performed on chest X-rays by unique CNN architectures without DenseNet-121 is used by the Attending Physician (AP) and Resident
transfer learning by Karakanis et al. [76]. Fellow (RF) networks, and MobileNetv2, ShuffleNetV2 and SqueezeNet
are used by the Medical Student (MS) network. The MS network has
3.3. Other approaches been designed to facilitate the deployment of the model on devices.
Transfer learning is conducted on the AP and RF networks, and the RF
A pretrained model is used to extract the deep features of the images network is used to train the MS network through knowledge distillation.
of a prepared custom dataset [77]. Then, the extracted deep features are An ensemble method with three steps, namely, feature extraction
input into a linear support vector machine (SVM) and OneVsAll SVM using Alexnet, feature selection using trial and error and classification
classifier for classification. Eleven established model architectures that using the SVM algorithm, is performed. The results are compared with
are pretrained on the ImageNet dataset [28] are used to extract the deep those of other deep learning methods, and the proposed solution has
features: AlexNet, DenseNet201, GoogleNet, InceptionV3, ResNet18, higher overall accuracy [88]. A multitask method is proposed by Rah­
ResNet50, ResNet101, VGG16, VGG19, XceptionNet, and man et al. [89], along with a new dataset, for image enhancement,
InceptionResNetV2. segmentation, and classification.
A slightly different approach is applied by the authors for the clas­ Fig. 3 presents an overview summary of all the methods that are
sification of X-ray images [77]. Similar to Ref. [77]., features are currently available for this application.
extracted from three networks, namely, VGG-16, GoogleNet and
ResNet-50 [78], for the classification of CT images. The features are 4. Datasets
fused, and to reduce the redundancy of the features, the t-test method is
used to rank the features based on frequency. The final constructed The size of the data is the key factor for the performance of any deep
feature vector is input into a binary SVM classifier for classification. A learning model. However, since COVID-19 is a recent disease, only a
depthwise separable convolution neural network (DWS-CNN) is used to limited number of datasets are publicly available. There is a repository
extract the features from the patient’s X-ray images. The extracted fea­ of COVID-positive lung X-ray images that is constantly updated [90],
tures are input into a deep support vector machine (DSVM) for classi­ solely for classification purposes. It also contains metadata and anno­
fication. Data acquisition occurs through Internet of things tations of the lung segments. Additionally, this repository contains only
(IoT)-enabled devices. The raw data are passed through a Gaussian filter a limited number of non-COVID images. Another commonly used
before feature extraction and classification [79]. A pretrained VGG16 dataset in this context is from Kaggle.2 Dr. Paul Mooney created a lung
network is used, and the output is upsampled to a depthwise separable
convolutional network, which is followed by a shallow 3D CNN block
and spatial pyramid pooling for COVID-19 detection [80]. 2
www.kaggle.com.

5
N. Subramanian et al. Computers in Biology and Medicine 143 (2022) 105233

Fig. 3. Methods and approaches. The surveyed literature works are grouped into three categories, namely, transfer learning and fine-tuning, novel architectures, and
other approaches. Three branches are included in the figure, and the methods under each category are listed.

image dataset with 5,863 pediatric images under three classes (normal, consists of 133 COVID + images, which include Middle East respiratory
viral pneumonia and bacterial pneumonia).3 Apart from these, the RSNA syndrome (MERS), acute respiratory distress syndrome (ARDS), and
Pneumonia Detection Challenge dataset,4 SIRM datasets,5 Covid Chest SARS images and 133 COVID-images from Refs. [90,97]. Four datasets
X-ray dataset [92], and Chexpert dataset [93] are notable datasets that are used for experiments in Ref. [38], namely, [90,92,93,98]. The ex­
are used for COVID classification. Another important consideration is periments are conducted separately, and the classes and the images are
that some of the methods use binary classes (COVID+ and COVID-) combined. Three datasets, namely [90,92,99], are used in Ref. [81].
whereas others use more than two classes (normal, COVID, viral pneu­ Multiple classes other than pneumonia are used, which include thorax
monia and bacterial pneumonia) for classification. diseases with COVID + images from Refs. [90,99,100]; pneumonia im­
The COVIDx datasets that are introduced in COVID-Net [60] include ages from Ref. [101]; and other thoracic disease images from Ref. [92].
13,975 CXR images across 13,870 patient cases that have been selected To add diversity, data augmentation is conducted.
and combined from publicly available datasets. The dataset consists of A dataset with a total of 1300 images, namely, 310 normal, 330
images in three classes, namely, normal, non-COVID infection and bacterial pneumonia, 327 viral pneumonia and 284 COVID images, is
COVID infection. A detailed study and the steps for generating the used in Ref. [71]. COVID-positive images are obtained from Ref. [90],
dataset can be found in Ref. [94]. A custom COVID-Xray-5k dataset is and normal, bacterial and viral pneumonia images are obtained from
built with 2,031 training images and 3,040 test images [41]. This dataset Ref. [91].
is a combination of COVID + images from the COVID Chest X-ray dataset In [46], four datasets are combined. COVID images are collected
[92] and ChexPert [93].6 from Refs. [90,100,101] by the authors of [46]. Normal and viral
Two datasets are used in Ref. [35] to evaluate transfer learning on pneumonia images are collected from Ref. [91]. A two-class dataset is
various models. A combination of normal, COVID and bacterial pneu­ created.
monia images from various sources, such as [90], and [95], are com­ Images from 5 sources are combined in Ref. [67]. [73] uses normal
bined into one dataset with 504 normal images, 700 bacterial and pneumonia images from Ref. [92] and COVID images from
pneumonia images and 224 confirmed COVID images. However, to Ref. [90]. To obtain a balanced dataset, only 500 random images from
fine-tune and improve the performance of the models, another class, Ref. [92] for both classes are selected. Fivefold cross-validation is con­
namely, viral pneumonia, is added to create another dataset. Dataset 2 ducted with two experiments – binary and three-class classification..7
consists of 504 normal images, 224 confirmed COVID images, 400 [68] uses [91] with images from normal, bacterial pneumonia and
bacterial pneumonia infection images and 314 viral pneumonia infec­ viral pneumonia for training, and testing is performed on COVID images
tion images. A black background is added, and the images are rescaled to from Ref. [90]. It is assumed that any infections that are caused by
dimensions of 200 × 266. Even after all these efforts, the number of COVID-19 are due to viruses; thus, the model has to predict the
samples in the dataset is small, and the classes are not balanced with the COVID-positive images under the viral pneumonia class. Covid images
minimum number of images for confirmed COVID cases. from Ref. [90] and pneumonia, no findings and normal images from
A few images that represent each class from the most commonly used Ref. [99] are used to form a dataset in Ref. [47].
datasets, namely [90,96], are presented in Fig. 4 [90]. has COVID+ and [36] uses 50 COVID infection images from Ref. [90], 50 normal
COVID-images, while [96] has normal, bacterial pneumonia, and viral healthy images from Ref. [91], 100 COVID images from Ref. [90], and
pneumonia images. Table 1 summarizes in detail the most commonly 1431 pneumonia infection images from Ref. [92] in Ref. [61]. A total of
used publicly available datasets. 130 COVID-19 and 130 normal X-ray images from Refs. [90,96,97] are
Similarly [35,77], develop two datasets for model training and used in Ref. [37]. A dataset that combines [90,91] is used in Ref. [69]..8
testing. The first dataset consists of 25 COVID + images, excluding Eighty normal images from Refs. [102,103] and 116 images from
MERS, ARDS and SARS, and 25 COVID-images. The second dataset Ref. [90], along with data augmentation, are used in Ref. [63]. In
Ref. [51], 624 images with two classes [101] are used [101]. is used in
Ref. [52] for binary classification using data augmentation [42]. uses
3
[90,104] for COVID and normal images, respectively. In Ref. [42], 25
https://www.kaggle.com/paultimothymooney/chest-xray-pneumonia/
version/2.
4
www.kaggle.com/c/rsna-pneumonia-detection-challenge.
5 7
https://www.sirm.org/category/senza-categoria/covid-19/. https://github.com/muhammedtalo/COVID-19.
6 8
https://github.com/shervinmin/DeepCovid. https://github.com/mr7495/covid19.

6
N. Subramanian et al. Computers in Biology and Medicine 143 (2022) 105233

Fig. 4. (a). Example images from two classes, namely, COVID+ and COVID-, from the [90] dataset. (b). Example images from three classes, namely, normal, viral
pneumonia and bacterial pneumonia, from the [91] dataset.

datasets.
Table 1
Summary of major publicly available datasets.
5. Evaluation
Dataset Size Type Classes

cohen Regularly updated Chest X-ray and CT images 5 As in any other classification tasks, the metrics that are used to
Paul Mooney 5856 Chest X-ray 3 evaluate the models are accuracy and precision, which are also called
Kaggle 97 Chest X-ray and CT images 2
positive prediction value (PPV) and negative prediction value (NPV),
COVIDx 104,009 CT images 3
ChestXray-8 108,948 Chest x-ray images 3 respectively; specificity; recall, which is also called sensitivity; and F1-
CheXpert 224,316 Chest radiographs 5 score; these are the most commonly used measures. To calculate these
Kaggle 2 2909 Chest X-ray images 4 measures, four main metrics are used: (a) correctly identified diseased
cases (true positives, TP), (b) incorrectly classified diseased cases (false
negatives, FN), (c) correctly identified healthy cases (true negatives,
normal and 25 COVID images are used [90,100,105]. are used to obtain
TN), and (d), incorrectly classified healthy cases (false positives, FP).
135 COVID images, and 320 pneumonia images from Ref. [106] are
The equations for calculating accuracy, specificity and sensitivity are
collected to form the dataset in Ref. [43]. To balance the dataset, only
presented in 1, 5, and 4.
102 images from both classes are considered, and 10-fold
cross-validation is performed as the dataset is small. TP + TN
Accuracy = (1)
First [92], is used to train COVID-CAPS. Then, transfer learning is TP + TN + FP + FN
performed on COVIDx [94] in COVID-CAPS [62]. The COVIDx dataset is /
used in Ref. [86]. Balanced and unbalanced datasets are considered for Precision PPV =
TP
(2)
experiments. CXR images and noisy snapshots of the lung images are the TP + FP
inputs that are used in Ref. [87]. Normal and pneumonia images are
TN
obtained from Ref. [18], and COVID images are obtained from Ref. [90]. NPV = (3)
TN + FN
Microsoft Office Lens is used to capture snapshots of the images on the
PC screen to create the noisy snapshot dataset. The captured images are /
TP
RGB images, which are converted to 8-bit grayscale images [78]. uses 53 Recall Sensitivity = (4)
TP + FN
COVID images from Ref. [100]. Two patch datasets are obtained from
these 53 images by selecting the COVID-infected and noninfected re­ TN
gions in the CT images. Two patch sizes are considered: 16 × 16 and 32 Specificity = (5)
TN + FP
× 32. A total of 3000 patches from COVID infection images and 3000
no-finding patches are used to form the dataset. 2 ∗ Precision ∗ Recall
F1 − Score = (6)
A few images that represent each class from the most commonly used Precision + Recall
datasets, namely [90,96], are presented in Fig. 4 [90]. uses COVID+ and Table 1 summarizes the methods and the accuracies that are realized
COVID-images, and [96] uses normal, bacterial and viral pneumonia in various papers. Additionally, methods for binary classification are
images. Table 2 summarizes the primarily used publicly available presented in the table. Fig. 5 presents a detailed comparison of the

7
N. Subramanian et al. Computers in Biology and Medicine 143 (2022) 105233

Table 2
Comparative analysis of the methods in terms of accuracy, precision/PPV, recall/sensitivity, specificity, NPV and F1-score.
Method Model/Backbone Accuracy Precision Sensitivity Specificity NPV F1-score

[35] VGG19 98.75 – 92.85 98.75 – –

MobileNetV2 97.40 – 99.10 97.09 – –
Inception 86.13 – 12.94 99.70 – –
Xception 85.57 – 0.08 99.99 – –
InceptionResNetV2 84.38 – 0.01 99.83 – –
[36] InceptionV3 95.4 73.4 90.6 96.0 – 81.1
ResNet50 96.1 76.5 91.8 96.6 – 83.5
ResNet101 96.1 84.2 78.3 98.2 – 81.2
ResNet152 93.9 74.8 65.4 97.3 – 69.8
InceptionResNetV2 94.2 67.7 83.5 95.4 – 74.8
[37] InceptionV3 100 100 100 100 100 100
[41] ResNet18 – – 98.0 90.7 – –
ResNet50 – – 98.0 89.6 – –
SqueezeNet – – 98.0 92.9 – –
DenseNet121 – – 98.0 75.1 – –
[42] VGG19 90.0 83.0 100 – – 91.0
DenseNet201 90.0 83.0 100 – – 91.0
ResNetV2 70.0 100 40.0 – – 57.0
InceptionResNetV2 80.0 100 60.0 – – 75.0
XceptionNet 80.0 100 60.0 – – 75.0
MobileNetV2 60.0 100 20.0 – – 33.0
[46] VGG19 99.6 99.2 98.6 99.8 – 98.9
ResNet18 99.6 99.6 99.6 99.3 – 99.6
DenseNet201 99.7 99.7 99.7 99.55 – 99.7
SqueezeNet 99.4 99.4 99.4 99.84 – 98.4
MobileNetV2 99.65 99.65 99.65 99.26 – 99.65
ResNet101 99.6 99.6 99.6 99.31 – 99.6
InceptionV3 99.40 98.80 98.33 99.7 – 98.56
[47] ResNet 89.0 67.0 89.0 85.0 – 76.0
InceptionV3 88.0 90.0 88.0 90.0 – 85.0
InceptionResNetV2 95.0 97.0 96.0 94 – 96.0
DenseNet169 92.0 94.0 96.0 95.0 – 95.0
NASNetLarge 98.0 95.0 91.0 98.0 – 98.0
[48] Inception 89.5 71.0 88.0 87.0 95.0 77.0
[51] AlexNet 96.1 96.52 95.37 – – 95.94
ResNet18 99.0 98.97 98.97 – – 98.97
GoogLeNet 96.8 98.63 98.31 – – 94.46
SqueezeNet 97.8 93.6 95.88 – – 98.47
[52] CNN 84.18 94.05 78.33 93.07 – 85.66
VGG16 86.26 87.73 85.22 87.36 – 86.46
VGG19 85.94 80.39 90.43 82.35 – 85.11
ResNet50 96.61 98.46 94.92 98.43 – 96.67
MobileNetV2 96.27 98.06 94.61 98.02 – 96.30
InceptionV3 94.59 93.75 95.35 93.85 – 94.54
InceptionResNetV2 96.09 98.61 93.88 98.53 – 96.19
DenseNet201 93.66 99.01 89.44 98.89 – 93.98
Xception 83.14 95.77 76.45 94.34 – 85.03
[56] VGG19+CNN 98.05 98.43 98.05 99.5 99.3 98.24
ResNet152V2 95.31 95.31 95.31 98.4 98.4 95.31
ResNet152V2+GRU 96.09 96.06 96.09 98.7 98.7 96.09
ResNet152V2+Bi-GRU 93.36 93.35 93.16 97.8 97.8 93.26

results of novel architectures and other approaches. Specificity and sensitivity are the measures used in Refs. [7,35,37,41,46,
The receiver operating characteristic (ROC) curve and area under the 48,52,60–63,70,73,77,78], and [85]. Precision/PPV, recall and f1-score
ROC curve (AUC) are the other evaluation metrics that are commonly are used in Refs. [37,42,46–48,51,52,60,73,78,81,83,86], and [85].
used. The ROC curve is used to show the performance of the proposed Additionally, a statistical analysis among the models is performed using
model by plotting the true positive rate (TPR), which is also called the the false-positive rate, F1, MCC and kappa measures in Ref. [77].
recall, against the false-positive rate (FPR), at various thresholds. The ROC-AUCs are used to measure the model performance in Refs. [38,41,
equation for calculating FPR is presented in Equation (7). Lowering the 46,47,62,83], and [60]. [69] uses the accuracy of all classes, accuracy of
classification threshold results in the classification of more items as each class, precision, recall and specificity. Overall accuracy, classwise
positive, thereby increasing both the number of false positives and the precision, recall and F-measure are the measures that are used in
number of true positives. AUC is an aggregate measure for evaluating a Ref. [71]. CPU (%), memory (MB), energy and AUC are used in Ref. [87].
model at various possible thresholds. It is the two-dimensional area The Matthews Correlation Coefficient (MCC) [107] is the extra metric
under the ROC curve between (0,0) and (1,1). AUC is the probability that is used in Ref. [78].
that the model ranks a random positive example higher than a random
negative example. 6. Discussion and future direction
FP
FPR = (7) From Table 2, Salman et al. [37] realize the best performance in
FP + TN
terms of accuracy, precision, specificity, sensitivity, NPV and F1-score,
Accuracy, by default, is the common metric that is used by almost with values of 100%. This method uses InceptionV3 as the model with
every method in the study except in Refs. [7,38,41,70,86], and [87]. transfer learning. However, the use of the same InceptionV3 architecture

8
N. Subramanian et al. Computers in Biology and Medicine 143 (2022) 105233

Fig. 5. Graphical representation of the accuracy results for novel architectures and other approaches.

in Refs. [36,46,47], and [52] did not produce the same results as in requirement of the ResNet model is so large that it is expensive and
Ref. [37]. It is observed that Salman et al. use data augmentation with impractical to deploy the trained model on a mobile device. Memory is
two classes of 130 images and 260 images in total. The models that compromised in the place of accuracy. The feasibility and portability of
produce the second- and third-best results are Denset201 and MobileNet the application for the detection of coronavirus is affected.
V2 in Ref. [46], which are designed by Chowdhury et al. The second and Accuracy is the common metric that is used to evaluate the perfor­
third best results are not far from the first result of 100%. DenseNet201 mances of models. VGG19 shows satisfactory performance with an ac­
achieves 99.7% accuracy, which is the second-best result. MobileNetV2 curacy of 98.75%. In addition, VGG19 has fewer parameters and a
realizes an accuracy of 99.65%, which is only 0.05% less than the shallower model, which makes it easily deployable even in mobile ap­
second-best accuracy and 0.35% less than the best accuracy results. plications and mobile devices. Although some of the other approaches,
Apart from accuracy, other evaluation metrics are used, namely, such as deep feature extraction [78] and hierarchical classification [81],
precision, recall/sensitivity, specificity, and F1-score. Hemdan et al. have been tested, they did not achieve better performance in
[42] produce a precision of 100%. Similar to accuracy, Denset201 comparison.
produces the second-best result with 99.7% precision, and MobileNetV2 As discussed earlier, deep learning methods need large amounts of
has the third-best precision of 99.65%. DenseNet201 and MobileNetV2 data to perform well. Although most of the methods have tried to
produce the second-best results in terms of sensitivity and F1-score. The overcome the shortage of data with various data augmentation methods,
sensitivity and F1-score values for DenseNet201 and MobileNetV2 are there is no proof of real-time detection. There is no proven evidence on
99.7% and 99.65%, respectively. However, DensetNet201 and Mobile­ the effectiveness of data augmentation in real-life and live images for the
Net V2 do not produce results with higher specificity. In terms of spec­ detection of coronavirus. Creating a public dataset with possible classes
ificity, SqueezeNet realizes the second-best value of 99.84%, and VGG19 requires help from medical experts, which is time-consuming. Since the
produces the third-best values of 99.8%, with a mere difference of availability of public datasets is low, studies have tailored custom
0.04%. NPV values are not used by many methods. The best NPV value datasets by combining two or three repositories based on the applica­
of 100% is produced by Inception V3 in Ref. [37]. Better results are tion. The popular representations are [90] for COVID images and [96]
produced by the combination of VGG19 and CNN, which realizes 99.3%, for normal, bacterial and viral pneumonia images.
and the combination of ResNet152 V2 and GRU, which realizes 98.7% A preprocessing step for resizing the input image to fit the archi­
[56]. tecture is conducted before training and testing the model. Careful
In summary, most of the methods utilize transfer learning on estab­ consideration must be taken when dealing with medical images. Medical
lished architectures for the classification of lung images. Even if novel images are easily prone to noise, and this noise has to be removed before
architectures are proposed, due to scarcity of available image data, passing them to the model; otherwise, the model will learn the noise
transfer learning on the ImageNet dataset is considered [60,71]. [109]. This may affect the performance of the model. An effective pre­
Different network architectures are used by different methods. Out of all processing step for removing artifacts and noise is essential for
of them, Inception, DenseNet, MobileNet, SqueezeNet and the VGG improving the model performance.
family outperform the other families. The major advantage in using the deep learning models is the ease of
To effectively detect coronavirus infection, an easy, fast, and accu­ using them without any requirements for manually picking the features.
rate application that can be deployed in hand-held devices has to be However, in the case of medical images, the selection and use of features
developed. Most of the architectures that are used in the literature have were of higher importance than any other tasks. The features that are
many layers and, hence, huge numbers of parameters to store and selected by the deep learning models are not interpretable by medical
compute. The ResNet50 architecture has 53 convolutions and one fully professionals, and hence, the reliability is not certain; hence, it is unclear
connected layer with over 23 million trainable parameters [108]. per­ how the application can help them.
formed a detailed analysis on the memory requirements of each model The privacy and security of confidential materials such as X-ray
before and after deployment in a hand-held device chip. The memory images, patient information and other details are of the utmost

9
N. Subramanian et al. Computers in Biology and Medicine 143 (2022) 105233

importance. [9] M. Asadi-Aghbolaghi, A. Clapes, M. Bellantonio, H.J. Escalante, V. Ponce-López,

X. Baró, I. Guyon, S. Kasaei, S. Escalera, A survey on deep learning based
In the future, more publicly available datasets with lung images can
approaches for action and gesture recognition in image sequences, in: 2017 12th
be collected and constructed for future use. Without the availability of IEEE International Conference on Automatic Face & Gesture Recognition (FG
quality data, the performance of the deep learning models cannot be 2017), IEEE, 2017, pp. 476–483.
improved. Other research directions include constructing and anno­ [10] I. Masi, Y. Wu, T. Hassner, P. Natarajan, Deep face recognition: a survey, in: 2018
31st SIBGRAPI Conference on Graphics, Patterns and Images (SIBGRAPI), IEEE,
tating data and providing metadata information. 2018, pp. 471–478.
[11] N. Subramanian, O. Elharrouss, S. Al-Maadeed, A. Bouridane, Image
7. Conclusion Steganography: A Review of the Recent Advances, IEEE Access, 2021.
[12] O. Elharrouss, N. Almaadeed, S. Al-Maadeed, An image steganography approach
based on k-least significant bits (k-lsb), in: 2020 IEEE International Conference on
The COVID-19 pandemic is a novel pandemic that is caused by the Informatics, IoT, and Enabling Technologies (ICIoT), 2020, pp. 131–135, https://
coronavirus, and the only preventive measures that are available thus doi.org/10.1109/ICIoT48696.2020.9089566.
[13] N. Subramanian, O. Elharrouss, S. Al-Maadeed, A. Bouridane, End-to-end image
far are social distancing and early detection. For early detection and steganography using deep convolutional autoencoders, IEEE Access (2021),
prevention of spread, deep learning models are trained to detect and https://doi.org/10.1109/ACCESS.2021.3113953, 1–1.
classify lung images. Since the spread of the COVID-19 pandemic started [14] A. Ulhaq, A. Khan, D. Gomes, M. Pau, Computer vision for covid-19 control: a
survey, arXiv preprint arXiv:2004 (2020), 09420.
recently in the last quarter of 2019, limited data are available for [15] F. Shi, J. Wang, J. Shi, Z. Wu, Q. Wang, Z. Tang, K. He, Y. Shi, D. Shen, Review of
training deep learning models. To overcome this scarcity, researchers artificial intelligence techniques in imaging data acquisition, segmentation and
created custom datasets by combining many repositories. Transfer diagnosis for covid-19, IEEE.Rev. Biomed. Eng. 14 (2020) 1–15.
[16] J. Gu, Z. Wang, J. Kuen, L. Ma, A. Shahroudy, B. Shuai, T. Liu, X. Wang, G. Wang,
learning on established architectures, novel architectures with transfer J. Cai, et al., Recent advances in convolutional neural networks, Pattern Recogn.
learning on the ImageNet dataset, and other approaches, such as deep 77 (2018) 354–377.
feature extraction using a deep learning architecture and hierarchical [17] J. Kunhoth, A. Karkar, S. Al-Maadeed, A. Al-Attiyah, Comparative analysis of
computer-vision and ble technology based indoor navigation systems for people
classification methods, are the methods that are available in the study.
with visual impairments, Int. J. Health Geogr. 18 (1) (2019) 1–18.
Among these available methods, transfer learning performs the best, and [18] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, Z. Wojna, Rethinking the inception
out of all the architectures, InceptionV3, DenseNet201, and Mobile­ architecture for computer vision, in: Proceedings of the IEEE Conference on
NetV2 realize higher accuracy, while SqueezeNet and VGG19 show Computer Vision and Pattern Recognition, 2016, pp. 2818–2826.
[19] M.A. Nielsen, Neural Networks and Deep Learning, vol. 25, Determination press,
better specificity. Although vaccine drives are occurring all around the San Francisco, CA, 2015.
world, supply chain logistics and fear of the vaccine are some of the [20] M. Sarıgül, B.M. Ozyildirim, M. Avci, Differential convolutional neural network,
major issues. The RT–PCR test that is currently used for the detection of Neural Network. 116 (2019) 279–287.
[21] A. Fakhrou, J. Kunhoth, S. Al Maadeed, Smartphone-based Food Recognition
coronavirus is expensive, time-consuming, and less sensitive. Chest X- System Using Multiple Deep Cnn Models, Multimedia Tools and Applications,
rays, CT scans, and ultrasound images of the lungs are primarily 2021, pp. 1–22.
considered for detecting coronavirus detection by health care officials. [22] S. Ioffe, C. Szegedy, Batch normalization: accelerating deep network training by
reducing internal covariate shift, in: International Conference on Machine
Deep learning methods can facilitate coronavirus detection using images Learning, PMLR, 2015, pp. 448–456.
at early stages. The best results of 100% accuracy, 100% precision, [23] H. Zhao, O. Gallo, I. Frosio, J. Kautz, Loss Functions for Neural Networks for
100% specificity, 100% sensitivity, 100% NPV, and 100% F1-score show Image Processing, arXiv preprint arXiv:1511.08861, 2015.
[24] S. Chakraborty, R. Mondal, P.K. Singh, R. Sarkar, D. Bhattacharjee, Transfer
the higher reliability of the deep learning methods. learning with fine tuning for human action recognition from still images,
Multimed. Tool. Appl. 80 (13) (2021) 20547–20578.
Declaration of competing interest [25] Z. Cai, C. Peng, A study on training fine-tuning of convolutional neural networks,
in: 2021 13th International Conference on Knowledge and Smart Technology
(KST), IEEE, 2021, pp. 84–89.
None declared. [26] K. Simonyan, A. Zisserman, Very deep convolutional networks for large-scale
image recognition, arXiv preprint arXiv (2014) 1409–1556. https://arxiv.org/
abs/1409.1556.
Acknowledgement [27] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang,
A. Karpathy, A. Khosla, M. Bernstein, A.C. Berg, L. Fei-Fei, ImageNet large scale
This publication was supported by Qatar University COVID19 visual recognition challenge, Int. J. Comput. Vis. 115 (3) (2015) 211–252,
https://doi.org/10.1007/s11263-015-0816-y.
Emergency Response Grant (QUERG-CENG-2020-1). The findings ach­ [28] J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, L. Fei-Fei, in: ImageNet: A Large-Scale
ieved herein are solely the responsibility of the authors. The contents of Hierarchical Image Database, 2009. CVPR09.
this publication are solely the responsibility of the authors and do not [29] C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan,
V. Vanhoucke, A. Rabinovich, Going deeper with convolutions, in: Proceedings of
necessarily represent the official views of the Qatar University. the IEEE Conference on Computer Vision and Pattern Recognition, 2015, pp. 1–9.
[30] C. Szegedy, S. Ioffe, V. Vanhoucke, A.A. Alemi, Inception-v4, inception-resnet and
References the impact of residual connections on learning, in: Thirty-first AAAI Conference
on Artificial Intelligence, 2017.
[31] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in:
[1] F. Wu, S. Zhao, B. Yu, Y.-M. Chen, W. Wang, Z.-G. Song, Y. Hu, Z.-W. Tao, J.-
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,
H. Tian, Y.-Y. Pei, et al., A new coronavirus associated with human respiratory
2016, pp. 770–778.
disease in China, Nature 579 (7798) (2020) 265–269.
[32] G. Huang, Z. Liu, L. van der Maaten, K.Q. Weinberger, Densely Connected
[2] R. Singh, S. Sarsaiya, T.A. Singh, T. Singh, L.K. Pandey, P.K. Pandey, N. Khare,
Convolutional Networks, 2016 arXiv:1608.06993.
F. Sobin, R. Sikarwar, M.K. Gupta, Corona virus (covid-19) symptoms prevention
[33] A.G. Howard, M. Zhu, B. Chen, D. Kalenichenko, W. Wang, T. Weyand,
and treatment: a short review, J. Drug Deliv. Therapeut. 11 (2-S) (2021)
M. Andreetto, H. Adam, Mobilenets, Efficient Convolutional Neural Networks for
118–120.
Mobile Vision Applications, 2017 arXiv preprint arXiv:1704.04861 arXiv:
[3] L. Zhou, Z. Wu, Z. Li, Y. Zhang, J.M. McGoogan, Q. Li, X. Dong, R. Ren, L. Feng,
1704.04861.
X. Qi, et al., One hundred days of coronavirus disease 2019 prevention and
[34] F. Chollet, Xception: deep learning with depthwise separable convolutions, in:
control in China, Clin. Infect. Dis. 72 (2) (2021) 332–339.
Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition,
[4] L.P. Samaranayake, C.J. Seneviratne, K.S. Fakhruddin, Coronavirus disease 2019
2017, pp. 1251–1258.
(covid-19) vaccines: a concise review, Oral Dis. (2021) 1–11.
[35] I.D. Apostolopoulos, T.A. Mpesiana, Covid-19: automatic detection from x-ray
[5] A. Rehman, T. Saba, U. Tariq, N. Ayesha, Deep learning-based covid-19 detection
images utilizing transfer learning with convolutional neural networks, Phys. Eng.
using ct and x-ray images: current analytics and comparisons, IT Professional 23
Sci. Med. (2020) 1.
(3) (2021) 63–68.
[36] A. Narin, C. Kaya, Z. Pamuk, Automatic detection of coronavirus disease (covid-
[6] economist.com, 2020. Academic articles, https://www.economist.com/graphi
19) using x-ray images and deep convolutional neural networks, Pattern Anal.
c-detail/2020/03/20/coronavirus-research-is-being-published-at-a-furious-pace.
Appl. (2021) 1–14.
[7] M. Hassaballah, A.I. Awad, Deep Learning in Computer Vision: Principles and
[37] F.M. Salman, S.S. Abu-Naser, E. Alajrami, B.S. Abu-Nasser, B.A. Alashqar, Covid-
Applications, CRC Press, 2020.
19 detection using artificial intelligence, Int. J. Acad. Eng. Res.(IJAER) 4 (2020)
[8] G. Ciaparrone, F.L. Sánchez, S. Tabik, L. Troiano, R. Tagliaferri, F. Herrera, Deep
18–25.
learning in video multi-object tracking: a survey, Neurocomputing 381 (2020)
61–88.

10
N. Subramanian et al. Computers in Biology and Medicine 143 (2022) 105233

[38] G. Maguolo, L. Nanni, A critic evaluation of methods for covid-19 automatic [67] H.S. Maghdid, A.T. Asaad, K.Z. Ghafoor, A.S. Sadiq, M.K. Khan, Diagnosing covid-
detection from x-ray images, arXiv preprint arXiv:2004 (2020) 12823. 19 pneumonia from x-ray and ct images using deep learning and transfer learning
[39] F.N. Iandola, S. Han, M.W. Moskewicz, K. Ashraf, W.J. Dally, K. Keutzer, algorithms, arXiv preprint arXiv:2004 (2020), 00038.
Squeezenet: Alexnet-level accuracy with 50x fewer parameters and< 0.5 mb [68] K. Hammoudi, H. Benhabiles, M. Melkemi, F. Dornaika, I. Arganda-Carreras,
model size, arXiv preprint arXiv:1602.07360 (2016). D. Collard, A. Scherpereel, Deep learning on chest x-ray images to detect and
[40] G. Huang, Z. Liu, L. Van Der Maaten, K.Q. Weinberger, Densely connected evaluate pneumonia cases at the era of covid-19, arXiv preprint arXiv:2004
convolutional networks, in: Proceedings of the IEEE Conference on Computer (2020), 03399.
Vision and Pattern Recognition, 2017, pp. 4700–4708. [69] M. Rahimzadeh, A. Attar, A new modified deep convolutional neural network for
[41] S. Minaee, R. Kafieh, M. Sonka, S. Yazdani, G.J. Soufi, Deep-covid: predicting detecting covid-19 from x-ray images, arXiv e-prints (2020) arXiv (2004).
covid-19 from chest x-ray images using deep transfer learning, Med. Image Anal. [70] I.D. Apostolopoulos, S.I. Aznaouridis, M.A. Tzani, Extracting possibly
65 (2020), 101794. representative covid-19 biomarkers from x-ray images with deep learning
[42] E.E.-D. Hemdan, M.A. Shouman, M.E. Karar, Covidx-net: a framework of deep approach and image data related to pulmonary diseases, J. Med. Biol. Eng. (2020)
learning classifiers to diagnose covid-19 in x-ray images, arXiv preprint arXiv: 1.
2003 (2020) 11055. [71] A.I. Khan, J.L. Shah, M.M. Bhat, Coronet: a deep neural network for detection and
[43] L.O. Hall, R. Paul, D.B. Goldgof, G.M. Goldgof, Finding covid-19 from chest x-rays diagnosis of covid-19 from chest x-ray images, Comput. Methods Progr. Biomed.
using deep learning on a small dataset, arXiv preprint arXiv:2004 (2020), 02060. (2020), 105581.
[44] M. Lin, Q. Chen, S. Yan, Network in network, arXiv preprint arXiv:1312.4400 [72] J. Redmon, A. Farhadi, Yolo9000: better, faster, stronger, in: Proceedings of the
(2013). IEEE Conference on Computer Vision and Pattern Recognition, 2017,
[45] W. Ma, J. Lu, An equivalence of fully connected layer and convolutional layer, pp. 7263–7271.
arXiv preprint arXiv:1712.01252 (2017). [73] T. Ozturk, M. Talo, E.A. Yildirim, U.B. Baloglu, O. Yildirim, U.R. Acharya,
[46] M.E. Chowdhury, T. Rahman, A. Khandakar, R. Mazhar, M.A. Kadir, Z. Automated detection of covid-19 cases using deep neural networks with x-ray
B. Mahbub, K.R. Islam, M.S. Khan, A. Iqbal, N. Al Emadi, et al., Can ai help in images, Comput. Biol. Med. (2020), 103792.
screening viral and covid-19 pneumonia? IEEE Access 8 (2020) 132665–132676. [74] F. Ozyurt, T. Tuncer, A. Subasi, An automated covid-19 detection based on fused
[47] N.S. Punn, S. Agarwal, Automated diagnosis of covid-19 with limited dynamic exemplar pyramid feature extraction and hybrid feature selection using
posteroanterior chest x-ray images using fine-tuned deep neural networks, Appl. deep learning, Comput. Biol. Med. 132 (2021), 104356.
Intell. 51 (5) (2021) 2689–2702. [75] T. Mahmud, M.A. Rahman, S.A. Fattah, Covxnet: a multi-dilation convolutional
[48] S. Wang, B. Kang, J. Ma, X. Zeng, M. Xiao, J. Guo, M. Cai, J. Yang, Y. Li, X. Meng, neural network for automatic covid-19 and other pneumonia detection from chest
et al., A deep learning algorithm using ct images to screen for corona virus disease x-ray images with transferable multi-receptive feature optimization, Comput.
(covid-19), Eur. Radiol. (2021) 1–9. Biol. Med. 122 (2020), 103869.
[49] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, [76] S. Karakanis, G. Leontidis, Lightweight deep learning models for detecting covid-
A. Courville, Y. Bengio, Generative adversarial nets, Adv. Neural Inf. Process. 19 from chest x-ray images, Comput. Biol. Med. 130 (2021), 104181.
Syst. 27 (2014). [77] P.K. Sethy, S.K. Behera, Detection of coronavirus disease (covid-19) based on
[50] A. Creswell, T. White, V. Dumoulin, K. Arulkumaran, B. Sengupta, A.A. Bharath, deep features, Preprints 2020030300 (2020), 2020.
Generative adversarial networks: an overview, IEEE Signal Process. Mag. 35 (1) [78] U. Ozkaya, S. Ozturk, M. Barstugan, Coronavirus (covid-19) classification using
(2018) 53–65. deep features fusion and ranking technique, arXiv preprint arXiv:2004 (2020),
[51] N.E.M. Khalifa, M.H.N. Taha, A.E. Hassanien, S. Elghamrawy, Detection of 03698.
coronavirus (covid-19) associated pneumonia based on generative adversarial [79] D.-N. Le, V.S. Parvathy, D. Gupta, A. Khanna, J.J. Rodrigues, K. Shankar, Iot
networks and a fine-tuned deep transfer learning model using chest x-ray dataset, enabled depthwise separable convolution neural network with deep support
arXiv preprint arXiv:2004 (2020), 01184. vector machine for covid-19 diagnosis and classification, Int. J.Mach.Learn.
[52] K. El Asnaoui, Y. Chawki, A. Idri, Automated methods for detection and Cybern. (2021) 1–14.
classification pneumonia based on x-ray images using deep learning, in: Artificial [80] K. Bayoudh, F. Hamdaoui, A. Mtibaa, Hybrid-covid: a novel hybrid 2d/3d cnn
Intelligence and Blockchain for Future Cybersecurity Applications, Springer, based on cross-domain adaptation approach for covid-19 screening from chest x-
2021, pp. 257–284. ray images, Phys. Eng.Sci. Med. 43 (4) (2020) 1415–1431.
[53] I.-M. Sintorn, L. Bischof, P. Jackway, S. Haggarty, M. Buckley, Gradient based [81] E.J.d.S. Luz, P.L. Silva, R. Silva, L. Silva, G. Moreira, D. Menotti, Towards an
intensity normalization, J. Microsc. 240 (3) (2010) 249–258. effective and efficient deep learning model for covid-19 patterns detection in x-
[54] G. Yadav, S. Maheshwari, A. Agarwal, Contrast limited adaptive histogram ray images, CoRR (2020) 1–14.
equalization based enhancement for real time video system, in: 2014 [82] M. Tan, Q. Le, Efficientnet: rethinking model scaling for convolutional neural
International Conference on Advances in Computing, Communications and networks, in: International Conference on Machine Learning, PMLR, 2019,
Informatics (ICACCI), IEEE, 2014, pp. 2392–2397. pp. 6105–6114.
[55] M. Toğaçar, B. Ergen, Z. Cömert, Covid-19 detection using deep learning models [83] D. Lv, W. Qi, Y. Li, L. Sun, Y. Wang, A cascade network for detecting covid-19
to exploit social mimic optimization and structured chest x-ray images using using chest x-rays, arXiv preprint arXiv:2005 (2020), 01468.
fuzzy color and stacking approaches, Comput. Biol. Med. 121 (2020), 103805. [84] R.R. Selvaraju, M. Cogswell, A. Das, R. Vedantam, D. Parikh, D. Batra, Grad-cam:
[56] D.M. Ibrahim, N.M. Elshennawy, A.M. Sarhan, Deep-chest: multi-classification visual explanations from deep networks via gradient-based localization, in:
deep learning model for diagnosing covid-19, pneumonia, and lung cancer chest Proceedings of the IEEE International Conference on Computer Vision, 2017,
diseases, Comput. Biol. Med. 132 (2021), 104348. pp. 618–626.
[57] S. Serte, H. Demirel, Deep learning for diagnosis of covid-19 using 3d ct scans, [85] O. Gozes, M. Frid-Adar, N. Sagie, H. Zhang, W. Ji, H. Greenspan, Coronavirus
Comput. Biol. Med. 132 (2021), 104306. detection and analysis on chest ct with deep learning, arXiv preprint arXiv:
[58] M.A. Alves, G.Z. Castro, B.A.S. Oliveira, L.A. Ferreira, J.A. Ramírez, R. Silva, F. 2004.02640 (2020).
G. Guimarães, Explaining machine learning based diagnosis of covid-19 from [86] M. Karim, T. Döhmen, D. Rebholz-Schuhmann, S. Decker, M. Cochez, O. Beyan, et
routine blood tests with decision trees and criteria graphs, Comput. Biol. Med. al., Deepcovidexplainer: explainable covid-19 predictions based on chest x-ray
132 (2021), 104335. images, arXiv preprint arXiv:2004 (2020) 4582.
[59] I. Shiri, M. Sorouri, P. Geramifar, M. Nazari, M. Abdollahi, Y. Salimi, B. Khosravi, [87] X. Li, C. Li, D. Zhu, Covid-mobilexpert: on-device covid-19 screening using
D. Askari, L. Aghaghazvini, G. Hajianfar, et al., Machine learning-based snapshots of chest x-ray, arXiv preprint arXiv:2004 (2020) 3042.
prognostic modeling using clinical data and quantitative radiomic features from [88] W. Jin, S. Dong, C. Dong, X. Ye, Hybrid ensemble model for differential diagnosis
chest ct images in covid-19 patients, Comput. Biol. Med. 132 (2021), 104304. between covid-19 and common viral pneumonia by chest x-ray radiograph,
[60] L. Wang, A. Wong, Covid-net: a tailored deep convolutional neural network Comput. Biol. Med. 131 (2021), 104252.
design for detection of covid-19 cases from chest x-ray images, arXiv preprint [89] T. Rahman, A. Khandakar, Y. Qiblawey, A. Tahir, S. Kiranyaz, S.B.A. Kashem, M.
arXiv:2003 (2020), 09871. T. Islam, S. Al Maadeed, S.M. Zughaier, M.S. Khan, et al., Exploring the effect of
[61] J. Zhang, Y. Xie, Y. Li, C. Shen, Y. Xia, Covid-19 screening on chest x-ray images image enhancement techniques on covid-19 detection using chest x-ray images,
using deep learning based anomaly detection, arXiv preprint arXiv:2003 (2020) Comput. Biol. Med. 132 (2021), 104319.
12338. [90] J.P. Cohen, P. Morrison, L. Dao, K. Roth, T.Q. Duong, M. Ghassemi, Covid-19
[62] P. Afshar, S. Heidarian, F. Naderkhani, A. Oikonomou, K.N. Plataniotis, image data collection: prospective predictions are the future, arXiv preprint
A. Mohammadi, Covid-caps: a capsule network-based framework for arXiv:2006.11988 (2020).
identification of covid-19 cases from x-ray images, Pattern Recogn. Lett. 138 [91] chest-xray pneumonia, Kaggle lung pneumonia dataset2. https://www.kaggle.
(2020) 638–643. com/paultimothymooney/chest-xray-pneumonia, 2020. (Accessed 11 November
[63] A. Abbas, M.M. Abdelsamea, M.M. Gaber, Classification of covid-19 in chest x-ray 2021).
images using detrac deep convolutional neural network, arXiv preprint arXiv: [92] X. Wang, Y. Peng, L. Lu, Z. Lu, M. Bagheri, R.M. Summers, Chestx-ray8: hospital-
2003 (2020) 13815. scale chest x-ray database and benchmarks on weakly-supervised classification
[64] M. Siddhartha, A. Santra, Covidlite: a depth-wise separable deep neural network and localization of common thorax diseases, in: 2017 IEEE Conference on
with white balance and clahe for detection of covid-19, arXiv preprint arXiv:2006 Computer Vision and Pattern Recognition (CVPR), Jul 2017, https://doi.org/
(2020) 13873. 10.1109/cvpr.2017.369.
[65] A.H. Panahi, A. Rafiei, A. Rezaee, Fcod: fast covid-19 detector based on deep [93] J. Irvin, P. Rajpurkar, M. Ko, Y. Yu, S. Ciurea-Ilcus, C. Chute, H. Marklund,
learning techniques, Informatics in Medicine Unlocked 22 (2021), 100506. B. Haghgoo, R. Ball, K. Shpanskaya, J. Seekins, D.A. Mong, S.S. Halabi, J.
[66] K.K. Singh, M. Siddhartha, A. Singh, Diagnosis of coronavirus disease (covid-19) K. Sandberg, R. Jones, D.B. Larson, C.P. Langlotz, B.N. Patel, M.P. Lungren, A.
from chest x-ray images using modified xceptionnet, Rom. J. Inf. Sci. Technol. 23 Y. Ng, Chexpert: a large chest radiograph dataset with uncertainty labels and
(657) (2020) 91–115. expert comparison, arXiv:1901 (2019) 7031.

11
N. Subramanian et al. Computers in Biology and Medicine 143 (2022) 105233

[94] COVID-NET, Covidx dataset. https://github.com/lindawangg/COVID-Net, 2020. Omar Elharrouss received a master’s degree in 2013 from the
(Accessed 11 November 2021). Faculty of Sciences, Dhar El Mehraz, Fez, Morocco. In 2017, he
[95] D.S. Kermany, M. Goldbaum, W. Cai, C.C. Valentim, H. Liang, S.L. Baxter, received a Ph.D. from the LIIAN Laboratory at USMBA-Fez
A. McKeown, G. Yang, X. Wu, F. Yan, J. Dong, M.K. Prasadha, J. Pei, M.Y. Ting, University. His research interests include pattern recognition,
J. Zhu, C. Li, S. Hewett, J. Dong, I. Ziyar, A. Shi, R. Zhang, L. Zheng, R. Hou, image processing, and computer vision.
W. Shi, X. Fu, Y. Duan, V.A. Huu, C. Wen, E.D. Zhang, C.L. Zhang, O. Li, X. Wang,
M.A. Singer, X. Sun, J. Xu, A. Tafreshi, M.A. Lewis, H. Xia, K. Zhang, Identifying
medical diagnoses and treatable diseases by image-based deep learning, e9, Cell
172 (5) (2018) 1122–1131, https://doi.org/10.1016/j.cell.2018.02.010. URL,
http://www.sciencedirect.com/science/article/pii/S0092867418301545.
[96] Kaggle, Kaggle lung pneumonia dataset. https://www.kaggle.com, 2020.
(Accessed 11 November 2021). Somaya Al-Maadeed received a Ph.D. degree in computer science from Nottingham, U.K.,
[97] NIH, Nih Dataset, 2020. https://openi.nlm.nih.gov/. (Accessed 11 November in 2004. She is the Coordinator of the Computer Vision and AI Research Group. She enjoys
2021). excellent collaboration with national and international institutions and industry. She is a
[98] Kaggle, Kaggle lung pneumonia dataset1. https://www.kaggle.com/andrewmvd principal investigator of several funded research projects that involve approximately five
/convid19-x-rays, 2020. (Accessed 11 November 2021). million people. She has published extensively in the field of pattern recognition and
[99] RSNA, Rsna pneumonia detection challenge dataset. https://www.kaggle.com/c/ delivered workshops on teaching programming for undergraduate students. She has
rsna-pneumonia-detection-challenge/data, 2020. (Accessed 11 November 2021). attended workshops that were related to higher education strategy, assessment methods,
[100] SIRM, S.i. s. o. m. a. i. radiology, covid-19 Database, 2020. https://www.sirm. and interactive teaching. In 2015, she was elected as the IEEE Chair for the Qatar Section.
org/category/senza-categoria/covid-19/. (Accessed 11 November 2021), 2020.
[101] D. Kermany, K. Zhang, M. Goldbaum, Labeled optical coherence tomography
(oct) and chest x-ray images for classification, Mendeley data 2 (2018).
Muhammad E. H. Chowdhury received B.Sc. and M.Sc. degrees
[102] S. Candemir, S. Jaeger, K. Palaniappan, J.P. Musco, R.K. Singh, Z. Xue,
from the Department of Electrical and Electronics Engineering,
A. Karargyris, S. Antani, G. Thoma, C.J. McDonald, Lung segmentation in chest
University of Dhaka, Bangladesh, and a Ph.D. degree from the
radiographs using anatomical atlases with nonrigid registration, IEEE Trans. Med.
University of Nottingham, U.K., in 2014. He worked as a Post­
Imag. 33 (2) (2013) 577–590.
doctoral Research Fellow and a Hermes Fellow with the Sir Peter
[103] S. Jaeger, A. Karargyris, S. Candemir, L. Folio, J. Siegelman, F. Callaghan, Z. Xue,
Mansfield Imaging Centre, University of Nottingham. He is
K. Palaniappan, R.K. Singh, S. Antani, et al., Automatic tuberculosis screening
currently working as an Assistant Professor with the Department
using chest radiographs, IEEE Trans. Med. Imag. 33 (2) (2013) 233–245.
of Electrical Engineering, Qatar University. Before joining Qatar
[104] Adrian, Medical images. https://www.pyimagesearch.com/category/medical/,
University, he worked in several universities in Bangladesh. He
2020. (Accessed 11 November 2021).
has two patents and published approximately 80 peer-reviewed journal articles, confer­
[105] Radiopedia, Radiopedia. https://radiopaedia.org/, 2020. (Accessed 11 November
ence papers, and four book chapters. His current research interests include biomedical
2021).
instrumentation, signal processing, wearable sensors, medical image analysis, machine
[106] Nih chestxrays, nih-chestxrays. https://www.kaggle.com/nih-chestxrays/data,
learning, embedded system design, and simultaneous EEG/fMRI. He is currently the
2020. (Accessed 11 November 2021).
recipient of several NPRP and UREP grants from QNRF and internal grants from Qatar
[107] D. Chicco, G. Jurman, The advantages of the matthews correlation coefficient
University, and he conducts academic and government projects. He has been involved in
(mcc) over f1 score and accuracy in binary classification evaluation, BMC Genom.
EPSRC, ISIF, and EPSRC-ACC grants, along with various national and international pro­
21 (1) (2020) 1–13.
jects. He has worked as a consultant for the project entitled Driver Distraction Manage­
[108] K. Siu, D.M. Stuart, M. Mahmoud, A. Moshovos, Memory requirements for
ment Using Sensor Data Clouds (2013–2014), Information Society Innovation Fund (ISIF)
convolutional neural network hardware accelerators, in: 2018 IEEE International
Asia. He is an Active Member of British Radiology, the Institute of Physics, ISMRM, and
Symposium on Workload Characterization (IISWC), IEEE, 2018, pp. 111–121.
HBM. He received the ISIF Asia Community Choice Award 2013 for a project entitled
[109] S. Rawat, K. Rana, V. Kumar, A novel complex-valued convolutional neural
Design and Develop-ment of a Precision Agriculture Information System for Bangladesh.
network for medical image denoising, Biomed. Signal Process Control 69 (2021),
He recently received the COVID-19 Dataset Award and National AI Com-petition awards
102859.
for his contribution to the fight against COVID-19. He is serving as an Associate Editor for
IEEE Access and a Topic Editor and Review Editor for Frontiers in Neuroscience.

Nandhini Subramanian received master’s degree in computing

from Qatar University, Doha, and bachelor’s degree in Electrical
and Electronics Engineering from the PSG College of Technol­
ogy, India. Her interests include computer vision, artificial in­
telligence, machine learning and cloud computing. She is
currently working as Research Assistant with Dr. Somaya Al-
Maadeed in Qatar University, Doha. She is the winner of the
national-level AI competition (Qatar) at the student level.

12