diagnostics

Article
CRV-NET: Robust Intensity Recognition of Coronavirus in
Lung Computerized Tomography Scan Images
Uzair Iqbal 1, * , Romil Imtiaz 2 , Abdul Khader Jilani Saudagar 3, * and Khubaib Amjad Alam 4

1 Department of Artificial Intelligence and Data Science, National University of Computer and Emerging Sciences,
Islamabad Campus, Islamabad 44000, Pakistan
2 Information and Communication Engineering, Northwestern Polytechnical University, Xi’an 710072, China;
[email protected]
3 Information Systems Department, College of Computer and Information Sciences, Imam Mohammad Ibn
Saud Islamic University (IMSIU), Riyadh 11432, Saudi Arabia
4 Department of Software Engineering, National University of Computer and Emerging Sciences,
Islamabad Campus, Islamabad 44000, Pakistan; [email protected]
* Correspondence: [email protected] (U.I.); [email protected] (A.K.J.S.)

Abstract: The early diagnosis of infectious diseases is demanded by digital healthcare systems.
Currently, the detection of the new coronavirus disease (COVID-19) is a major clinical requirement.
For COVID-19 detection, deep learning models are used in various studies, but the robustness is still
compromised. In recent years, deep learning models have increased in popularity in almost every
area, particularly in medical image processing and analysis. The visualization of the human body’s
internal structure is critical in medical analysis; many imaging techniques are in use to perform
this job. A computerized tomography (CT) scan is one of them, and it has been generally used for
the non-invasive observation of the human body. The development of an automatic segmentation
method for lung CT scans showing COVID-19 can save experts time and can reduce human error. In
this article, the CRV-NET is proposed for the robust detection of COVID-19 in lung CT scan images.
A public dataset (SARS-CoV-2 CT Scan dataset), is used for the experimental work and customized
according to the scenario of the proposed model. The proposed modified deep-learning-based U-Net
model is trained on a custom dataset with 221 training images and their ground truth, which was
Citation: Iqbal, U.; Imtiaz, R.;
labeled by an expert. The proposed model is tested on 100 test images, and the results show that
Saudagar, A.K.J.; Alam, K.A.
the model segments COVID-19 with a satisfactory level of accuracy. Moreover, the comparison of
CRV-NET: Robust Intensity
the proposed CRV-NET with different state-of-the-art convolutional neural network models (CNNs),
Recognition of Coronavirus in Lung
Computerized Tomography Scan
including the U-Net Model, shows better results in terms of accuracy (96.67%) and robustness (low
Images. Diagnostics 2023, 13, 1783. epoch value in detection and the smallest training data size).
https://doi.org/10.3390/
diagnostics13101783 Keywords: deep learning; machine learning; computerized tomography; convolutional neural
network; U-Net
Academic Editor: Saif Ul Islam

Received: 16 April 2023

Revised: 1 May 2023
Accepted: 10 May 2023 1. Introduction
Published: 18 May 2023 The detection of infectious diseases is a continuous job for modern digital healthcare
systems. The report of coronavirus (COVID-19) was in the city of Wuhan in China, and it
was initially discovered in the transmission from animals to humans [1]. COVID-19 is a
deadly virus that has killed hundreds of people every day across the globe since December
Copyright: © 2023 by the authors.
2019 [2]. It is a communicable disease (transferable from one human body to another
Licensee MDPI, Basel, Switzerland.
This article is an open access article
human body), and the discovery of robust methods to diagnosis its intensity is still open to
distributed under the terms and
different scientist’s communities. At the first stage, nucleic acid amplification tests (NAATs),
conditions of the Creative Commons
antigen tests and self-tests are normally in practice for diagnostic purposes, but intelligent
Attribution (CC BY) license (https:// diagnostic tools are still on the way for robust COVID-19 intensity measurement [3]. As
creativecommons.org/licenses/by/ a way to ensure human survival, one of the most successful strategies for infection tree
4.0/). trimming is the early detection of this illness [2,3]. The increasing number of COVID-19

Diagnostics 2023, 13, 1783. https://doi.org/10.3390/diagnostics13101783 https://www.mdpi.com/journal/diagnostics

Diagnostics 2023, 13, 1783 2 of 15

patients is overwhelming in many countries’ healthcare systems. As a result, a reliable

automated methodology for recognizing and measuring infected lung areas would be
extremely beneficial. From a clinical perspective, radiologists have detected three forms of
COVID-19-related abnormalities in CT images of lungs that used ground glass opacification
(GGO), consolidation and pleural effusion [2,4].
A number of computational techniques are used to recognize different human diseases
by using medical images of the lungs and chest [2,3]. For the accurate recognition of
anomalies in medical images, deep learning models have highlighted a standard strategy to
build the network’s capacity for the simulation of higher-order records. Similarly, tumors (in
the infected area) are the clear target of deep-learning-based medical segmentation systems.
From one instance in [5], the authors attempted to assess a tumor in the brain with MRI
images by using the combination of U-Net along with SegNet, achieving a 0.99 accuracy.
Similarly, [6] proposed DL structures that were used to develop a lung cancer screening
tool to reduce the false positive rate in CT scan images. CT scan images are also used for
breast tumors detection with image segmentation techniques [7], with a mean accuracy
of 0.90 with generative adversarial networks (GANs) and convolutional neural networks
(CNNs). For the robust diagnosis of human diseases, different researchers attempted to
segment the medical image by highlighting human body parts, such as kidneys, with
images segmented in [8]; the lungs, with images segmented in [9,10]; the liver, with images
segmented in [11]; the brain, with images segmented in [12,13]; and temporal bones, with
images segmented in [14].
The literature further delivers the various clinical decision support platforms, which
used the DL-based medical image segmentation technique very efficiently in, for example,
breast cancer classification [15,16]. Similarly, few studies provide a better solution for patch
classification in CT scan images [17,18]. Along with the diagnosis of different diseases
via machine learning techniques, many research initiatives have also been carried out
for the identification of COVID-19 by employing deep learning models in X-Ray and CT
scan images and obtaining substantial findings [19,20]. For instance, in [18], a stacked
sparse autoencoder (SSAE) model for the efficient segmentation of CT scans with the
identification of foreground and background patches was conducted. Furthermore, in [20],
extensive experiments were performed on CT scan images for the accurate classification
of COVID-19 through the encoder–decoder convolution neural networks (ED-CNNs),
U-Net, and the Feature Pyramid Network (FPN). Conversely, for the early diagnosis of
COVID-19, semantically segmenting X-Ray and CT scans is still an open window for
medical computational research communities [21,22].
Different computational researchers have explored a variety of DL structures to detect
COVID-19-infected individuals using medical images of X-Ray and CT scans. For the
accurate detection of COVID-19, one study used a CNN by using X-Ray images as an
input to create a binary classifier [23] and a multi-class classifier, achieving a 0.98 accuracy
in binary and a 0.87 accuracy in multi-class sections [24,25]. Similarly, another research
work used the Xception and ResNet50V2 models to recognize the COVID-19 cases in CT
images, with 0.99 accuracy [26]. Furthermore, recent studies [27–29] employed different DL
structures, especially CNN, on medical images (CT Scan and X-Rays) and found accuracy
scores ranging from 83% to 98%.
Recently, a few other attempts to identify COVID-19 cases in the segmenting of medi-
cal images of COVID-19 patients have been reported by using X-Rays and CT scans [30,31].
Implementing semantically segmentations would be a critical component of any health-
care system that prioritizes the patients based on severity [32]. Along with the patient’s
condition, it would diagnose the infection and its intensity early. In one instance [32], a
deep CNN was used as a binary segment, and it was compared to different U-Net struc-
tures. In another instance [32], the authors obtained a 0.73 dice value, 0.75 sensitivity,
and 0.73 precision value for the detection of COVID-19. Another study [24] highlights
different applications for a binary segmentation tool for the detection of COVID-19 and
obtained fair efficiency parameters including a Dice value of 78%, a precision of 86%, and a
Diagnostics 2023, 13, 1783 3 of 15

sensitivity of 94%. Using binary segmentation tools, the authors in [33] used a U-Net along
with a fully convolutional network model (FCN); their attempt fared well in terms of effi-
ciency parameters (precision and accuracy) but not so well in Dice parameters. Moreover,
Ref. [34] describes the creation of two unique deep neural network structures (Inf-Net and
Semi-Inf-Net) for segmenting infected areas, as well as segmenting GGO and consolidation.
One of the most serious difficulties in COVID-19 diagnosing difficulties is the lack
of synchronized patterns in data [35,36]. Another constraint of the pattern recognition
of COVID-19 is the restricted availability of specialists to segment CT images [37]. An
increasing number of experts can validate the accuracy of the retrieved lesions’ boundaries,
which might eventually lead to an improvement in model accuracy. Therefore, in the
presence of these hardware constraints, it is also difficult to test and analyze the effect of
altering various parameters and hyperparameters on a neural model’s performance [38].
In this article, we proposed a new and robust method for the detection of COVID-19
using limited data from lung CT scan images. The customized limited dataset of 321 CT
scan images is prepared using the public SARS-CoV-2 CT Scan dataset [39] after the pre-
processing stage. The limited customized dataset includes 221 CT scan ground truth images
and 100 CT images. The primary objective of this study is to build a model that is well
trained for the robust accurate detection of COVID-19 images by using limited data sources.
The major contributions of this research work are summarized as follows:
1. This research work delivered the accurate detection and automated segmentation of
COVID-19 infected areas in lung CT scan images.
2. The proposed CRV-NET architecture is the modified version of generic U-Net architecture
that has included the weight pruning approach on both encoder and decoder sides for
robust COVID-19 detection using limited data sources of lung CT scan images.
3. The proposed CRV-NET architecture is compared to different state-of-the-art variants
of U-Net architectures in terms of accuracy and robustness (least time complexity) for
COVID-19 detection.
4. This study delivers the future concept of federated learning-based CRV-NET for
robust COVID-19 detection in intra-patient hospital cases.
The rest of the article is organized into a few sections. Section 2 highlights the method-
ology, which is further classified into four subsections: Section 2.1 belongs to deep CNN
based classification, Section 2.2 highlights the general flow of the pre-processing stage, and
Section 2.3 covers the operation of generic U-Net architecture in terms of the detection of
COVID-19, and Section 2.4 discusses the COVID-19 detection workflow using the proposed
architecture, which is a modified version of generic U-architecture. The Results are thor-
oughly discussed in Section 3 and highlight the significance of the proposed architecture in
the context of effectiveness and robustness. Furthermore, state-of-the-art comparisons are
also performed in Section 3. The robustness of the proposed architecture is highlighted in
Section 4, Discussion. Finally, Section 5 highlighted the conclusion of this study.

2. Materials and Methods

This study proposed a unique method for the accurate and robust detection of
COVID-19. A proposed CRV-NET, generic U-Net and CNN classifier are used to detect
and segment images of COVID-19. According to Figure 1, a public dataset, SARS-CoV-2 CT
Scan dataset is used to formulate the customized dataset. The customized dataset is a com-
bination of ground truth images after pre-processing phase (labeled by radiologists) and
general lung CT scan images. In the pre-processed stage, binary operations are performed
for the removal of unnecessary minor details in ground truth CT scan images. According
to Figure 1, the unlabeled CT scan images and ground truth CT scan images are processed
in two different paths. At first, the CT scan images of the public dataset are processed by a
well-tuned CNN model for classifying COVID-19. Concurrently, after pre-processing stage,
the 221 CT scan ground truth images and 100 raw CT scan images of customized dataset
are employed for training and testing with different architectures of U-Net (U-Net, U-Net++
and proposed CRV-NET). Thus, after the classification of COVID-19 by the well-tuned
 customized dataset are employed for training and testing with different architectures of
U-Net (U-Net, U-Net++ and proposed CRV-NET). Thus, after the classification of COVID-
19 by the well-tuned CNN model, these classified images are processed using U-Net, U-
Net++
Diagnostics 2023,and proposed CRV-NET for accurate and robust segmentation of COVID-19 to
13, 1783 4 of 15

measure the intensity of the virus.

Figure 1 highlights the work scheme for robust and accurate detection of COVID-19.
CNN model, these classified images are processed using U-Net, U-Net++ and proposed
It is a composition ofCRV-NET
two main parts, namely, the CNN-based classification and pre-pro-
for accurate and robust segmentation of COVID-19 to measure the intensity of
cessing stage along with image segmentation using different architectures of U-Net.
the virus.

Figure 1. Proposed workflow

Figure 1. for COVID-19
Proposed detection
workflow anddetection
for COVID-19 intensity
andmeasurement.
intensity measurement.

Figure 1 highlights the work scheme for robust and accurate detection of COVID-19.
2.1. Deep-CNN-BasedItClassification
is a composition of two main parts, namely, the CNN-based classification and pre-
In the classification phase of the proposed
processing stage along with imagemethod, a CNN-based
segmentation classification
using different architectures ofdeliv-
U-Net.
ers the efficient and 2.1.
robust detection Classification
Deep-CNN-Based of COVID-19. In the proposed method, a deep con-
volutional neural network model
In the is employed
classification phase of that contains
the proposed 16 layers.
method, Furthermore,
a CNN-based thede-
classification
structure of CNN with kernel
livers size 3and
the efficient × 3 robust
and filters 16, of
detection 32,COVID-19.
64, 128 and 256proposed
In the are employed
method, afordeep
classification. Figure 2 presents the layer structure of the CNN model that is used in thethe
convolutional neural network model is employed that contains 16 layers. Furthermore,
structure of CNN with kernel size 3 × 3 and filters 16, 32, 64, 128 and 256 are employed for
classification of COVID-19.
classification. Figure 2 presents the layer structure of the CNN model that is used in the
classification of COVID-19.
Diagnostics 2023,
Diagnostics 2023,12,
13,x1783
FOR PEER REVIEW 5 of 15 5 of 1

Figure2.2.Layered
Figure Layered view
view of deep
of deep convolution
convolution neural
neural network.
network.

According
According toto
Figure
Figure2 above,
2 above,the structure of deep
the structure CNN is
of deep highlighted
CNN as first using
is highlighted as first using
the 3 convolution layers, a 3 × 3 kernel and 16 filters. After that, a
the 3 convolution layers, a 3 × 3 kernel and 16 filters. After that, a max poolmax pool layerlayer
with with 2 ×
2 × 2 strides, 3 more convolution layers, 3 × 3 kernel and 32 filters are employed. Further
2 strides, 3 more convolution layers, 3 × 3 kernel and 32 filters are employed. Further in
in the depth of the structure, add another max pool layer with strides 2 × 2 and 3 more
the depth oflayers
convolution the structure,
along withadd 128 another max pool
filters. Similarly, thelayer withpool
final max strides
layer2 and
× 2 and 3 more con
the final
volution layers
convolution layeralong with
with 256 128are
filters filters. Similarly,
further connectedthewith
final2 hidden
max pool layer
layers and
with the final con
230,400
and 460,800 perceptrons. Lastly, the two output layers are classified by the existence230,400
volution layer with 256 filters are further connected with 2 hidden layers with or and
460,800of
absence perceptrons.
COVID-19 inLastly,
CT scan the two output layers are classified by the existence or absence
images.
of COVID-19 in CT scan images.
2.2. Pre-Processing
In pre-processing stage, according to Figure 1, the ground truth images are labeled
2.2. Pre-Processing
by radiologists. At the start of this stage, set the threshold point with the help of rescaling
In pre-processing
intensity, which ranges from stage,
92 toaccording
98. Next,to Figure
assign the1, the ground
binary truth
operations, images the
including are labeled
binary opening (0.5 > set 1 value) and the binary closing (0.5 < set 0 value), and removerescaling
by radiologists. At the start of this stage, set the threshold point with the help of
intensity,
small objectswhich ranges
with dice rangefrom
of 2 to924 from
to 98.theNext, assign
defined image.theThebinary
definedoperations, including
setting of the pre- the
binary opening
processing (0.5 a>vital
stage plays set 1role
value)
in theand
nextthe binary
two-way closing (0.5
classification for<the
setidentification
0 value), and of remove
the infected
small objectsarea (thedice
with detection
rangeofof COVID-19)
2 to 4 from in athe
CTdefined
scan of the lungs.The defined setting of the
image.
After the pre-processing phase, the customized small-scale
pre-processing stage plays a vital role in the next two-way classification dataset of 331 forimages is
the identifica
employed for image segmentation to highlight the infected areas
tion of the infected area (the detection of COVID-19) in a CT scan of the lungs. in CT scan images. In
the segmentation phase, the COVID-19 classified images are processed with the proposed
After the pre-processing phase, the customized small-scale dataset of 331 images i
CRV-NET architecture and generic U-Net for segmentation of the COVID-19 infected areas to
employed
highlight thefor imageofsegmentation
intensity to highlightofthe
the virus. The combination infected
proposed areas in
CRV-NET, CT scan
generic U-Net,images. In
and U-Net++ are used to validate the segmentation results of COVID-19 in CT scan images.proposed
the segmentation phase, the COVID-19 classified images are processed with the
CRV-NET architecture and generic U-Net for segmentation of the COVID-19 infected ar
2.3.
eas U-Net Architecture
to highlight the intensity of the virus. The combination of proposed CRV-NET, generi
U-Net
U-Net, andis U-Net++
a CNN-based architecture
are used that plays
to validate a vital role in tumor
the segmentation detection
results in MRI and
of COVID-19 in CT scan
CT scan images [40,41]. In U-Net, performing the image segmentation for detection of the
images.
infected area without losing the feature mapping is the most significant part. For semantic
segmentation, especially in the CT scan and MRI images, different U-Net structures are
2.3. U-Net Architecture
used [5,40]. CNN in U-Net is focused on learning representations of data and hierarchical
U-Net
feature is a For
learning. CNN-based architecture
feature extraction, CNNthat playsan
employs a vital role in tumor
arrangement detection
of several layers in MR
and
of CT scanprocessing
nonlinear images [40,41]. In U-Net,
identities, and the performing
output of each thesequential
image segmentation
layer becomesfor thedetection
of the infected area without losing the feature mapping is the most significant part. Fo
input of the next one, which aids in data abstraction. The standard U-Net design is shown
in Figure 3 segmentation,
semantic [41]. especially in the CT scan and MRI images, different U-Net struc
tures are used [5,40]. CNN in U-Net is focused on learning representations of data and
hierarchical feature learning. For feature extraction, CNN employs an arrangement of sev
eral layers of nonlinear processing identities, and the output of each sequential layer be
comes the input of the next one, which aids in data abstraction. The standard U-Net design
is shown in Figure 3 [41].
 Diagnostics 2023, 12, x FOR PEER REVIEW
Diagnostics 2023, 13, 1783 6 of 15

Figure 3. Generic U-Net Architecture [41].

Figure 3. Generic U-Net Architecture [41].
The standard U-Net design in Figure 3 highlights the dense structure. The denser part
of U-Net is the 32The standard
× 32 U-Netresolution)
pixels (lowest design inalong
Figure
with3 highlights the dense
a multi-channel structure.
function map The
part of U-Net
that is represented is the
in the blue 32 Furthermore,
box. × 32 pixels (lowest resolution)
the number along is
of channels with a multi-channel
shown on the fu
map boxes
box’s end, white that isshow
represented in maps
the feature the blue
thatbox.
haveFurthermore, thearrows
been copied and number of channels is
represent
on the
the upsampling andbox’s end, whiteoperations.
downsampling boxes show the feature maps that have been copied and
represent the upsampling and downsampling operations.
2.4. Proposed CRV-NET Architecture
The proposed CRV-Net
2.4. Proposed is the Architecture
CRV-NET modified version of generic U-Net, which works on
limited data with a fair accuracy level. In CRV-NET, the CNN model is operated in
The proposed CRV-Net is the modified version of generic U-Net, which wo
upsampling and downsampling operations in a similar way to U-Net. The reduction in
limited data with a fair accuracy level. In CRV-NET, the CNN model is operated
computational cost in terms of minimum time consumption and accurate detection of
sampling and downsampling operations in a similar way to U-Net. The reduction i
COVID-19 is the primary objective of the proposed CRV-NET. Therefore, the CRV-NET
increases the putational
level of COV costand
in terms
pooling of minimum
layers by usingtimeupconsumption and
to 2048 filters. accurate detection
Additionally, the of C
19 is the primary objective of the proposed CRV-NET.
weight dropout scheme follows in each operational phase of CRV-NET. Figure 3 shows the Therefore, the CRV-NET in
the level of COV and
structural view of the proposed CRV-NET. pooling layers by using up to 2048 filters. Additionally, the
CRV-NET dropout schemeisfollows
architecture dividedinintoeachtwooperational phaseparts,
state-of-the-art of CRV-NET. Figure 3 shows th
namely, downsam-
tural view of
pling and upsampling. the proposed
In Figure 4, the leftCRV-NET.
side highlights the downsampling and is known
as a contracting CRV-NET
path. The architecture is dividedtointo
image size continues two in
reduce state-of-the-art
downsampling parts,
withnamely,
the dow
filter’s increment.
pling andThe number of features
upsampling. would 4,
In Figure bethe
doubled everyhighlights
left side time in thethe
contracting
downsampling
direction. The process
known as aofcontracting
downsampling path. is The
accomplished
image size bycontinues
repeating to two 3 × 3 convolu-
reduce in downsamplin
tions with thetheusage of rectified
filter’s increment. linear
Theunit (ReLU)ofactivation
number features function
would be and max-pooling
doubled everybytime in th
setting the kernel size of 2 × 2 and a stretch of 2. Conversely, the upsampling
tracting direction. The process of downsampling is accomplished by repeating tw represented
on the rightconvolutions
side of Figure 3 isthe
with alsousage
known as an expanding
of rectified linear unit path. In upsampling,
(ReLU) the
activation function an
size of the image is gradually enhanced by decreasing the size of the filter. The same
pooling by setting the kernel size of 2 × 2 and a stretch of 2. Conversely, the upsa
configuration is used in the upsampling process as was used in the downsampling, namely,
represented on the right side of Figure 3 is also known as an expanding path. In u
two 3 × 3 convolutions, followed by a ReLU activation function, except that the layers in
pling, the size of the image is gradually enhanced by decreasing the size of the filt
the expansive path are concatenated with the corresponding layers in the contracting path.
At the final same
stage, configuration
a 1 × 1 convolutionis used is in
usedtheinupsampling
the last layerprocess
to map as was
with used inclass.
a desired the downsam
namely, two 3 × 3 convolutions, followed by a ReLU activation function, except t
layers in the expansive path are concatenated with the corresponding layers in th
tracting path. At the final stage, a 1 × 1 convolution is used in the last layer to map
desired class.
 The CRV-NET architecture was trained on 321 images with the ground truth images
and minimum epoch values that reflected the least execution time (computation cost). This
Diagnostics 2023, 13, 1783
optimal training process of CRV-NET plays a vital role in the robust detection of COVID- 7 of 15
19. In addition, the padding is equal to that of the input image in order to resize the image
after upsampling.

Figure 4. Proposed CRV-NET architecture.

Figure 4. Proposed CRV-NET architecture.
The CRV-NET architecture was trained on 321 images with the ground truth images
Moreover, Table 1 highlighted the hyperparameters list used in training the proposed
and minimum epoch values that reflected the least execution time (computation cost). This
CRV-NET. In CRV-NET, Adam is used, which is the replacement optimization algorithm
optimal training process of CRV-NET plays a vital role in the robust detection of COVID-19.
for stochastic gradient descent for training CNN models. Therefore, CRV-NET is the pref-
In addition, the padding is equal to that of the input image in order to resize the image
erable architecture for the robust detection of COVID-19 in limited data sources.
after upsampling.
Moreover, Table 1 highlighted the hyperparameters list used in training the proposed
Table 1. Hyper-parameter values used in model training (CRV-NET and U-Net).
CRV-NET. In CRV-NET, Adam is used, which is the replacement optimization algorithm for
stochastic gradient descent forCRV-NET Hyper-Parameters
training CNN models. Therefore, CRV-NET is the preferable
architecture for the robust detection
Batch size of COVID-19 in limited data sources. 16
Train Validation Test split 0.7, 0.1, 0.2
Table 1. Hyper-parameter values used in model training (CRV-NET and U-Net).
Optimizer Adam
Input shape
CRV-NET Hyper-Parameters 128 × 128 × 3
Dropout
Batch size 16
0.2
Zoom range 0.2
Train Validation Test split 0.7, 0.1, 0.2
U-Net Hyper-parameters
Optimizer Adam
Batch size 64
Input shape
Train Validation Test split 128 × 128 × 3 0.1
0.8, 0.1,
Optimizer
Dropout 0.2 Adam
Input
Zoom range shape 512 × 512 × 1
0.2
U-Net Hyper-parameters
Along with accuracy, the short computational time of CRV-NET is the key parameter
Batch
in the diagnosis of size
COVID-19. 64
In CRV-NET, the weight pruning policy in each stage of
Train Validation Test split 0.8, 0.1, 0.1
Optimizer Adam
Input shape 512 × 512 × 1

Along with accuracy, the short computational time of CRV-NET is the key parameter in the
diagnosis of COVID-19. In CRV-NET, the weight pruning policy in each stage of upsampling
Diagnostics 2023, 12, x FOR PEER REVIEW

Diagnostics 2023, 13, 1783 8 of 15

upsampling (encoder), downsampling (decoder) and dense convolution layers p

key role downsampling
(encoder), in the optimization
(decoder)ofand
accurate results. Figure
dense convolution layers 5 presents
plays the in
a key role component’
the
of CRV-NET.
optimization of accurate results. Figure 5 presents the component’s view of CRV-NET.

Figure5. 5.
Figure Basic
Basic components
components of CRV-NET
of CRV-NET architecture.
architecture.

For performance improvement of convolution neural structures, in terms of the combi-

For performance improvement of convolution neural structures, in terms of th
nation of activation functions, the input layer uses the ReLU activation function, and the
bination
output layerofuses
activation functions,
the sigmoid function.the
Theinput
reasonlayer usesthe
for using the ReLU
ReLU activation
function function, a
over other
output layer
activation usesisthe
functions thatsigmoid
it does notfunction.
activate allThe reason at
the neurons fortheusing
same the
time.ReLU function
Equation (1) ove
shows the standard discrete nature of the ReLU activation function:
activation functions is that it does not activate all the neurons at the same time. Eq
(1) shows the standard discretef (nature of(0,the
x ) = max x ) ReLU activation function: (1)
Similarly, at the output layer, the sigmoid 𝑓 𝑥 function
= 𝑚𝑎𝑥is0,used 𝑥 to perform the binary
classification. The virus-infected area is quickly identified via binary classification. The
range ofSimilarly,
the sigmoidat the output
function is 0.0layer,
to 1.0,the
andsigmoid
Equation function is used
(2) shows the to perform
non-linear the binar
nature of
sification.
the The virus-infected area is quickly identified via binary classification. The
sigmoid function:
1
of the sigmoid function is 0.0 to f ( x1.0,
) = and −( Equation (2) shows the non-linear (2) nature
1 + e x)
sigmoid
In the function:
CRV-NET, class 0 belongs to the downsample because the majority of the pixels
belong to it, and it does little to boost distinguishing ability. 1 In the CRV-NET, only select
𝑓 𝑥 =
the slices in the image that contain infected-area pixels in the downsampling process. As
1+ 𝑒
discussed above, the downsampling process continuously shrinks the image size and
In the
increases the CRV-NET, class
size of the filter to02048.
belongs to the downsample
Furthermore, in each step of because the majority
downsampling, the of the
image size reduces by up to half when we are applying filters.
belong to it, and it does little to boost distinguishing ability. In the CRV-NET, only
Similarly, the upsampling process is executed in a reverse way, such as reducing
the slices in the image that contain infected-area pixels in the downsampling proc
the number of filters by enlarging the image size and maintaining a connection with
discussed above,
downsampling the
values todownsampling
obtain features. At process
the lastcontinuously
layer, the final shrinks the image
output image was size a
creases the size of the filter to 2048. Furthermore, in each
converted into a binary image, and the sigmoid function was used to obtain a more step of downsampling, the
accurate value of
size reduces byCOVID-19
up to half detection.
when we are applying filters.
Similarly, the upsampling process is executed in a reverse way, such as reduc
number of filters by enlarging the image size and maintaining a connection
downsampling values to obtain features. At the last layer, the final output image w
verted into a binary image, and the sigmoid function was used to obtain a more ac
value of COVID-19 detection.
Diagnostics 2023, 13, 1783 9 of 15

3. Results
To evaluate the performance of the proposed CRV-NET, efficiency parameters, includ-
ing mean accuracy, sensitivity and the epoch ratio, are calculated along with the dice score.
A dice score is computed for the identification of the feature recognition of the infected
area. In feature recognition, the dice score coefficient is used to measure the accuracy
between the ground truth images and segmented images. The dice coefficient is a fixed
similarity measure function that is typically used to determine the similarity of two samples
(ground truth images and segmented images), with a value range of [0, 1]. Furthermore,
dice loss will be smaller in the dice coefficient in the case of two samples that highlight a
high similarity index.
2| A ∩ B |
Dice Loss( A, B) = 1 − (3)
| A| + | B|
Equation (3) highlights the standard form of dice loss in which set A belongs to the
ground truth images and set B represents the segmented images. The dice loss ratio in
Equation (3) highlights the ratio of feature loss between the ground truth images of CT
scans and the segmented images of CT scans. Dice loss is the efficiency measurement gauge
of CRV-NET. The minimum ratio of dice loss highlights the least feature mapping loss
between the ground truth images and the segmented images. Conversely, if the dice loss
ratio is large, it reflects a huge feature mapping loss.
Additionally, to cross-validate the performance of the proposed CRV-NET, it is rec-
ommended to measure the statistical parameters, including TP (true positive), TN (true
negative), FP (false positive) and FN (false negative), and then, with the help of these
parameters, compute the efficiency gages in terms of specificity (Sp), sensitivity (Se) and
accuracy (Acc). Equation (4) represents the specificity that delivers the correct analysis of a
person not having a tumor:
TN
Sp(%) = × 100 (4)
TN + FP
Similarly, Equation (5) highlights the sensitivity (Se) that highlights the correct analysis
of a person having a disease:
TP
Se(%) = × 100 (5)
TP + FN
Equation (6) represents the accuracy factor (Acc) of the proposed method in terms of
accurate classification:
TP + TN
Acc(%) = × 100 (6)
TP + TN + FP + FN
The proposed CRV-NET architecture was validated on the limited dataset (321 images)
and the minimum epoch ratio. The proposed architecture obtained an accuracy of 96.22%,
which is better than the state-of-the-art generic U-net, and the least computation complexity
in comparison (less computation cost in terms of execution time). Therefore, according to
Figure 1, the second track of the work scheme is executed that contains the generic U-Net
architecture and proposed CRV-NET architecture.
According to Figure 1, after the pre-processing stage on the dataset, the execution
of the generic U-Net and proposed CRV-NET architectures are operated on lung CT scan
images and ground truth images. The effectiveness of generic U-NET architecture and
the proposed CRV-NET architecture is validated by efficiency gages, including the epoch
ratio, the accuracy of COVID-19 detection, specificity and sensitivity. The robustness factor
in the context of computational cost is also a core parameter in the accurate detection of
COVID-19. Hence, the epoch ratio is measured as a computational cost parameter for
COVID-19 detection. Figure 6 presents the smallest epoch ratio of the proposed CRV-NET
in terms of accuracy and loss for the detection of COVID-19.
Diagnostics
Diagnostics2023,
2023,12,
13,x1783
FOR PEER REVIEW 10 10
ofof1615

Figure 6. Epoch incremental variation in accuracy and loss parameters of CRV-NET.

Figure 6 shows the incremental variation of the CRV-NET epoch ratio. Figure 6a,b
highlight the accuracy and loss parameters of CRV-NET in the COVID-19 detection pro-
cess. Figure 6a presents the least increment variation in epoch values, enhancing the accu-
racy ratio up to 96% with limited input data, which means the model is well trained in
CRV-NET architecture. Similarly, Figure 6b highlights the gradual decline of the loss pa-
rameter with a similar pattern of the epoch incremental ratio. Figure 6b also represents
the model that is fine tuned in CRV-NET. Figure 7 represents the findings for the robust
detection of COVID-19 via segmentation and detection.
Epochincremental
Figure6.6.Epoch
Figure incrementalvariation
variationininaccuracy
accuracyand
andloss
lossparameters
parametersofofCRV-NET.
CRV-NET.
In a sample of four images, Figure 7a presents the CT scan images, which are seg-
mented and 6masked
Figure showsthe in Figure
the 7b after
incremental the implementation
variation of image processing
Figuretech-
Figure
niques, 6 shows incremental variation ofofthe
theCRV-NET
CRV-NET epochratio.
epoch ratio.Figure 6a,b
6a,b
highlight the accuracy and loss parameters of CRV-NET in the COVID-19 detectionclosing.
including the removal of small objects, the binary opening and the binary process.
highlight
Finally, the accuracy
Figure 7c shows and
theloss parameters
detection of CRV-NET in theinyellow
the COVID-19 detection pro-
Figure 6a presents the least incrementofvariation
COVID-19 in epoch values, part of the
enhancing CT
thescan im-
accuracy
cess.
ages. Figure 6a presents the least increment variation in epoch values, enhancing the accu-
ratio According
up to 96% with to Figure
limited7,input
image 3 shows
data, whichthemeansworstthedice
modelscore after
is well the execution
trained in CRV-NETof
racy ratio up
CRV-NET to 96%
because thewith limited
ground truthinput
doesdata,
not which accurately
match means the with
model theis detected
well trained in
result.
architecture. Similarly, Figure 6b highlights the gradual decline of the loss parameter
CRV-NET
However, architecture.
the proposed Similarly, Figure 6b3highlights the gradual decline ofsegments
the loss pa-
with a similar pattern ofCRV-NET
the epochimages
incrementalandratio.
image 4 also
Figure 6bhighlight the
also represents of
the model
rameter
infected with
areas a similar
that are pattern
not of
present the
in epoch
the incremental
ground truth ratio.
image. Figure
Similarly, 6b also
Figure represents
8 presents
that is fine tuned in CRV-NET. Figure 7 represents the findings for the robust detection of
the
the model that is fine tuned
dice segmentation result in CRV-NET.
with Figure 7 represents the findings for the robust
COVID-19 via segmentation and the bar chart.
detection.
detection of COVID-19 via segmentation and detection.
In a sample of four images, Figure 7a presents the CT scan images, which are seg-
mented and masked in Figure 7b after the implementation of image processing tech-
niques, including the removal of small objects, the binary opening and the binary closing.
Finally, Figure 7c shows the detection of COVID-19 in the yellow part of the CT scan im-
ages. According to Figure 7, image 3 shows the worst dice score after the execution of
CRV-NET because the ground truth does not match accurately with the detected result.
However, the proposed CRV-NET images 3 and image 4 also highlight the segments of
infected areas that are not present in the ground truth image. Similarly, Figure 8 presents
the dice segmentation result with the bar chart.

(a) (b) (c)

Figure 7. Detection of COVID-19 using CRV-NET. (a) Raw CT scan images of the public dataset;
(b) raw CT scan images after image processing is implemented in the context of the removal of small
objects, the binary opening, and the binary closing; (c) raw CT scan images with the yellow region
highlighting the detected area of COVID-19.

In a sample of four images, Figure 7a presents the CT scan images, which are seg-
mented and masked in Figure 7b after the implementation of image processing techniques,
including the removal of small objects, the binary opening and the binary closing. Finally,
Figure 7c shows the detection of COVID-19 in the yellow part of the CT scan images.
(a) (b) (c)
According to Figure 7, image 3 shows the worst dice score after the execution of CRV-NET
Diagnostics 2023, 13, 1783 11 of 15
Diagnostics 2023, 12, x FOR PEER REVIEW 11 of 16

because the ground truth does not match accurately with the detected result. However,
Figure 7. Detection of COVID-19 using CRV-NET. (a) Raw CT scan images of the public dataset; (b)
the
raw proposed CRV-NET
CT scan images images
after image 3 and image
processing 4 also highlight
is implemented the segments
in the context of infected
of the removal of small
areas that are not present in the ground truth image. Similarly, Figure 8 presents
objects, the binary opening, and the binary closing; (c) raw CT scan images with the yellowtheregion
dice
segmentation result with the bar chart.
highlighting the detected area of COVID-19.

Figure 8.
Figure 8. Dice
Dice loss
loss ratio
ratio of
of COVID-19
COVID-19 detected
detected images.
images.

Figure 8 shows
showsthe
thedice
dicesimilarity
similarityindex
indexonon four
four COVID-19
COVID-19 detected
detected images.
images. The
The dice
dice similarity index covers the exact ratio of COVID-19 infected areas
similarity index covers the exact ratio of COVID-19 infected areas in images. in images. Figure 88
shows
shows the
the 64.45%
64.45% dice similarity index of COVID-19 detection in image 3, which is easily
observed
observed in the third section
section of
of Figure
Figure 8.
8. Similarly, images 2 and image 4 represent
represent the
the
63.38%
63.38% dice
dice score,
score, and
and image
image 11 highlights
highlights the
the 61.21%
61.21% dice
dice score.
score.

4.
4. Discussion
Discussion
In
In operations
operations ofof the
the proposed
proposed work
work scheme,
scheme, the
the trained
trained model
model accuracy
accuracy of
of the
the generic
generic
U-Net is 95.81% using the ground truth CT scan image dataset, and the validation
U-Net is 95.81% using the ground truth CT scan image dataset, and the validation accu- accuracy
using the CT
racy using thescan image
CT scan dataset
image is 94.50%.
dataset In the
is 94.50%. Inproposed CRV-NET,
the proposed the accuracy
CRV-NET, has
the accuracy
been increased with the trained model accuracy reaching up to 96.67% by using
has been increased with the trained model accuracy reaching up to 96.67% by using the the ground
truth CT scan image dataset, and the validation accuracy is 96% when using CT scan images
ground truth CT scan image dataset, and the validation accuracy is 96% when using CT
dataset. The first level of comparison highlights that a CRV-NET is a step higher than the
scan images dataset. The first level of comparison highlights that a CRV-NET is a step
generic U-Net in terms of accuracy. For the cross-validation of the proposed CRV-NET
higher than the generic U-Net in terms of accuracy. For the cross-validation of the pro-
architecture, the state-of-the-art comparison is performed in Table 2. For the state-of-the-art
posed CRV-NET architecture, the state-of-the-art comparison is performed in Table 2. For
comparison, different efficiency parameters, including sensitivity, accuracy, epoch value
the state-of-the-art comparison, different efficiency parameters, including sensitivity, ac-
and the dice coefficient, are considered.
curacy, epoch value and the dice coefficient, are considered.
Table 2. State-of-the-art comparison of proposed CRV-NET.
Table 2. State-of-the-art comparison of proposed CRV-NET.
Ref
Ref Method Dice Dice
Method DatasetDataset
Size Size Sensitivity Specificity
Sensitivity Specificity Accuracy
Accuracy Epoch
Epoch
[42]
[42] U-NET
U-NET 83.10% 83.10% 473 CT scan images
473 CT scan images 86.70%
86.70% 99.00%
99.00% 50
50
[40] SegNet 74.90% 21,658 CT scan images 94.50% 95.40% 95% 160
[40]
[43] SegNet
COVID-NET 74.90%13,975 21,658
CT scanCT scan images 80%
images 94.50% 95.40% 95%
93.3% 160
22
[44]
[43] U-Net
COVID-NET 92% 639 CT scan images
13,975 CT scan images 81.8%80% 95.2% 94.0%
93.3% 100
22
[45] 3D U-Net 76.1% 20 CT scan images 95.56% 99.8% 95.56% 312
[44] U-Net 92% 639 CT scan images 81.8% 95.2% 94.0% 100
[46] CovidDenseNet 4173 CT scan images 86.14% 95.46% 95.76% 150
Purposed[45]
Method 3D U-Net 64.45% 76.1% 331 CT scan
CRV-NET 20 CT scan images
images 95.56%
96.67% 99.8%
90% 95.56%
96.67% 312
7
[46] CovidDenseNet 4173 CT scan images 86.14% 95.46% 95.76% 150
Purposed Method CRV-NET Table 2 highlights
64.45% 331 the larger
CT scan dataset of96.67%
images lung CT scan90% images in 96.67%
previous methods,
7
showing that the current dice score and specificity parameters are high for measuring the
intensity of COVID-19 detection compared to the accuracy and sensitivity parameters of
Tablemedical
previous 2 highlights thediagnostic
imaging larger dataset of lung
studies CTThe
[47,48]. scanproposed
images in previousdominates
CRV-NET methods,
showing that the current dice score and specificity parameters are high for measuring
in terms of efficiency parameters such as the accuracy of COVID-19 detection, sensitivity the
intensity of COVID-19 detection compared to the accuracy and sensitivity
and robustness (low epoch value). According to the target of COVID-19 detection in theparameters of
previous medical imaging diagnostic studies [47,48]. The proposed CRV-NET
limited data source, the high accuracy and sensitivity percentage in the limited dataset dominates
in terms
size of efficiency
highlight that theparameters such as the
proposed CRV-NET is accuracy
better than ofprevious
COVID-19 detection,
U-Net sensitivity
architecture. Ad-
and robustness
ditionally, (low epoch
the small epoch value
value).
ofAccording
the proposed to the target ofhighlights
CRV-NET COVID-19 thedetection in the
significance of
limited data source, the high accuracy and sensitivity percentage in the limited
the low computation cost compared to previous studies that used different U-Net archi- dataset size
tecture.
Diagnostics 2023, 13, 1783 12 of 15

highlight that the proposed CRV-NET is better than previous U-Net architecture. Addition-
ally,
Diagnostics 2023, 12, x FOR PEER REVIEW the small epoch value of the proposed CRV-NET highlights the significance of12 theoflow
16
computation cost compared to previous studies that used different U-Net architecture.
Furthermore, U-Net++ is another refactored version of U-Net architecture that delivers
optimized accurateU-Net++
Furthermore, results in CT scan refactored
is another images for version
the diagnosis of different
of U-Net diseases
architecture [41,49]
that deliv-
U-Net++ is a composition of nest U-Nets and operates with node pruning policy via
ers optimized accurate results in CT scan images for the diagnosis of different diseases nested
and dense skip connections between upsampling and downsampling. The effectiveness of
[41,49] U-Net++ is a composition of nest U-Nets and operates with node pruning policy
U-Net++ is validated in the test environment of CRV-NET. Table 3 shows the performance
via nested and dense skip connections between upsampling and downsampling. The ef-
comparison of CRV-NET and standard U-Net++ using the customized dataset of 331 CT
fectiveness of U-Net++ is validated in the test environment of CRV-NET. Table 3 shows
scan images.
the performance comparison of CRV-NET and standard U-Net++ using the customized
dataset of 331 CT scan images.
Table 3. CRV-NET compared to U-Net++ using the customized dataset.
Table 3. CRV-NET compared to U-Net++ using the customized dataset.
Training Loss Training Accuracy Test Loss Test Accuracy
Epoch
Training Loss
U-Net++ CRV-NET Training
U-Net++ Accuracy
CRV-NET Test LossCRV-NET
U-Net++ Test Accuracy
U-Net++ CRV-NET
Epoch
1 U-Net++
0.6006 CRV-NET
0.5569 U-Net++
0.3994 CRV-NET
0.4431 U-Net++
0.8247 CRV-NET
0.5017 U-Net++
0.1753 CRV-NET
0.4983
1 2 0.6006
0.4745 0.5569
0.3036 0.3994
0.5255 0.4431
0.6964 0.8247
0.784 0.5017
0.4242 0.1753
0.216 0.4983
0.5758
2 0.4745 0.3036 0.5255 0.6964 0.784 0.4242 0.216 0.5758
3 0.3936 0.2046 0.6064 0.7954 0.7414 0.2599 0.2586 0.7401
3 0.3936 0.2046 0.6064 0.7954 0.7414 0.2599 0.2586 0.7401
4 0.3452 0.1544 0.6548 0.8456 0.7159 0.2522 0.2841 0.7478
4 0.3452 0.1544 0.6548 0.8456 0.7159 0.2522 0.2841 0.7478
5 0.3101 0.1427 0.6899 0.8573 0.6527 0.1751 0.3473 0.8249
5 0.3101 0.1427 0.6899 0.8573 0.6527 0.1751 0.3473 0.8249
6 6 0.2841
0.2841 0.1338
0.1338 0.7159
0.7159 0.8662
0.8662 0.5596
0.5596 0.1705
0.1705 0.4404
0.4404 0.8295
0.8295
7 7 0.2642
0.2642 0.1175
0.1175 0.7358
0.7358 0.966
0.966 0.4401
0.4401 0.236
0.236 0.5599
0.5599 0.8974
0.8974

According to Table 3, the customized dataset of 331 CT scan images is operated on

According to Table 3, the customized dataset of 331 CT scan images is operated on
both
both U-Net++and
U-Net++ andCRV-NET
CRV-NETarchitectures.
architectures.InInCRV-NET,
CRV-NET,the theminimum
minimumepoch’s
epoch’sincrement
increment
generates
generatesbetter
bettertraining
trainingdata
dataaccuracy
accuracyand
andtesting
testingdata
dataaccuracy
accuracyalong
alongwith
withminimum
minimum
loss
lossvalues.
values.Conversely,
Conversely,in inU-Net++,
U-Net++,large
largeincremental
incrementalvariation
variationininepochs
epochsdelivers
deliversthethelow
low
training and testing accuracies compared to
training and testing accuracies compared to CRV-NET.CRV-NET.
Nonetheless,
Nonetheless,the theaccurate
accurateearly
earlydiagnosis
diagnosisofofCOVID-19
COVID-19isisthe theprimary
primaryobjective
objectiveof of
modern healthcare systems, but the transformation of such early diagnoses
modern healthcare systems, but the transformation of such early diagnoses using wearable using weara-
ble gages
gages is the
is the future
future vision.To
vision. Toconcatenate
concatenatethe thefuture
future vision,
vision, the
the integration
integration of
offederated
federated
learning and CRV-NET is needed for the robust recognition of COVID-19
learning and CRV-NET is needed for the robust recognition of COVID-19 via wearable via wearable
gages.
gages.Figure
Figure 9 presents
9 presentsthethe
unique
uniquework
workscheme
schemeof federated learning
of federated for the
learning for robust de-
the robust
tection of COVID-19 that will be implemented in the
detection of COVID-19 that will be implemented in the future. future.

Figure 9. Future robust COVID-19 prediction wearable gages with embedded CRV-NET feder-
Figure 9. Future robust COVID-19 prediction wearable gages with embedded CRV-NET federated
ated learning.
learning.

A trustworthy federated learning structure will be used in the future for the detection
of COVID-19 with the integration of CRV-NET [50,51]. According to Figure 9, real-time
data streams that are fetched from standard wearable gages [52,53] are trained on local
CRV-NET, which is further tested and validated by a cloud-based Master CRV-NET and
Diagnostics 2023, 13, 1783 13 of 15

A trustworthy federated learning structure will be used in the future for the detection
of COVID-19 with the integration of CRV-NET [50,51]. According to Figure 9, real-time data
streams that are fetched from standard wearable gages [52,53] are trained on local CRV-NET,
which is further tested and validated by a cloud-based Master CRV-NET and an active reposi-
tory data center [54,55]. The robustness and accurate factors of CRV-NET will help the early
diagnosis [56,57] of COVID-19 via federated learning, which is highlighted in Figure 9.

5. Conclusions
The improvement of accuracy and robustness in detecting COVID-19 is the top clinical
requirement following the emergence of this new global virus. Previous studies have
delivered a fair contribution to the diagnosis of COVID-19, but the robustness factor is
compromised. This research work proposed a unique structure of U-Net in terms of CRV-
Net for the accurate and early diagnosis of COVID-19 with the least computational cost
(the smallest epoch ratio). The proposed CRV-NET architecture worked for the robust
detection of COVID-19, which is a refactored version of generic U-Net architecture. A
public SARS-CoV-2 CT scan dataset is used for experimental purposes and the customized
dataset. The customized dataset is used for training the model with 331 ground truth
images and 100 images. The proposed CRV-NET’s accuracy, 96.67%, is better than that
of the generic U-Net architecture. Similarly, the epoch value of CRV-NET is seven, which
is quite impressive and supportable for robustness (least computational cost). Moreover,
wearable gages will be used in the future for the early diagnosis of COVID-19 by using the
trustworthy federated learning work scheme with an embedded part for CRV-NET.

Author Contributions: U.I. and R.I. created an experimental environment and worked on the structure
of the refactored U-Net in terms of CRV-NET. U.I. and K.A.A. cross-checked the findings of CRV-NET
and created a write-up of the manuscript. K.A.A. validated the architecture scheme of CRV-NET. A.K.J.S.
financially supported the experimental setup and performed the data integrity of the different COVID-19
datasets. All authors have read and agreed to the published version of the manuscript.
Funding: This work was supported by the Deputyship for Research and Innovation, Ministry of
Education in Saudi Arabia, for funding this research work through Project Number 959.
Institutional Review Board Statement: The institutional ethical board of research cleared the pro-
ceeding of this experimental study. The authors declare that the work described has been carried out
according to the Declaration of Helsinki of the World Medical Association, revised in 2013.
Informed Consent Statement: Not applicable.
Data Availability Statement: All the datasets used in this study are publicly available; no ex-
clusive dataset is used in this study. Readers can extend this research for the diagnosis of dif-
ferent communicable diseases using the structure of CRV-NET and reproduce the experimen-
tal setup of CRV-NET to measure the intensity of COVID-19 by using the following link: https:
//github.com/UZAIRIQBAL-dev/CRV-NET-.git (accessed on 4 June 2022).
Acknowledgments: The authors extend their appreciation to the Deputyship for Research and Innovation,
Ministry of Education in Saudi Arabia for funding this research work via project number 959.
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Guan, W.J.; Ni, Z.Y.; Hu, Y.; Liang, W.H.; Qu, C.Q.; He, J.X.; Liu, L.; Shan, H.; Lei, C.L.; Hui, D.S.C.; et al. Clinical Characteristics
of coronavirus disease 2019 in China. N. Engl. J. Med. 2020, 382, 1708–1720. [CrossRef] [PubMed]
2. Ye, Z.; Zhang, Y.; Wang, Y.; Huang, Z.; Song, B. Chest CT manifestations of new coronavirus disease 2019 (COVID-19): A pictorial
review. Eur. Radiol. 2020, 30, 4381–4389. [CrossRef] [PubMed]
3. Chen, M.; Tu, C.; Tan, C.; Zheng, X.; Wang, X.; Wu, J.; Huang, Y.; Wang, Z.; Yan, Y.; Li, Z.; et al. Key to successful treatment of
COVID-19: Accurate identification of severe risks and early intervention of disease progression. MedRxiv 2020. [CrossRef]
4. Shi, H.; Han, X.; Jiang, N.; Cao, Y.; Alwalid, O.; Gu, J.; Fan, Y.; Zheng, C. Radiological findings from 81 patients with COVID-19
pneumonia in Wuhan, China: A descriptive study. Lancet Infect. Dis. 2020, 20, 425–434. [CrossRef] [PubMed]
5. Daimary, D.; Bora, M.B.; Amitab, K.; Kandar, D. Brain Tumor Segmentation from MRI Images using Hybrid Convolutional Neural
Networks. Procedia Comput. Sci. 2020, 167, 2419–2428. [CrossRef]
Diagnostics 2023, 13, 1783 14 of 15

6. Causey, J.L.; Guan, Y.; Dong, W.; Walker, K.; Qualls, J.A.; Prior, F.; Huang, X. Lung cancer screening with low-dose CT scans using
a deep learning approach. arXiv 2019, arXiv:1906.00240.
7. Singh, V.K.; Rashwan, H.A.; Romani, S.; Akram, F.; Pandey, N.; Sarker, M.K.; Saleh, A.; Arenas, M.; Arquez, M.; Puig, D.; et al.
Breast tumor segmentation and shape classification in mammograms using generative adversarial and convolutional neural
network. Expert Syst. Appl. 2019, 139, 112855. [CrossRef]
8. Zhao, W.; Jiang, D.; Queralta, J.P.; Westerlund, T. MSS U-Net: 3D segmentation of kidneys and tumors from CT images with a
multi-scale supervised U-Net. Inform. Med. Unlocked 2020, 19, 100357. [CrossRef]
9. Skourt, B.A.; El Hassani, A.; Majda, A. Lung CT Image Segmentation Using Deep Neural Networks. Procedia Comput. Sci. 2018,
127, 109–113. [CrossRef]
10. Huidrom, R.; Chanu, Y.J.; Singh, K.M. Automated Lung Segmentation on Computed Tomography Image for the Diagnosis of
Lung Cancer. Comput. Sist. 2018, 22, 907–915. [CrossRef]
11. Almotairi, S.; Kareem, G.; Aouf, M.; Almutairi, B.; Salem, M.A.-M. Liver Tumor Segmentation in CT Scans Using Modified
SegNet. Sensors 2020, 20, 1516. [CrossRef] [PubMed]
12. Kumar, P.; Nagar, P.; Arora, C.; Gupta, A. U-Segnet: Fully Convolutional Neural Network Based Automated Brain Tissue
Segmentation Tool. In Proceedings of the 2018 25th IEEE International Conference on Image Processing (ICIP), Athens, Greece,
7–10 October 2018; pp. 3503–3507. [CrossRef]
13. Akkus, Z.; Kostandy, P.; Philbrick, K.A.; Erickson, B.J. Robust brain extraction tool for CT head images. Neurocomputing 2019,
392, 189–195. [CrossRef]
14. Li, X.; Gong, Z.; Yin, H.; Zhang, H.; Wang, Z.; Zhuo, L. A 3D deep supervised densely network for small organs of human
temporal bone segmentation in CT images. Neural Netw. 2020, 124, 75–85. [CrossRef]
15. Qadri, S.F.; Shen, L.; Ahmad, M.; Qadri, S.; Zareen, S.S.; Khan, S. OP-convNet: A Patch Classification-Based Framework for CT
Vertebrae Segmentation. IEEE Access 2021, 9, 158227–158240. [CrossRef]
16. Hirra, I.; Ahmad, M.; Hussain, A.; Ashraf, M.U.; Saeed, I.A.; Qadri, S.F.; Alghamdi, A.M.; Alfakeeh, A.S. Breast Cancer
Classification From Histopathological Images Using Patch-Based Deep Learning Modeling. IEEE Access 2021, 9, 24273–24287.
[CrossRef]
17. Qadri, S.F.; Ai, D.; Hu, G.; Ahmad, M.; Huang, Y.; Wang, Y.; Yang, J. Automatic Deep Feature Learning via Patch-Based Deep
Belief Network for Vertebrae Segmentation in CT Images. Appl. Sci. 2018, 9, 69. [CrossRef]
18. Qadri, S.F.; Zhao, Z.; Ai, D.; Ahmad, M.; Wang, Y. Vertebrae segmentation via stacked sparse autoencoder from computed
tomography images. In Proceedings of the Eleventh International Conference on Digital Image Processing (ICDIP 2019),
Guangzhou, China, 10–13 May 2019; p. 160. [CrossRef]
19. Duran-Lopez, L.; Dominguez-Morales, J.P.; Corral-Jaime, J.; Vicente-Diaz, S.; Linares-Barranco, A. COVID-XNet: A Custom Deep
Learning System to Diagnose and Locate COVID-19 in Chest X-ray Images. Appl. Sci. 2020, 10, 5683. [CrossRef]
20. Qiblawey, Y.; Tahir, A.; Chowdhury, M.E.H.; Khandakar, A.; Kiranyaz, S.; Rahman, T.; Ibtehaz, N.; Mahmud, S.; Al Maadeed, S.;
Musharavati, F.; et al. Detection and Severity Classification of COVID-19 in CT Images Using Deep Learning. Diagnostics 2021,
11, 893. [CrossRef]
21. Zhao, X.; Zhang, P.; Song, F.; Fan, G.; Sun, Y.; Wang, Y.; Tian, Z.; Zhang, L.; Zhang, G. D2A U-Net: Automatic segmentation of
COVID-19 CT slices based on dual attention and hybrid dilated convolution. Comput. Biol. Med. 2021, 135, 104526. [CrossRef]
22. Shi, F.; Wang, J.; Shi, J.; Wu, Z.; Wang, Q.; Tang, Z.; He, K.; Shi, Y.; Shen, D. Review of Artificial Intelligence Techniques in Imaging
Data Acquisition, Segmentation, and Diagnosis for COVID-19. IEEE Rev. Biomed. Eng. 2020, 14, 4–15. [CrossRef]
23. Jadon, S. COVID-19 detection from scarce chest x-ray image data using few-shot deep learning approach. In Proceedings of the
Imaging Informatics for Healthcare, Research, and Applications, Online, 15–19 February 2021. [CrossRef]
24. Amyar, A.; Modzelewski, R.; Li, H.; Ruan, S. Multi-task deep learning based CT imaging analysis for COVID-19 pneumonia:
Classification and segmentation. Comput. Biol. Med. 2020, 126, 104037. [CrossRef] [PubMed]
25. Ozturk, T.; Talo, M.; Yildirim, E.A.; Baloglu, U.B.; Yildirim, O.; Acharya, U.R. Automated detection of COVID-19 cases using deep
neural networks with X-ray images. Comput. Biol. Med. 2020, 121, 103792. [CrossRef] [PubMed]
26. Rahimzadeh, M.; Attar, A. A modified deep convolutional neural network for detecting COVID-19 and pneumonia from chest X-ray
images based on the concatenation of Xception and ResNet50V2. Inform. Med. Unlocked 2020, 19, 100360. [CrossRef] [PubMed]
27. Xu, X.; Jiang, X.; Ma, C.; Du, P.; Li, X.; Lv, S.; Yu, L.; Chen, Y.; Su, J.; Lang, G. Deep Learning System to Screen novel Coronavirus
Disease 2019 Pneumonia. Engineering 2020, 6, 1122–1129. [CrossRef]
28. Wang, S.; Kang, B.; Ma, J.; Zeng, X.; Xiao, M.; Guo, J.; Cai, M.; Yang, J.; Li, Y.; Meng, X.; et al. A deep learning algorithm using CT
images to screen for Corona Virus Disease (COVID-19). Eur. Radiol. 2020, 31, 6096–6104. [CrossRef]
29. Zheng, C.; Deng, X.; Fu, Q.; Zhou, Q.; Feng, J.; Ma, H.; Liu, W.; Wang, X. Deep Learning-based Detection for COVID-19 from
Chest CT using Weak Label. MedRxiv 2020. [CrossRef]
30. Apostolopoulos, I.D.; Mpesiana, T.A. Covid-19: Automatic detection from X-ray images utilizing transfer learning with convolu-
tional neural networks. Phys. Eng. Sci. Med. 2020, 43, 635–640. [CrossRef]
31. Narin, A.; Kaya, C.; Pamuk, Z. Automatic detection of coronavirus disease (COVID-19) using X-ray images and deep convolutional
neural networks. Pattern Anal. Appl. 2021, 24, 1207–1220. [CrossRef]
32. Yan, Q.; Wang, B.; Gong, D.; Luo, C.; Zhao, W.; Shen, J.; Shi, Q.; Jin, S.; Zhang, L.; You, Z. COVID-19 Chest CT Image
Segmentation—A Deep Convolutional Neural Network Solution. arXiv 2020, arXiv:2004.10987.
Diagnostics 2023, 13, 1783 15 of 15

33. Voulodimos, A.; Protopapadakis, E.; Katsamenis, I.; Doulamis, A.; Doulamis, N. Deep learning models for COVID-19 infected area
segmentation in CT images. In Proceedings of the 14th PErvasive Technologies Related to Assistive Environments Conference,
Corfu, Greece, 29 June–2 July 2021; pp. 404–411. [CrossRef]
34. Fan, D.-P.; Zhou, T.; Ji, G.-P.; Zhou, Y.; Chen, G.; Fu, H.; Shen, J.; Shao, L. Inf-Net: Automatic COVID-19 Lung Infection
Segmentation From CT Images. IEEE Trans. Med. Imaging 2020, 39, 2626–2637. [CrossRef]
35. Azarpazhooh, M.R.; Morovatdar, N.; Avan, A.; Phan, T.G.; Divani, A.A.; Yassi, N.; Stranges, S.; Silver, B.; Biller, J.; Belasi, M.T.; et al.
COVID-19 Pandemic and Burden of Non-Communicable Diseases: An Ecological Study on Data of 185 Countries. J. Stroke
Cerebrovasc. Dis. 2020, 29, 105089. [CrossRef] [PubMed]
36. Ahsan, M.; Alam, T.E.; Trafalis, T.; Huebner, P. Deep MLP-CNN Model Using Mixed-Data to Distinguish between COVID-19 and
Non-COVID-19 Patients. Symmetry 2020, 12, 1526. [CrossRef]
37. Lalmuanawma, S.; Hussain, J.; Chhakchhuak, L. Applications of machine learning and artificial intelligence for Covid-19
(SARS-CoV-2) pandemic: A review. Chaos Solitons Fractals 2020, 139, 110059. [CrossRef]
38. Ismael, A.M.; Şengür, A. Deep learning approaches for COVID-19 detection based on chest X-ray images. Expert Syst. Appl. 2020,
164, 114054. [CrossRef] [PubMed]
39. Soares, E.; Angelov, P.; Biaso, S.; Froes, M.H.; Abe, K. SARS-CoV-2 CT-scan dataset: A large dataset of real patients CT scans for
SARS-CoV-2 identification. MedRxiv 2020. [CrossRef]
40. Saood, A.; Hatem, I. COVID-19 lung CT image segmentation using deep learning methods: U-Net versus SegNet. BMC Med.
Imaging 2021, 21, 19. [CrossRef]
41. Siddique, N.; Paheding, S.; Elkin, C.P.; Devabhaktuni, V. U-Net and Its Variants for Medical Image Segmentation: A Review of
Theory and Applications. IEEE Access 2021, 9, 82031–82057. [CrossRef]
42. Zhou, T.; Canu, S.; Ruan, S. Automatic COVID -19 CT segmentation using U-Net integrated spatial and channel attention
mechanism. Int. J. Imaging Syst. Technol. 2020, 31, 16–27. [CrossRef]
43. Wang, L.; Lin, Z.Q.; Wong, A. COVID-Net: A tailored deep convolutional neural network design for detection of COVID-19 cases
from chest X-ray images. Sci. Rep. 2020, 10, 19549. [CrossRef]
44. Feng, Y.-Z.; Liu, S.; Cheng, Z.-Y.; Quiroz, J.C.; Rezazadegan, D.; Chen, P.-K.; Lin, Q.-T.; Qian, L.; Liu, X.-F.; Berkovsky, S.; et al.
Severity Assessment and Progression Prediction of COVID-19 Patients Based on the LesionEncoder Framework and Chest CT.
Information 2021, 12, 471. [CrossRef]
45. Müller, D.; Rey, I.S.; Kramer, F. Automated Chest CT Image Segmentation of COVID-19 Lung Infection based on 3D U-Net. arXiv
2020, arXiv:2007.04774.
46. Alshazly, H.; Linse, C.; Abdalla, M.; Barth, E.; Martinetz, T. COVID-Nets: Deep CNN architectures for detecting COVID-19 using
chest CT scans. PeerJ Comput. Sci. 2021, 7, e655. [CrossRef] [PubMed]
47. Bernard, O.; Lalande, A.; Zotti, C.; Cervenansky, F.; Yang, X.; Heng, P.-A.; Cetin, I.; Lekadir, K.; Camara, O.; Ballester, M.A.G.; et al.
Deep Learning Techniques for Automatic MRI Cardiac Multi-Structures Segmentation and Diagnosis: Is the Problem Solved?
IEEE Trans. Med. Imaging 2018, 37, 2514–2525. [CrossRef] [PubMed]
48. Sharan, T.S.; Tripathi, S.; Sharma, S.; Sharma, N. Encoder Modified U-Net and Feature Pyramid Network for Multi-class
Segmentation of Cardiac Magnetic Resonance Images. IETE Tech. Rev. 2021, 39, 1092–1104. [CrossRef]
49. Zhou, Z.; Rahman, S.M.M.; Nima, T.; Liang, J. UNet++: A Nested U-Net Architecture for Medical Image Segmentation. In
Proceedings of the Multimodal Learning for Clinical Decision Support, Granada, Spain, 20 September 2018; Volume 11045 LNCS,
pp. 3–11. [CrossRef]
50. Chen, Y.; Qin, X.; Wang, J.; Yu, C.; Gao, W. FedHealth: A Federated Transfer Learning Framework for Wearable Healthcare. IEEE
Intell. Syst. 2020, 35, 83–93. [CrossRef]
51. Kamišalić, A.; Fister, I.; Turkanović, M.; Karakatič, S. Sensors and Functionalities of Non-Invasive Wrist-Wearable Devices: A
Review. Sensors 2018, 18, 1714. [CrossRef]
52. Kumar, Y.; Koul, A.; Singla, R.; Ijaz, M.F. Artificial intelligence in disease diagnosis: A systematic literature review, synthesizing
framework and future research agenda. J. Ambient. Intell. Humaniz. Comput. 2022, 1–28. [CrossRef]
53. Dash, S.; Verma, S.; Kavita; Bevinakoppa, S.; Wozniak, M.; Shafi, J.; Ijaz, M.F. Guidance Image-Based Enhanced Matched Filter
with Modified Thresholding for Blood Vessel Extraction. Symmetry 2022, 14, 194. [CrossRef]
54. Srinivasu, P.N.; Ahmed, S.; Alhumam, A.; Kumar, A.B.; Ijaz, M.F. An AW-HARIS Based Automated Segmentation of Human
Liver Using CT Images. Comput. Mater. Contin. 2021, 69, 3303–3319. [CrossRef]
55. Mandal, M.; Singh, P.K.; Ijaz, M.F.; Shafi, J.; Sarkar, R. A tri-stage wrapper-filter feature selection framework for disease
classification. Sensors 2021, 21, 5571. [CrossRef]
56. Muhammad, A.; Muhammad, F.I.; Son, Y. Data-Driven Cervical Cancer Prediction Model with Outlier Detection and Over-
Sampling Methods. Sensors 2020, 20, 2809.
57. Long, G.; Xie, M.; Shen, T.; Zhou, T.; Wang, X.; Jiang, J. Multi-center federated learning: Clients clustering for better personalization.
World Wide Web 2022, 26, 481–500. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.