Biomedical Signal Processing and Control 86 (2023) 104951

Contents lists available at ScienceDirect

Biomedical Signal Processing and Control

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

Kidney stone detection using an optimized Deep Believe network by

fractional coronavirus herd immunity optimizer
Chaohua Yan a, Navid Razmjooy b, *
a
School of Digital Media and Art Design, Nanyang Institute of Technology, Nanyang, Henan, 473004, China
b
Department of Computer Science and Engineering, Division of Research and Innovation, Saveetha School of Engineering, SIMATS, Chennai 602105, Tamil Nadu, India

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

Keywords: In this study, a computer-assisted kidney stone diagnosis system based on CT images has been proposed. The
Computer-aided diagnosis method is based on a combination of deep training and metaheuristics. The method aims to provide a customized
Kidney stone Deep Believe Network (DBN) based on a fractional version of the coronavirus herd immunity enhancer to provide
Deep Believe Network
an efficient and reliable kidney stone diagnosis system. The designed method is then authenticated by running a
Fractional order version of coronavirus herd
immunity optimizer
standard benchmark called a “CT kidney dataset”. Subsequently, a comparison is made between the results and
some other state-of-the-art methods. Simulations show that the recommended DBN/FO-CHIO outperforms the
other studied approaches in terms of efficiency with an accuracy of 97.98%. Moreover, the proposed DBN/FO-
CHIO recall outperforms others with 92.99%, demonstrating its excellent accuracy compared to other compar­
ison algorithms. Moreover, the higher specificity of the proposed method compared to the other evaluated ap­
proaches indicates its advanced event-independent value.

1. Introduction In recent years, 3-dimensional image processing in medicine has

performed an increasingly pivotal role in computer-assisted identifica­
Due to “today’s lifestyle, kidney stones have been among the most tion of disease, assisting radiologists in analyzing medical imaging and
frequent ailments. A change in the equilibrium between solubility and identifying aberrant findings in medical images [7]. Both imaging tests
salt deposition in renal causes kidney stones. Kidney stones are a hard (MRI, Ultrasound, and CT (Computer Tomography)) and traditional
build-up of crystals within the renal, ureter, bladder, or urethra. It’s ones (blood tests, urine tests, and biopsies) are utilized in Kidney stone
possible that kidney stones won’t be discovered until you attempt to pass diagnosis. Imaging tests employing CT have been converted into the
them [1]. The use of ultrasound as the primary diagnostic phase in the most prevalent diagnostic test regarding the acquired data, time, and
identification of kidney stones is common. cost consumption in this kind of test.
A prevalent kind of urological illness is kidney stone disease (neph­ However, owing to the numerous structures in abdominal CT, stone
rolithiasis) [2]. In mainland China, the prevalence of kidney stones has diagnosis on CT is still a difficulty in segmentation. Because of the
risen from 5.95 percent to 10.63 percent [3,4]. Due to vague symptoms, heterogeneous nature of tissues, the lack of distinct borders, the simi­
diagnosing the problem in the initial phase of the disease is impossible. larity of nearby organs, noise, and the partial volume effect, precise
The disease is visible only after the onset of symptoms of kidney injury organ segmentation is a tough challenge. As a result, the preprocessing
[5]. procedure is critical for increasing the 3D picture segmentation
Furthermore, renal illness is a chronic condition that damages the performance.
kidneys, resulting in a persistent and unresolved issue. As a result, In recent years, there has been a lot of research into using magnetic
detecting kidney stone illness before serious damage occurs is critical. resonance imaging (MRI) to detect kidney cancer and automatically
Kidney ailment could be efficiently cured providing that the detection of diagnose and classify kidney stones [8]. However, lots of noise in the
the stone problem is implemented earlier [6]. As a result, not only for the MRI pictures, human-operated processing of the images failed to detect
treatment of renal illness, but also for the management of recurring kidney stones, and was difficult and unreliable for huge volumes of data
stone development, stone identification is critical. [9]. To address this issue, artificial intelligence techniques have been

* Corresponding author.
E-mail address: [email protected] (N. Razmjooy).

https://doi.org/10.1016/j.bspc.2023.104951
Received 24 July 2022; Received in revised form 3 March 2023; Accepted 5 April 2023
Available online 19 April 2023
1746-8094/© 2023 Elsevier Ltd. All rights reserved.
C. Yan and N. Razmjooy Biomedical Signal Processing and Control 86 (2023) 104951

Fig. 1. Graphical abstract of the proposed method.

proposed for the detection and categorization of kidney stones, which urine system, is computed tomography (CT). The resemblance of in­
can deliver accurate findings. The automatic detection process reduces tensity and form of gem and non-stone structures is a problem in com­
the noise generated by the operator, which makes detection easier. This puter aided detection (CAD) for urinary stones, just like heavy high-
helps identify and classify existing kidney stones. resolution CT problems. In this work, the mentioned obstacles are cur­
An automated classification of kidney stones using neural network bed by employing a CNN which directly operates on the high-resolution
models and imaging data processing methods can provide significant CT data used in this investigation. A test set was implemented on eighty-
results. Human examination is a traditional method for classifying eight scans and the results endorsed that the optimal modeling using
medical resonance imaging and stone diagnosis. For large volumes of anatomic information and 2.5D input data had a mean of 2.68 false
data, operator-assisted classification systems are impractical. Medical positives for each sick person and 100 percent sensitivity.
MRI scans contain operating system noise, which can lead to inaccura­ Bhatt et al. [16] employed a Fuzzy Logic Analysis in Chronic K­
cies in classification. In this field, artificial intelligence approaches such idney Disease Detection. Chronic kidney disease (CKD) is a no commu­
as neural network models and feature extraction have shown consider­ nicable illness that is becoming more widespread across the world. The
able promise. frequency of this condition is on the rise, with roughly 10 % of persons of
Despite the fact that the shape of true kidney stones varies depending any age suffering from CKD. Accordingly, an intelligent system based on
on the period of aggravation, they are almost non-spherical. Recognizing computer that can identify kidney illness and its phases might be very
stones from imaging is challenging and time-consuming because worthwhile and helpful. A study of medical expert systems designed for
abdominal ultrasound images display bright areas and weak artifacts. As illness diagnosis utilizing various methodologies is presented in this
a result, information such as the texture, shape, and position of the stone publication. The results of Fuzzy Logic Analysis in Chronic Kidney
in the kidney is required to precisely diagnose kidney stones. By showed the accuracy of the fuzzy method in diagnosing the disease.
recognizing integrated and independent data, neural network models Vikas et al. [17] examined Image Analysis and Neural Networks for
minimize time and money [10–12]. The most extensively utilized ap­ Kidney Stone Detection. Due to the obvious presence of noise, there are
proaches for treating this condition are neural network models. The mistakes in the categorization of kidney stones. A Back - propagation
ability of neural networks to withstand mistakes, generalization, and learning network and imaging and data analysis techniques are used to
contextual learning allows for more accurate medical diagnosis. We automate kidney stone categorization.
might highlight the following studies among researchers that have This study uses the Fuzzy C-Mean technique to separate ultrasound
focused on this topic. imaging data to detect kidney stones early on. The feature extraction
A multikernel k-means clustering and NN model are utilized for ul­ approach used by the GLCM is a statistical tool. It has showed a lot of
trasound pictures by Nithya et al. [13] in order to discover kidney potential in terms of feature extraction, with a classification perfor­
illness. The main goal of this study is to create a technique for detection mance of 98.8 % accuracy [18–20]. The fuzzy C-means method out­
and segmentation of renal illness based on a classification and clustering performs K-Means clustering in the situation of overlapping results.
procedures. Preprocessing, feature extraction, classification, and seg­ Others result in a reduction in feature size, which may lead to the loss of
mentation are the four components that make up the system. In com­ key characteristics. According to the literature, there are various studies
parison to all other approaches, the suggested system as a segmentation that have been offered for the detection of kidney stones using deep
based on quadratic + linear division obtains a max accuracy of ninety- learning and artificial intelligence. However, there are many areas that
nine points sixty-one percent. may be changed to produce efficient results. The goal of this research is
Somasundaram et al. [14] analyzed an HDVS Deep Learning Method to provide an efficient way for diagnosing kidney stones and to establish
for Effective Renal Stone Diagnosis. Kidney crystals are one of the most an appropriate approach for this purpose. To do so, after getting the
common problems that individuals face, causing severe discomfort and input CT kidney images, they have injected with the DBN for the diag­
necessitating immediate medical attention. There are numerous imaging nosis application. Proper selection of a network’s biases and weights is
methods that can be used to identify KDS at an early stage and diagnose the main part of the DBN and has a noteworthy impact on its training
it. The use of DL algorithm to accomplish segmentation and feature and final diagnosis. So, here, a new improved version of Coronavirus
extraction for the detection of KDS is given in this study. These results of Herd Immunity Optimizer based on fractional theory has been proposed
the proposed HDVS technique are compared with those of previous for optimal selecting of these decision variables. The final models has
advanced procedures in DL methods such as SqueezeNet, Alexnet, been then applied to the “CT Kidney Dataset” as a standard database to
ResNet, and GoogleNet. The results revealed that the proposed scheme validate the proposed method. The graphical abstract of the method is
performed the best. indicated in Fig. 1.
Längkvist et al. [15] used Convolutional Neural Networks to help in Therefore, the main contributions of this study can be briefly high­
kidney stones’ identification in CT images. The most common approach lighted as follows:
for identifying ureteral stones, which are kidney stones that clog the

2
C. Yan and N. Razmjooy Biomedical Signal Processing and Control 86 (2023) 104951

Fig. 2. Different examples from the CT Kidney Dataset [21].

- A modified Deep believed Network (DBN) based on metaheuristics to stone?resource=download.”.

get higher efficiency. Fig. 2 shows different examples from the CT Kidney Dataset [21].
- Using the fractional order version of coronavirus herd immunity The present study uses this dataset for authentication of the sug­
optimizer to provide the idea. gested kidney stone diagnosis system.
- Validated the methodology by comparing it with some published
works from the literature. 3. Deep Belief Networks

2. Description of the dataset A Restricted Boltzmann Machine (RBM) can extract features and
rebuild the input data, but it can’t handle the blur of gradient yet.
The kidney images used in this study are from “CT Kidney Dataset” However, it is possible to design a neural network that can tackle this
[21]. The images contain 12,446 unique data points, consisting of 5,077 problem using a creative mix of many RBMs and a classifier [22]. The
normal data points, 3,709 cysts, 2,283 tumor data points, and 1,377 Deep Belief Network is the name given to this network. Jeff Hinton also
stone data points. Nazmul Islam and Humaion Kabir have generated this founded the Deep Belief Network, or DBN. Google is thought to employ
dataset, which has been gathered from the “PACS dataset (Picture these strong networks in its image recognition research. A deep
archiving and communication system)” archived in Bangladesh from the conviction and multilayer perceptron (MLP) are structurally similar, but
patients with different kidney problems including kidney tumors, cysts, their sole resemblance is in structure. To deal with gradient blur, a deep
stone findings, and people that are healthy. Dicom format was then belief network uses a whole new training strategy.
extracted from the images and then they converted to “*.jpg” format to The DBN is a multilayer superposition of RBM which is capable of
provide a proper database for validation in different image processing extracting the deep properties of the original data. The low-level RBM
applications. A physician and a radiologist double-checked the images launches the training procedure, which accepts DBN original data and
for labeling and for ensuring that the data has been accurately marked. slowly rises the hierarchy, while by far the highest level of RBM are
The link to accessing the presented dataset is: “https://www.kaggle. ultimately gained, where productions of DBN are trained, because the
com/datasets/nazmul0087/ct-kidney-dataset-normal-cyst-tumor-and- aim of the DBN is to make the possibility of training data maximum. This
method combines many and simpler (RBM) models that have been
learnt in order to create an effective technique to study (Precisely, a
complicated method).
Layer-wise Unsupervised Greedy Pre-Training is the name of this
technique. In essence, a DBN is an RBM that is taught to two layers at the
same time. The hidden layer of an RBM works as the input layer of an
adjacent RBM across the network. As a result, the first RBM is taught,
and its results are utilized as the input for the second RBM. This method
is continued until you reach the output layer.
Following this training, the DBN can be able to recognize patterns in
the data. In other terms, DBN is multilayered, complex feature extractor
[23]. This sort of network is unusual in that it learns the whole archi­
tecture of the input at the conclusion of each layer.
Other forms of deep networks have layers that gradually learn
complicated patterns; for example, primary levels can identify edges,
and succeeding layers can integrate their results to produce facial
characteristics. A DBN, on the other hand, learns hidden patterns at a
Fig.3. Deep Belief Network (DBN) Training process by the RBM. broad level, similar to how a camera gradually focuses a picture.

3
C. Yan and N. Razmjooy Biomedical Signal Processing and Control 86 (2023) 104951

Table 1 intended to locate a region in the weights where generalization may be

Parameter set values for the studied optimization algorithms. advanced while reducing overfitting, which happens whenever the test
Algorithm Parameter Set value error is high while the error of the training is modest. One of the most
significant advantages of DBN is its ability to derive features from high
WHO [34] α1 0.8
β1 0.2 level data and enhance the differentiation ability among multiple
α2 0.2 classes.
β2 0.8 By considering Fig. 3, it is clear that between the hidden and visible
FA [35] α 0.3 layers, there have been unadjusted biases and weights. The energy
0.4
function is constructed as follows to explain the joint distribution
β
γ 0.9
AOA [33] Protection probability 0.1 function of layers:
Elimination probability 0.25 ( ) /
c1 1.3 ρ Ly , Lh = e− E(Lv ,Lh ) fp (1)
c2 1.3
∑
fp = e− E(Ly ,Lh ) (2)
Ly ,Lh

Following this training, the DBN is able to recognize patterns in the

where, fp denotes the partition function for the hidden and visible
data. In other words, DBN is a multilayered, complex feature extractor.
layers attained by the sum of possible pairs, the binary formal for the jth
This sort of network is unusual in that it learns the whole structure of the ( )
input at the conclusion of each layer [24]. Other forms of deep networks hidden layer and ith visible layer are depicted by Lhj and Lyi , and E Ly , Lh
have layers that gradually learn complicated patterns; for example, denotes the energy of the joint formation of visible and invisible layers
primary levels can identify edges, and succeeding layers can integrate which can be obtained from equation (3):
their results to produce facial characteristics [25–27]. A DBN, on the ∑ ∑ ∑
E(Lv , Lh ) = − ai Lvi − bj Lhj − Lvi Lhj wij (3)
other hand, learns hidden patterns at a broad level, similar to how a i=1 j=1 i,j
camera gradually focuses a picture. Finally, a DBN requires labels to be
added to the result templates. In the above equation, ai describes the visible layer’s biases, bj
The DBN is then set up using supervised learning and a limited batch specifies the hidden layer’s biases, and the weight between visible and
of labeled samples as a final step. The network accuracy improves invisible elements of the RBM is represented by wij , weights updates as
somewhat after small weight and bias manipulations. Using the GPU, followed:
this entire process may be completed in a small chunk of time, and the ( ) (
Δwij = Et Lyi Lhj − Em Lyi Lhj
)
(4)
result is usually highly accurate. As a result, DBN is a viable solution to
( ) ( )
the gradient blur problem [28]. In the actual world, the training pro­ In equation (4), Em Lyi Lhj and Et Lyi Lhj illustrates the anticipation
cedure only necessitates a modest amount of labeled data. Fig. 3 in­
for the model and training data.
dicates the process of Deep Belief Network (DBN) training by the RBM.
Nonetheless, the DBN gives good classification results, constructing
where, Lv is the visible layer that includes binary stochastic visible
an ideal DBN structure can boost its efficiency even further. Various
units specified by v as a visible vector, whereas Lh as a hidden layer
strategies have been used to achieve this improvement. For this reason,
includes binary stochastic hidden units determined by h as hidden
this study employs a new enhanced metaheuristic technique that is
vector. Visible units have been entirely connected to the hidden layers
explained in the following.
situated in hidden levels with specified weights in the visible layer.
There are no links between the units in this network’s identical layers.
4. Coronavirus Herd Immunity Algorithm (CHIA)
As the idea of DBN is to make the possibility of training data
maximum, Low-level RBM has been considered for starting the training
4.1. Coronavirus
that receives DBN inputs and works its way up the hierarchy until it
reaches the highest level for DBN output training. By integrating simpler
Viruses multiply in the host body after being transmitted between
and numerous models (RBMs) that are learnt sequentially, the suggested
people. In 2019, the world was the eyewitness of the new virus outbreak
strategy provides an efficient approach to learning. The output of the
which spread from Wuhan, China. Who denounced the epidemic of
primary machine was already applied as the original data for the 2nd
COVID-19. As of March 27, 2020, there were “532,279” cases in 199
layer after learning it (2-nd RBM). The procedure is repeated while all
countries and around the world.
layers have been trained.
Covid-19 can stay in the body for 2 to 11 days without any symp­
The neural network is trained individually with no employing
toms. Due to the fact that there is still no cure for this disease, the
observer and generally with backward propagation. This phase was

Table 2
Validation of the proposed fractional order coronavirus herd immunity optimizer compared with other procedures based on average and standard deviation values.
Function FO-CHIO CHIO [36] WHO [34] FA [35] AOA [33]

AVE STD AVE STD AVE STD AVE STD AVE STD

F1 0.12 0.102 4.29×10+1 3.825×10+1 4.06 4.13×10+1 3.86 3.73×10+1 3.35 3.02×10+1
F2 5.49 4.67 5.32×10+1 4.84×10+1 4.49×10+21 4.03×10+1 3.98×10+1 3.34×10+1 3.82×10+1 3.33×10+1
F3 5.29 4.54 4.86 4.21 4.12 3.87 3.79 3.51 3.35 3.22
F4 0.11 0.09 1.25×10− 1 1.55×10− 1 1.45×10− 1 1.62×10− 1 1.88×10− 1 2.18×10− 1 2.38×10− 1 3.62×10− 1
F5 4.82 4.26 3.79×10+1 3.21×10+1 3.46×10+1 2.96×10+1 3.23×10+1 2.82×10+1 2. 34×10+1 2.03×10+1
F6 5.52×10− 2 5.89×10− 2 5.24×10− 1 6.14×10− 1 5.18×10− 1 6.35×10− 1 4.32×10− 1 6.83×10− 1 3.85×10− 1 6.81×10− 1
F7 6.01×10− 2 6.23×10− 2 6.56×10− 1 6.98×10− 1 6.69×10− 1 7.15×10− 1 6.99×10− 1 7.34×10− 1 7.32×10− 1 7.67×10− 1
F8 6.42×10+1 5.85×10+1 6.19×10+1 5.23×10+1 5.83×10+1 4.93×10+1 4.63×10+1 4.14×10+1 3.81×10+1 3.11×10+1
F9 4.85×10+2 4.43×10+2 4.69×10+2 3.82×10+2 3.31×10+2 2.36×10+2 2.89×10+2 2.25×10+2 1.94×10+2 1.35×10+2
F10 4.22×10+2 4.17×10+2 3.85×10+2 3.36×10+2 3.57×10+2 2.25×10+2 3.34×10+2 2.11×10+2 2.0.38×10+2 2.15×10+2

4
C. Yan and N. Razmjooy Biomedical Signal Processing and Control 86 (2023) 104951

Table 3
Computational complexity of the investigated methods on the benchmark functions.
Algorithm Function FO-CHIO CHIO [36] WHO [34] FA [35] AOA [33]

F1 1.76 s 1.45 s 1.82 s 1.94 s 2.50 s

F2 2.92 s 2.33 s 3.11 s 4.25 s 5.82 s
F3 3.05 s 2.71 s 3.42 s 3.94 s 6.61 s
F4 3.68 s 3.46 s 3.78 s 4.59 s 4.59 s
F5 2.57 s 2.14 s 2.66 s 2.81 s 2.67 s
F6 3.28 s 2.96 s 3.49 s 4.34 s 4.49 s
F7 2.16 s 1.95 s 2.24 s 3.68 s 3.79 s
F8 2.19 s 2.05 s 2.34 s 3.08 s 3.97 s
F9 4.59 s 4.38 s 4.73 s 5.37 s 7.38 s
F10 3.49 s 3.33 s 3.69 s 5.19 s 4.93 s

mortality rate of this disease is between 0.25 and 3.0 %. by n. The upper and lower bounds of gene are specified by ubi and lbi ,
The spread of the disease is ceased by herd immunity, at which time and xi is between lbi and ubi .
many persons in the community are immune to the disease by either CHIO contains 4 algorithmic variables which are described below:
natural infection or vaccination [29]. One of the methods that reduces
the transmission of coronavirus and controls the spread of this disease is • The number of the primary diseased case which is shown by C0 . Here,
herd immunity because in this case higher than 65 % of the population is it starts with one item.
improved from the disease. In this method, the principle of survival of • The maximum number of iterations that is indicated by Max Itr.
the fittest (in Darwinian theory) is used. • The population size that is indicated by HIS.
The infection with the corona virus occurs in two ways: • The problem dimensionality that is expressed by n.

When people are in close contact (less than 2 m) with an infected CHIO also includes two main control variables, which are quantified
person, they become infected with droplets of cough or sneezing. at this stage:
When the surface or device is a carrier of the virus and after contact
with it, the person touches their nose, eyes, or mouth. Herd immu­ • Basic reproduction rate (BRr ): It is used to control CHIO operators,
nity and quarantine in the country are methods used by governments which is done by distribution the virus among people.
to prevent the COVID-19 epidemic until the vaccine is discovered. • The maximum age of infected people (MaxAge ). This variable defines
the status of infected people. Infected people who reach MaxAge is
Every-one who is diseased can spread the virus to others. If a person either die or improve.
has a strong immune system, he or she will defeat the virus and become
immune, otherwise he or she will die. Older people have a weaker im­ 4.2. Coronavirus herd immunity optimizer (CHIO)
mune system or may have other illnesses such as cancer, diabetes, or
heart disease. As a result, older people are less likely to defeat the corona In this work, CHIO algorithm is utilized to improve Adaboost’s pa­
virus. According to research, the following steps are necessary to attain rameters. This algorithm is useful for minimizing computational
herd immunity: complexity and positive predictive value (precision) enhancement in the
evaluation of the AdaBoost performance. Prevention of COVID-19
• Many infected people transmit the virus to more people and infect outbreak provided the inspiration for the concept of CHIO. The spread
them. rate of coronavirus infection is intimately associated with the social
• Many infected people have improved and few have died. interaction of the affected individuals. Social distancing is advised in
• Over time, most people will be immune to the illness. order to avoid the COVID-19 spread. When the vast majority of the
population is immune, herd immunity is achieved by that species. Such
In the CHIA, the method of herd immunity is modeled. The contents immunity plays a pivotal role in prevention of disease spread. Optimi­
of COVID-19 in the optimization context are presented in Table 2. A set zation principles are utilized to clarify these concepts. The integration of
of stages is described below, and these stages are specified by CHIO: social distancing and herd immunity schemes leads to the proposed al­
Stage (1): the initialization of the optimization problem and CHIO’s gorithm (CHIO). According to the classification, there are three kinds of
variables are implemented in this stage. The following formula repre­ human cases for herd immunity: vulnerable, immune, and contami­
sents the fitness function at this stage. nated. That is implemented to discover whether the proposed method
applies social distancing schemes for genes’ update. In the following, the
minf (x) (5)
x numerical model of the offered algorithm is illustrated:
Where, x is between lb and ub. The degree of immunity of each in­ Stage 1: the initialization of the optimization problem and CHIO
dividual or the fitness function of each case is determined by f(x). The variables are done in this stage. Equation (12) depicts the fitness func­
decision variable, or each person’s gene, is denoted by x being from x1 to tion at stage one.
xn (x1 ,x2 ,⋯,xn ), and the whole number of genes in each case is expressed minf (x) (6)
x

Table 4
The classification effectiveness of the suggested DBN/FO-CHIO compared with DBN and DBN/CHIO.
Method Precision Accuracy Recall F1-score Specificity

DBN 93.99 84.99 77.67 82.99 92.45

DBN/CHIO 95.99 95.99 80.99 88.15 93.99
DBN/FO-CHIO 96.98 97.98 92.99 91.54 96.99

5
C. Yan and N. Razmjooy Biomedical Signal Processing and Control 86 (2023) 104951

Fig. 4. The classification efficiency of the proposed DBN/FO-CHIO compared with DBN and DBN/CHIO.

Here, f(x) is a function that determines each person’s immunity CHIO in the number of HIS. According to equation (13), a set of pro­
degree or each item’s fitness function in which x varies from lb to ub. x duced elements is stored as a matrix with n columns and HIS rows in the
denotes each individual’s gene (the decision variable), which is from HIP.
this set (x1 ,x2 ,⋯,xn ). Where, n dose represent the genes’ whole number ⎡ ⎤
in every case. ubi and lbi indicate the upper and lower boundaries of
⎢ x11 x12 ⋯ x1n ⎥
gene, and xi differs from lbi to ubi . There are 4 variables in CHIO algo­ ⎢
⎢ x2
⎥
rithm as following: ⎢ x22 ⋯ x2n ⎥
⎥
HIP = ⎢ 1 ⎥ (7)
⎢ ⋮ ⋮ ⋯ ⋮ ⎥
⎢ ⎥
- C0 : the primary contaminated case’s number that commences with 1 ⎣ xHIS xHIS ⋯ HIS ⎦
xn
1 2
case.
- Max Itr: the maximum quantity of repetitions.
Where, xj is achieved from this formula:
- HIS: the size of the population
- n: the dimensionality of the problem. xji
= lbi +(ubi − lbi ) × U(0, 1), ∀i = 1, 2, ⋯, n. In this formula j is row
number and each row describes a case xj . As previously mentioned,
There is also 2 major control variables in CHIO method, which are Formula 1 is used to obtain the fitness function for each case. For all
measured at stage1: cases in HIP, S (the status vector) is launched by infected individual (1)
or susceptible ones (zero).What is more, the number of infected in­
- BRr : (Basic reproduction rate) it is utilized to regulate CHIO opera­ dividuals (1) in S is randomly commenced as many as C0 .
tives, implemented by spreading the virus in a society. Stage (3): This step is the basic development loop of CHIO, which can
j
- MaxAge : (infected individuals’ maximum age) This explains the be addressed during CHI development. For each xj (individual), xi (
diseased individuals’ status. Sick people, who are MaxAge years old, is related gene ) is either remained stable or obtained from equation (14),
either pass away or recover. which is achieved under social distancing circumstances via 3 rules
according to BRr :
Stage (2): HIP (Herd Immunity Population) is almost made in this
stage in a way that a haphazard set of cases (individuals) is created by

Fig. 5. Confusion matrix created using the test data.

6
C. Yan and N. Razmjooy Biomedical Signal Processing and Control 86 (2023) 104951

Table 5
Classification efficiency of the proposed DBN/FO-CHIO in comparison with some diagnosis systems in the literature.
Technique Precision Accuracy Recall F1-score Specificity

LS/NN [37] 78.29 84.99 68.51 85.99 72.45

GLCM [38] 81.45 81.53 72.99 83.50 83.15
PPs [40] 85.99 76.99 78.09 61.14 86.99
kmeans/NN [13] 89.99 84.06 81.66 68.51 85.99
DEEP [39] 94.99 86.61 93.12 89.99 92.99
DBN/FO-CHIO 96.98 97.98 92.99 91.54 96.99

⎧ ⎫
⎪
⎪ xij (t) r ≥ BRr ⎪
⎪ ⎧
⎪
⎪ ⎪
⎪ ⎪
⎪ ) f (x)j (t + 1) ( )
⎪ // ⎪ ⎪ f (xj (t + 1) < ∧ Sj = 0 ∧ is Corona xj (t + 1)
⎪
⎪
⎪
⎪
( j ) 1 ⎪
⎪
⎪
⎪ ⎨1
⎪
Δf (x)
⎨ C xi (t)
⎪ r < × BRr .
3
infectedcase ⎪
⎬ Sj ← (15)
xji (t + 1)← // (8) ⎪
⎪
⎪
⎪ j
) f (x)j (t + 1)
⎪
⎪
⎪
( ) 2 ⎪
⎪
⎪ ⎩2 f (x (t + 1) > ∧ Sj = 1
⎪
⎪ N xji (t) r < × BRr . susceptiblecase ⎪
⎪ Δf (x)
⎪
⎪ 3 ⎪
⎪
⎪ ( j )
⎪ ⎪
⎪
⎪
⎩ R xi (t) ⎪
⎭ is Corona(xj (t +1)) is binary value equal to one, whereas xj (t +1) (the
r < BRr .//immunedcase
new item) takes a value from infected items. The average immune rate of
∑HIS
In this equation, r creates a haphazard number from zero to one. f(xi )
the population is defined by Δf (x), for instance, j=1 . The immunity
Mentioned rules are clarified as follows: HIS
[
0,1
]
j
rate of individuals is gained in accordance with social distance whenever
Infected individual: When r ∈ 3BR r
, xi (t +1) (the new gene number) the immunity rate of the new case is far better than the mean population
is obtained by the disparity between a gene resulted from an infected immunity. Therefore, there will be a safer population. We will be on the
individual (xc ) and the present gene, under social distancing circum­ verge of herd immunity if the new population is adequately safe against
stance, for instance: the disease.
( ) Stage 5: The penultimate stage is fatalness of individuals. If the im­
xji (t = 1) = C xji (t) (9) ( )
munity rate f xj (t + 1) of the existing infected person (Sj = 1) is not
Where able to recuperate after a specific number of repetitions as computed by
( ) ( ) the Max Age (i.e., Aj ≥ Max Age), the case is assumed to die. Afterwards,
C xji (t) = xji (t) + r × xji (t) − xic (t) (10) j
it is reformed from the start by this equation”xi (t +1) = lbi + (ubi − lbi )
xci (t) is chosen from some infected cases xc in accordance with S (the × U(0, 1), ∀i = 1, 2, ⋯, n.” This implementation is for diversifying the
status vector ). Therefor,c = (i|Si = 1). existent population, and most importantly, for evading local
[ ) j
Susceptible case: When r ∈ 13BRr , 23BRr , xi (t + 1) is obtained by the optimization.
disparity between a gene resulted from a susceptible case (xv ) and the Stage 6: The eventual step is halt condition. In CHIO, phase 3 through
existing gene, under social distancing circumstance, for instance 6 are repeated until the ending condition is concluded. The eventual
( ) state is usually overshadowed by the max number of repetitions. Having
xji (t + 1) = N xji (t) (11) reached the ending condition, the safe cases overcome the population
and the infected cases disappear.
Where
( ) ( )
N xji (t) = xij (t) + r × xij (t) − xvi (t) (12) 4.3. The fractional order-based coronavirus herd immunity optimizer
(FO-CHIO)
xvi (t) is haphazardly gained from some susceptible item xv according
to S, Thus, v = (i|Si = 0).
[ ) j Although various and diverse implementations of the coronavirus
Immune case: When r = 23BRr , BRr , xi (t + 1) it is achieved by the herd immunity optimizer have been conducted, all researches demon­
disparity between a gene resulted from an immuned case xz and the strate that the original coronavirus herd immunity optimizer has diffi­
existing one, for example culty in tackling specific problems. These issues often involve the local
optimization of problems, which yields erroneous results. This motivates
xji (t + 1)= R(xji (t) (13)
researchers to develop updated versions of the method, such as Modified
Where CHIO [30], and Improved CHIO [31].
( ) ( ) A novel design of Coronavirus Herd Immunity Optimizer is presented
R xji (t) = xij (t) + r × xij (t) − xvi (t) (14) here in order to obtain better outcomes. We employed fractional cal­
xzi (t)
is obtained from the best immune case x according to S, for
z culus to produce this adjustment. The fractional calculus is a relatively
instance:f(xz ) = argj min[k|Sk =2] f(xj ). new mechanism that may be used as an acceptable adjustment for
optimization strategies. We employed the Grunwald-Plotnikov (GL)
Stage 4: This stage is associated with HIP updating. In this step,
technique to offer the Fractional Order (FO) alteration in this case.
f(xj (t +1)) of each produced case xj (t + 1) is arranged and xj (t) (the
Considering this process, the following equation yields a FO-based dif­
existing item) is replaced with the generated item xj (t +1) if an
ferential equation:
improvement yield, like f(xj (t + 1)) < f(xj (t)). Aj (age) increases by one ( )
providing that Sj = 1. ( ) 1 ∑∞ t δ
Dδ xji (t) = lim δ ( − 1) x(l − jh) (16)
Equation (12) is utilized for updating Sj (the status vector) . h→0 h j=0 j

where, δ signifies the derivative order operator, and:

( )
δ δ(δ − 1)(δ − 2)⋯(δ − j + 1)
= (17)
j j!

7
C. Yan and N. Razmjooy Biomedical Signal Processing and Control 86 (2023) 104951

1700,1600, and 2100, and finally for F8-F10 (composite functions for N
where, the fractional order derivative (Dδ (x(l) ) of order δ is achieved by
= 3, N = 4, and N = 5) are 2200, 2400, and 2500.
the following equation:
For a fair validation of the proposed methodology, a comparison
[ ] 1 ∑N ( − 1)j Γ(δ + 1)xi (l − jT) with some techniques in the literature is implemented, which includes:
Dδ xji (t) = δ (18)
T j=0 Γ(j + 1)Γ(δ − j + 1) Archimedes Optimization Algorithm (AOA)[33], Wildebeest herd opti­
mization (WHO) [34], Firefly algorithm (FA) [35], and the original
where, Γ is gamma function, T stands for the sampling time, and N de­ coronavirus herd immunity optimizer [36]. Table 1 shows the parameter
scribes the window memory. set values for the studied optimization algorithms.
By assuming δ = 1, the above equation is changed to the following Simulations have been established on a laptop running Microsoft
formula: Windows 10 with MATLAB R2018b environment. For all algorithms, the
[ ] population number and the max number of repetitions are assumed 45
x ji (t + 1)
D1 xij (t) = xji (t + 1) − ̂ (19)
and 200, respectively. Each algorithm runs 25 times independently to
[ ] provide consistent results. Consequently, a total of 5000 function eval­
j
where, D1 xi (t) describes the distance difference definition. uations are established. The minimum and the maximum candidate
The following formula is employed to update the candidates’ constraint for the functions are − 100 and 100, respectively. Table 2
positions: records the final validation of the proposed fractional order coronavirus
( ) herd immunity optimizer compared with other procedures in the liter­
− xji (t) = r × xji (t) − xvi (t) (20) atures based on average and standard deviation values.
And the total formulation of the equation is achieved by the As it is clear from Table 2, we can conclude that the proposed frac­
following equation: tional order coronavirus herd immunity optimizer beats other compar­
ative methods, i.e., WHO, FA, AOA, and the original coronavirus herd
[ j ] ∑m ( − 1)j Γ(δ + 1)xj (t + 1) ( ) immunity optimization algorithm. The findings disclose that, in addition
x i (t + 1) = xji +
Dδ ̂ i
= r × xij (t) − xiv (t)
j=1 Γ(j + 1)Γ(δ − j + 1) to being superior to the original version, the proposed technique per­
(21) forms far better than other procedures regarding its accuracy in solving
the examined functions. Furthermore, a smaller standard deviation
By considering equation (21), the new updated equation is achieved
score based on the proposed technique against the others indicates
as follows:
greater consistency across several runs. On the benchmark functions
∑m ( − 1)j Γ(δ + 1)xj (t + 1) ( ) under consideration, Table 3 shows the computational complexity of the
̂ l+1 (k) = − i
+ r × xji (t) − xvi (t) (22)
X j j=1 Γ(j + 1)Γ(δ − j + 1) proposed fractional order coronavirus herd immunity optimizer and the
other investigated approaches.
At last, by assuming the explained equations, the first four terms of Table 3 shows that the CHIO has the fastest speed when compared to
the updated susceptible case are obtained by the following equation: the others. The planned FO-CHIO is slower than the CHIO, despite
1 j+1 1 1 certain improvements being made. By keeping in mind that the
̂ l+1 δ(1 − δ)xij+1 (t) + δ(1 − δ)(2 − δ)xj+1
X j (k) = δxi (t + 1) +
1! 2! 3! i (t − 1) approach offers a decent trade-off between accuracy and convergence
1 ( ) speed, this may be disregarded.
+ δ(1 − δ)(2 − δ)(3 − δ)xij+1 (t − 2) + r × xij (t) − xvi (t)
4!
(23) 5. Optimizing the DBN based on FO-CHIO
For the initialization stage, the objective function evaluation and the
Proper selection of a network’s biases and weights has a significant
solution update have been 3 aspects that effect on the computational
influence on its superior learning and final classification, that is based on
complexity of the fractional coronavirus herd immunity optimizer. The
minimizing the error between the prediction and the network data. Data
initialization procedure is of O(N) complexity, where N is the individual
assignment in the associated class is a crucial task due to the large in­
size of the virus. The complexity of the objective function varies ac­
fluence of the last layer in system diagnostics. This may be met if the
cording to the task, from O(N2 ) to O(N4 ). Finally, updating the solutions
relationship between the weights in the layers is appropriately chosen.
produces a complexity of O(MN) + O(MNL), where M is the number of
We applied the suggested FO-CHIO for training the final weight vector to
iterations and L is the number of parameters of the problem. For the
improve the classification accuracy. The number of optimization in­
studied functions, the computational complexity of the proposed frac­
dividuals and the max number of repetitions are considered 50 and 200
tional coronavirus herd immunity optimizer is O(N) +
( ( )) ( ( )) in this case.
O (NML) +O N2 or O(N) + O (NML) +O N4 .
As previously stated, the proposed FO-CHIO is used here for optimal
selection of biases (b) and the weights (W) so that the value of error,
4.4. Authentication of the algorithm between the anticipated value and network result, is reduced. These
cases are defined as follows:
As aforementioned, the presented fractional order of coronavirus [ ]
W = w1 , w2 , ..., wp bn = [b1n , b2n , ..., bLn ]l = 1, 2, ..., Ln = 1, 2, ..., A (24)
herd immunity optimizer can resolve the problem of local optimization
for this algorithm. To authenticate this advantage in the improved
X = [x1 , x2 , ..., xA ] (25)
version, we should authenticate the method efficiency by applying it to
some standard test functions. Here, we used the 10 first functions of the wn = [w1n , w2n , ..., wLn ] (26)
“CEC-BC-2020 test suite” [32]. This benchmark contains unimodal,
composite, hybrid, and multimodal functions that simulate real-world where, l specifies the index of layer. X describes the number of solutions.
situations. Due to this benchmark’s similarity with real-world models, n represents the individuals’ number. L expresses the whole quantity of
it is widely used in the authentication optimization problems. For all layers, and the weight for the ith layer is defined by win .
functions, the variable number is set 10. The optimal value for F1 The error value that is used for the DBN optimization is formulated
(rotated and shifted bent cigar) is one hundred, for F2 (rotated and below:
shifted Schwefel’s) is 1100, for F3 (rotated and shifted Lunacek bi-
Rastrigin) is seven hundred, for F4 (Expanded Griewangk’s and Rose­
nbrock’s) is 1900, for F5-F7 (Hybrid functions 1–3 (N = 3)) are

8
C. Yan and N. Razmjooy Biomedical Signal Processing and Control 86 (2023) 104951

(⃒ j ⃒)
⃒V (i) − V j (i) ⃒ 2 respectively. This indicator is mathematically defined by the following
1∑ N ∑ M
(27) equation:
Emp Des
OF = j ,
T j=1 i=1 VEmp (i)
Precision × Sensitivity
F1 = 2 × × 100 (31)
j
where, M and N describe the number of data and output layers. VEmp (i) Precision + Sensitivity
j
and VDes (i) represent the empirical and desired output voltage of the
network from the time line t. - Specificity
The DBN optimization procedure was continued until the required
minimal value for the error was obtained. The trained network was then This indicator determines the ability of a test for correctly diagnosing
applied to the database in order to validate the method’s correctness. people without the disease. This indicator is achieved by the following
The fundamental rationale for employing the suggested FO-CHIO is that equation:
utilizing traditional approaches for error minimization results in local TN
minimum difficulties in some circumstances. Specificity = × 100 (32)
TN + FP

6. Simulation Results To show the prominence of the suggested methodology, its results
are compared with simple Deep believed Network (DBN), DBN/CHIO,
The proposed method has been executed by MATLAB R2018b. The and the proposed DBN/FO-CHIO. The results are indicated in Table 4.
experiment has been performed on a core i7-6700HQ Asus N552 pro­ To Provide more clarification, Fig. 4 records the classification effi­
cessor and 12 GB RAM. Here, some different measurement indexes have ciency of the proposed DBN/FO-CHIO compared with DBN and DBN/
been utilized to examine the proposed method’s efficiency. CHIO in a bar plot.
In a classification problem, merely 1 statistic like accuracy, it is not In terms of accuracy, precision, and sensitivity, the recommended
able to properly evaluate the overall model efficiency. As a result, five DBN/FO-CHIO method outperforms the conventional DBN and the
factors, namely precision, F1 score, accuracy, specificity, and recall are traditional DBN/FO-CHIO. The suggested technique with 96.98 percent
assessed for each category of kidney stone condition. Moreover, CM (the Precision, 97.98 percent Accuracy, and 96.99 percent Specificity, pro­
Confusion Matrix) is plotted to see the performance of the model across vides the best combination of system correctness and dependability. In
all classes. We’ll now look at how the metrics described above are addition, the proposed DBN/FO-CHIO with 92.99 percent as highest
calculated and what they imply. All measurement indicators are based value of sensitivity, delivers how good this method is at detecting a
on four terms: True Positive and Negative = TP and TN; False Positive positive kidney stone. Fig. 5 displays the confusion matrix created using
and Negative = FP and FN. the test data.
For clarification, the case is termed a TP instance providing that the As can be observed from Fig. 5, the confusion matrix demonstrates
kidney stone picture is labeled with stone and the model guesses it. False that the model accurately predicted 1322 kidney stone pictures (true
negative occurs when an image is labeled with stone but is classed as positives, TP), and 626 kidney stone images in the normal class (false
healthy. A false positive instance occurs when the classification model negative, FN). Moreover, the model accurately identified 11,820 true
indicates that a kidney has stones, but it rightfully healthy. If the clas­ negative (TN) photos as belonging to the normal class. On the other
sifier suggests that a healthy image has not a stone, this is an example of hand, it incorrectly identified 55 benign pictures as kidney stones (False
true negative. Positive, FP). For providing more analyses, the efficiency of the pro­
posed method has been validation by applying it to the predetermined
- Precision measurement indicators and a comparison with some published
methods in the literature is implemented, including LS/NN (level Set /
This measure addresses that what fraction of positive diagnosis has Artificial Neural Network classification) [37], GLCM (Gray-Level Co-
been truly corrected as the following equation: Occurrence Matrix) [38], DEEP (Deep model) [39], PPs (Learning
Part-Prototypes) [40], and multi-kernel k-means /Artificial Neural
Precision =
TP
× 100 (28) Network clustering (kmeans/NN) [13]. Table 5 represents the classifi­
TP + FP cation efficiency of the suggested DBN/FO-CHIO in comparison with
some other diagnosis systems in literature.
- Accuracy As shown in Table 5, the recommended DBN/FO-CHIO with 97.98
percent accuracy outperforms the other examined approaches in terms
This measure defines the fraction of predictions that is accurately of efficiency. Furthermore, DEEP’s [39] approach was placed second in
determined. Accuracy is achieved by the following equation: terms of accuracy, with 94.99 percent. Furthermore, the suggested
DBN/FO-CHIO recall with 92.99 percent outperforms the others,
TP + TN
Accuracy = × 100 (29) demonstrating its excellent accuracy in contrast to other comparator
TP + TN + FP + FN algorithms. Moreover, the suggested method’s higher specificity in
comparison to the other approaches evaluated indicates its advanced
- Recall (Sensitivity): occurrence-independent value.

This measure is also defined as True Positive Rate (TPR). This mea­ 7. Conclusions
sure indicates that what fraction of the real positives has been diagnosed
correctly. This indicator is achieved by the following equation: Kidney stones (Urolithiasis) are different types of hard deposits that
TP are made of acidic salts inside the kidneys and cause pain in the lower
Recall = × 100 (30) back and flanks, which will also vary depending on the type of treatment
TP + FN
stone. Kidney stones have different causes such as family history and/or
malnutrition occurs which is treated according to the type of stone and
- F1 Score the patient’s condition. Kidney stones are hard deposits that are made up
of acidic salts and salts inside your kidneys. Accurate diagnosis of kidney
This indicator is defined as a weighted mean of the recall and pre­ stone can improve the chance of healing of this disease. In this paper, a
cision. 1 and 0 are the finest and poorest values of this indicator, novel automatic method was suggested for accurately diagnosing of the

9
C. Yan and N. Razmjooy Biomedical Signal Processing and Control 86 (2023) 104951

kidney stone from the CT images. The proposed method presented a [10] W. Jiang, et al., Optimal economic scheduling of microgrids considering renewable
energy sources based on energy hub model using demand response and improved
novel procedure using metaheuristics and deep learning. The major
water wave optimization algorithm, J. Storage Mater. 55 (2022), 105311.
objective was to provide a modified version of deep belief network [11] E. Han, N. Ghadimi, Model identification of proton-exchange membrane fuel cells
(DBN) based on an improved version of a metaheuristic technique, based on a hybrid convolutional neural network and extreme learning machine
named Fractional order coronavirus herd immunity optimizer to deliver optimized by improved honey badger algorithm, Sustainable Energy Technol.
Assess. 52 (2022), 102005.
a reliable detection system for kidney CT images. The designed system [12] L. Chen, et al., Optimal modeling of combined cooling, heating, and power systems
was then authenticated by applying to a standard benchmark, called “CT using developed African Vulture Optimization: a case study in watersport complex,
kidney dataset” and a comparison with some published techniques was Energy Sources Part A 44 (2) (2022) 4296–4317.
[13] A. Nithya, et al., Kidney disease detection and segmentation using artificial neural
made, including LS/NN classification, GLCM, DEEP, PPS, and k-means/ network and multi-kernel k-means clustering for ultrasound images, Measurement
NN clustering. Concluding results revealed that the suggested technique 149 (2020), 106952.
can provide more efficient results in diagnosing the kidney stone from [14] Somasundaram, K. and P. Sivakumar, An Efficient Detection of Kidney Stone Based on
HDVS Deep Learning Approach. 2021.
the CT images. [15] M. Längkvist, et al., Computer aided detection of ureteral stones in thin slice
computed tomography volumes using Convolutional Neural Networks, Comput.
CRediT authorship contribution statement Biol. Med. 97 (2018) 153–160.
[16] M. Bhatt, T. Kasbe, A Survey on Chronic Kidney DiseAse DiAgnosis Using Fuzzy
Logic, IEEE, 2019.
Chaohua Yan: Conceptualization, Data curation, Writing – original [17] A.L. Vikas, Kidney stone detection using image processing and neural networks,
draft, Writing – review & editing. Navid Razmjooy: Conceptualization, Annals of RSCB 25 (6) (2021) 13112–13119.
[18] M. Gheydi, et al., Planning in microgrids with conservation of voltage reduction,
Data curation, Writing – original draft, Writing – review & editing. IEEE Syst. J. 12 (3) (2016) 2782–2790.
[19] M. Ghiasi, et al., An analytical methodology for reliability assessment and failure
analysis in distributed power system, SN Appl. Sci. 1 (1) (2019) 1–9.
Declaration of Competing Interest [20] D. Yu, N. Ghadimi, Reliability constraint stochastic UC by considering the
correlation of random variables with Copula theory, IET Renew. Power Gener. 13
The authors declare that they have no known competing financial (14) (2019) 2587–2593.
[21] CT KIDNEY DATASET: Normal-Cyst-Tumor and Stone, N.I.a.H. Kabir, Editor. 2021.
interests or personal relationships that could have appeared to influence [22] Z. Guo, et al., Novel computer-aided lung cancer detection based on convolutional
the work reported in this paper. neural network-based and feature-based classifiers using metaheuristics, Int. J.
Imaging Syst. Technol. (2021).
[23] M. Rezaie, et al., Model parameters estimation of the proton exchange membrane
Data availability
fuel cell by a Modified Golden Jackal Optimization, Sustainable Energy Technol.
Assess. 53 (2022), 102657.
The authors are unable or have chosen not to specify which data has [24] H. Guo, W. Gu, M. Khayatnezhad, N. Ghadimi, Parameter extraction of the SOFC
been used. mathematical model based on fractional order version of dragonfly algorithm, Int.
J. Hydrogen Energy 47 (2022) 24059–24068.
[25] M. Khayyatnezhad, R. Gholamin, H. Keynoos, Investigation of the effect of end
season drought stress on morphological on durum wheat genotypes, Advancements
Acknowledgements:
in Life Sciences 9 (2022) 284–288.
[26] M. Khayyatnezhad, H. Keynoos, Incorporating Conditional Uncertainty into
Henan Provincial Department of Science and Technology Provincial Decision-making for Forecasting Actual Evapotranspiration in Semi-arid Area,
Key Research and Development and Promotion Project, Project Number: Advancements in Life Sciences 9 (2022) 498–503.
[27] J. Zhang, M. Khayatnezhad, N. Ghadimi, Optimal model evaluation of the proton-
222102320375, Project Title: Research on Key Technologies for Plant exchange membrane fuel cells based on deep learning and modified African
Image Recognition in the Danjiang Basin based on Multi-scale Con­ Vulture Optimization Algorithm, Energy Sources Part A 44 (2022) 287–305.
volutional Neural Network. [28] M. Ghiasi, et al., A comprehensive review of cyber-attacks and defense mechanisms
for improving security in smart grid energy systems: past, present and future,
Nanyang Science and Technology Tackling Key Problems Project, Electr. Pow. Syst. Res. 215 (2023), 108975.
Project Number: KJGG023, Project Title: Research on Rapid Retrieval [29] Z. Yuan, et al., Probabilistic decomposition-based security constrained
Technology for Distributed Medical Image Data based on Cloud transmission expansion planning incorporating distributed series reactor, IET
Gener. Transm. Distrib. 14 (17) (2020) 3478–3487.
Platform. [30] S.N. Makhadmeh, et al., A modified coronavirus herd immunity optimizer for the
power scheduling problem, Mathematics 10 (3) (2022) 315.
References [31] A. Naderipour, et al., An improved corona-virus herd immunity optimizer
algorithm for network reconfiguration based on fuzzy multi-criteria approach,
Expert Syst. Appl. 187 (2022), 115914.
[1] X. Cai, et al., Breast cancer diagnosis by convolutional neural network and
[32] A.K. Mohamed, A.A. Hadi, A.W. Mohamed, Generalized Adaptive Differential
advanced thermal exchange optimization algorithm, Comput. Math. Methods Med.
Evolution Algorithm for Solving CEC 2020 Benchmark Problems, IEEE, 2020.
(2021).
[33] F.A. Hashim, et al., Archimedes optimization algorithm: a new metaheuristic
[2] N. Razmjooy, et al., Computer-aided diagnosis of skin cancer: a review, Current
algorithm for solving optimization problems, Appl. Intell. 51 (3) (2021)
medical imaging 16 (7) (2020) 781–793.
1531–1551.
[3] S. Mahdinia, et al., Optimization of PEMFC Model Parameters Using Meta-
[34] D. Amali, M. Dinakaran, Wildebeest herd optimization: a new global optimization
Heuristics, Sustainability 13 (22) (2021) 12771.
algorithm inspired by wildebeest herding behaviour, J. Intell. Fuzzy Syst. (2019
[4] F.R.S. Navid Razmjooy, N. Ghadimi, A hybrid neural network – world cup
(Preprint):) 1–14.
optimization algorithm for melanoma detection, Open Medicine 13 (2018) 9–16.
[35] Yang, X.-S., Firefly algorithm. 2008.
[5] Z. Xu, et al., Computer-aided diagnosis of skin cancer based on soft computing
[36] M.A. Al-Betar, et al., Coronavirus herd immunity optimizer (CHIO), Neural
techniques, Open Medicine 15 (1) (2020) 860–871.
Comput. & Applic. 33 (10) (2021) 5011–5042.
[6] N. Razmjooy, et al., Imperialist competitive algorithm-based optimization of neuro-
[37] K. Viswanath, R. Gunasundari, Design and Analysis Performance of Kidney Stone
fuzzy system parameters for automatic red-eye removal, Int. J. Fuzzy Syst. 19 (4)
Detection From Ultrasound Image by Level Set Segmentation and ANN
(2017) 1144–1156.
Classification, IEEE, 2014.
[7] N. Razmjooy, et al., A hybrid neural network–world cup optimization algorithm for
[38] R. Goel, A. Jain, Improved detection of kidney stone in ultrasound images using
melanoma detection, Open Medicine 13 (1) (2018) 9–16.
segmentation techniques, in: Advances in Data and Information Sciences, Springer,
[8] M. Ghiasi, et al., Evolution of smart grids towards the Internet of energy: concept
2020, pp. 623–641.
and essential components for deep decarbonisation, IET Smart Grid. 6 (1) (2023)
[39] N. Thein, et al., An Image Preprocessing Method for Kidney Stone Segmentation in
86–102, https://doi.org/10.1049/stg2.12095.
CT ScAn Images, IEEE, 2018.
[9] N. Ghadimi, et al., An innovative technique for optimization and sensitivity
[40] Flores-Araiza, D., et al., Interpretable Deep Learning Classifier by Detection of
analysis of a PV/DG/BESS based on converged Henry gas solubility optimizer: a
Prototypical Parts on Kidney Stones Images. arXiv preprint arXiv:2206.00252, 2022.
case study, IET Gener. Transm. Distrib. (2023).

10