Medical Image Classifications for 6G IoT-Enabled Smart Health Systems

Abstract

As day-to-day-generated data become massive in the 6G-enabled Internet of medical things (IoMT), the process of medical diagnosis becomes critical in the healthcare system. This paper presents a framework incorporated into the 6G-enabled IoMT to improve prediction accuracy and provide a real-time medical diagnosis. The proposed framework integrates deep learning and optimization techniques to render accurate and precise results. The medical computed tomography images are preprocessed and fed into an efficient neural network designed for learning image representations and converting each image to a feature vector. The extracted features from each image are then learned using a MobileNetV3 architecture. Furthermore, we enhanced the performance of the arithmetic optimization algorithm (AOA) based on the hunger games search (HGS). In the developed method, named AOAHG, the operators of the HGS are applied to enhance the AOA’s exploitation ability while allocating the feasible region. The developed AOAG selects the most relevant features and ensures the overall model classification improvement. To assess the validity of our framework, we conducted evaluation experiments on four datasets, including ISIC-2016 and PH2 for skin cancer detection, white blood cell (WBC) detection, and optical coherence tomography (OCT) classification, using different evaluation metrics. The framework showed remarkable performance compared to currently existing methods in the literature. In addition, the developed AOAHG provided results better than other FS approaches according to the obtained accuracy, precision, recall, and F1-score as performance measures. For example, AOAHG had 87.30%, 96.40%, 88.60%, and 99.69% for the ISIC, PH2, WBC, and OCT datasets, respectively.

Full text

Citation: Elaziz, M.A.; Dahou, A.;
Mabrouk, A.; Ibrahim, R.A.; Aseeri,
A.O. Medical Image Classifications
for 6G IoT-Enabled Smart Health
Systems. Diagnostics 2023,13, 834.
https://doi.org/10.3390/
diagnostics13050834
Academic Editor: Ayman El-Baz
Received: 11 January 2023
Revised: 3 February 2023
Accepted: 19 February 2023
Published: 22 February 2023
Copyright: © 2023 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license (https://
creativecommons.org/licenses/by/
4.0/).
diagnostics
Article
Medical Image Classifications for 6G IoT-Enabled Smart
Health Systems
Mohamed Abd Elaziz 1,2,3,4,* , Abdelghani Dahou 5, Alhassan Mabrouk 6, Rehab Ali Ibrahim 1
and Ahmad O. Aseeri 7,*
1Department of Mathematics, Faculty of Science, Zagazig University, Zagazig 44519, Egypt
2Artificial Intelligence Research Center (AIRC), Ajman University, Ajman 346, United Arab Emirates
3Faculty of Computer Science & Engineering, Galala University, Suze 435611, Egypt
4Department of Electrical and Computer Engineering, Lebanese American University,
Byblos P.O. Box 36, Lebanon
5Mathematics and Computer Science Department, University of Ahmed DRAIA, Adrar 01000, Algeria
6Mathematics and Computer Science Department, Faculty of Science, Beni-Suef University,
Beni-Suef 62521, Egyptalhassanmohamed@science.bsu.edu.eg
7Department of Computer Science, College of Computer Engineering and Sciences,
Prince Sattam Bin Abdulaziz University, Al-Kharj 11942, Saudi Arabia
*Correspondence: abd_el_aziz_m@yahoo.com (M.A.E.); a.aseeri@psau.edu.sa (A.O.A.)
Abstract:
As day-to-day-generated data become massive in the 6G-enabled Internet of medical things
(IoMT), the process of medical diagnosis becomes critical in the healthcare system. This paper presents
a framework incorporated into the 6G-enabled IoMT to improve prediction accuracy and provide a
real-time medical diagnosis. The proposed framework integrates deep learning and optimization
techniques to render accurate and precise results. The medical computed tomography images are
preprocessed and fed into an efficient neural network designed for learning image representations and
converting each image to a feature vector. The extracted features from each image are then learned
using a MobileNetV3 architecture. Furthermore, we enhanced the performance of the arithmetic
optimization algorithm (AOA) based on the hunger games search (HGS). In the developed method,
named AOAHG, the operators of the HGS are applied to enhance the AOA’s exploitation ability
while allocating the feasible region. The developed AOAG selects the most relevant features and
ensures the overall model classification improvement. To assess the validity of our framework, we
conducted evaluation experiments on four datasets, including ISIC-2016 and PH2 for skin cancer
detection, white blood cell (WBC) detection, and optical coherence tomography (OCT) classification,
using different evaluation metrics. The framework showed remarkable performance compared to
currently existing methods in the literature. In addition, the developed AOAHG provided results
better than other FS approaches according to the obtained accuracy, precision, recall, and F1-score as
performance measures. For example, AOAHG had 87.30%, 96.40%, 88.60%, and 99.69% for the ISIC,
PH2, WBC, and OCT datasets, respectively.
Keywords:
Internet of medical things; 6G networks; deep learning; feature selection; arithmetic
optimization algorithm; hunger games search; metaheuristic
1. Introduction
The emergence of the Internet of medical things (IoMT) and 6G technologies has
provided the medical field with a new opportunity and methodologies to improve the
diagnosis and prediction of diseases [
1
–
4
]. A large quantity of data, including computed
tomographic (CT) images, are generated quickly at a short timescale, raising the problem
of efficiently processing such images in real-time to help the medical field detect cancerous
diseases in their early stages [
5
–
7
]. However, the availability and accessibility of medical
images have always been limited for researchers due to privacy concerns, holding back the
Diagnostics 2023,13, 834. https://doi.org/10.3390/diagnostics13050834 https://www.mdpi.com/journal/diagnostics

Diagnostics 2023,13, 834 2 of 26
desired rapid advancement in the healthcare domain. Meanwhile, CT images are of low
resolution, noisy, and difficult to process, which challenges routine diagnosis in terms of
accuracy and precision [8–11].
In the era of advanced communication technologies such as 6G, providing a real-time
medical diagnosis is a critical issue [
12
]. An early detection of diseases affecting sensitive
human body areas, such as blood, breast, lung, and skin, can help limit the spread of the
disease and protect the affected body parts. Without providing an accurate and quick
diagnosis, the spread of diseases and tumors can be enormously infectious, which can
cause a high rate of mortality [
13
]. For instance, skin cancer detection and prediction is
a significant challenge in medical imaging, which is still under development. The health
sector can benefit from the rapid development of medical equipment, communication
technologies, and the IoMT to provide quality service at an efficient scale.
The IoMT is a collection of internet-related devices that assist healthcare procedures
and activities [
14
]. The IoMT relates to employing the intelligent Internet of things (IoT) and
modern communications to service medical personnel, medicines, and medical equipment
and facilities to enable the gathering, monitoring, controlling, and a faster access to personal
health data. IoT techniques’ software in medicine covers nearly all field areas, including
physician ID, remote hospital emergency services, home healthcare of medical products
and replacement parts, hospital instruments and the clinical waste surveillance of medical
devices, blood management staff, infectious disease control, and others. In addition, the
6G-enabled IoMT provides ultrafast and accurate responses while reducing the workload
and the cost of the research and development of the medical field. It is anticipated that
the 6G communications system will play an essential role in providing the necessary
transmission rate, stability, accessibility, and architecture [
15
]. Compared to the traditional
diagnosis methods used for cancerous disease detection at an early stage, the 6G-enabled
IoMT provides a necessary platform for processing enormous healthcare data, including
hundreds of slices of CT scans [16,17].
Although the 6G-enabled IoMT has proven helpful for developing embedded systems
that can detect illness with the same precision as an expert, it relies heavily on the developed
algorithms based on deep learning (DL) and optimization algorithms [
18
]. The embedded
systems in the 6G-enabled IoT can benefit from the capability of DL methods in medical
image processing and cancerous disease identification. The adoption of DL methods may
assist in avoiding recurrent problems that require a considerable time to solve and the need
for a large number of well-labelled training data. Therefore, transfer learning (TL) can
assist in overcoming some of these issues by integrating pretrained DL models [
19
]. DL
models incorporate several techniques, such as structure design, model training, model size,
feature representation, and hyperparameter optimization. On the other hand, metaheuristic
(MH) optimization techniques have proven effective in addressing various complicated
optimization issues for computer-aided diagnosis. For instance, Silva et al. [20] improved
the hyperparameters of a CNN using particle swarm optimization (PSO) to reduce the
false positives (FPs), while detecting lung nodules in lung scans owing to their similar
patterns and low population density, which can produce misleading data. Moreover,
Surbhi et al. [
21
] utilized adaptable PSO to automatically diagnose brain tumors to reduce
noise and enhance image quality.
This paper introduces a framework to improve the diagnostic imaging identification
efficiency, designed to be integrated into the 6G-enabled IoMT. In addition, it aims to
overcome several problems, including (1) the curse of dimensionality, (2) the slow infer-
ence, and (3) the low performance. The framework comprises two phases: (1) a feature
extraction using a deep learning model with transfer learning and (2) a feature selection
using a developed optimization algorithm. In the first phase, a developed deep learning
architecture is constructed based on MobileNetV3, which acts as a core component for
feature extraction. The pretrained MobileNetV3 is employed to learn and extract medical
image representations during the deep learning model training using CT images. The
pretrained MobileNetV3 has been selected as the deep-learning-based model of choice

Diagnostics 2023,13, 834 3 of 26
due to its lightweight design that can be operated on resource-constrained devices with
limited energy and resource consumption. In the second phase, a newly developed feature
selection method is introduced to enhance the behaviour of the arithmetic optimization
algorithm using hunger games algorithms, named AOAHG. The AOAHG method selects
only the most relevant features and ensures the overall model classification improvement
and efficiency. A thorough assessment of the suggested framework is presented and com-
pared to various state-of-the-art methods utilizing four real-world datasets. In general, the
main motivation for using this combination between the AOA, HGS, and MobileNetV3
was based on the shown performance of each performance on different applications. HGS
has been applied to handle engineering problems [
22
], crisis events detection [
23
], node
clustering and multihops [
24
], feature selection [
25
], and others [
26
]. AOA has been applied
to solve structural design optimization [
27
], functionally graded material [
28
], robot path
planning [
29
], human activity recognition [
30
], and others [
31
]. Therefore, this combination
can improve the performance of medical image classifications in a 6G IoT-enabled smart
health environment. To the best our knowledge, this is the first time the HGS, AOA, and
MobileNetV3 have been integrated under a single framework for IoMT.
The main contributions of this work can be summarized as follows:
1.
Proposing a 6G-enabled IoMT method that reduces human involvement in medical
facilities while providing rapid diagnostic results. The new method is designed to be
integrated into resource-constrained systems.
2. Using the transfer learning approach to extract the features from the medical images.
3.
Enhancing the ability of the arithmetic optimization algorithm as a feature selection
technique using operators of the hunger games search.
4.
Evaluating the developed 6G-IoMT model using four datasets and comparing its
performance with other state-of-the-art techniques.
The rest of this paper provides a background on transfer learning for extracting fea-
tures. Section 4provides the developed 6G-enabled IoMT framework. Section 5discusses
the image diagnosis framework’s outcomes. Lastly, in Section 6, the conclusion and future
scope are discussed.
2. Related Works
The power of classification to aid in medical diagnosis makes it an important field of
research. As a result of classification optimization, researchers have improved classification
performance by applying deep learning and transfer learning in the IoMT. In addition, using
metaheuristic optimization algorithms in conjunction with convolution neural networks
(CNN) for medical image classification is presented in this section. Table 1summarizes the
literature review on the used datasets in our study.
2.1. IoMT-Based Deep Learning
Due to the spread of contagious diseases which can cause a pandemic, a reliable
infrastructure offering conventional diagnostic tools and systems has emerged in the IoMT.
The IoMT relies on the IoT infrastructure, which lowers the information transmission
latency and the complexity of centralized diagnosis processes. The IoMT offers several
solutions for the medical field, including monitoring systems, medical information-sharing
mediums, remote consulting, and automatic report generation. Thus, technology facilitates
the life of patients by offering to monitor and consult systems while helping the medical
staff by reducing human intervention and human faults. The IoMT collects information
from the patient using different types of sensors, devices, and clinical records, which
can be stored and shared on a cloud-based centre [
32
]. For instance, computed aided
diagnosis (CAD) technologies rely on IoT technologies to offer medical image classification,
which can be built using several IoT and deep learning techniques [
33
]. Furthermore,
self-monitoring systems can be seen as valuable components in the IoMT system such as
weight and activity monitoring in diet, cardiovascular fitness, heartbeat, and nutrition
planning programs [34–36].

Diagnostics 2023,13, 834 4 of 26
Recently, the IoMT technologies have been evolving with the development of the arti-
ficial intelligence field, especially with the breakthrough of DL algorithms [
18
]. As a result,
DL has enhanced both the specialist and the patient experience in the IoMT ecosystem
providing accurate and fast diagnosis reports and helping the early prevention of disease
spread. For instance, Rodrigues et al. [
37
] developed a vital healthcare system based on DL
techniques, such as transfer learning, to classify skin lesions. Han et al. [
38
] investigated
using DL techniques to process CT scan images and perform lung and stroke region segmen-
tation. The authors established a communication channel with the patient relying on IoT
technologies to provide diagnosis reports and consultations.
Bianchetti et al. [39]
proposed
an automated ML system using tumour histotypes of dPET (dynamic positron emission
tomography) data for adenocarcinoma lung cancer classification.
Hossen et al. [40]
de-
veloped a framework based on a federated learning approach and a convolution neural
network (CNN) to classify human skin diseases and preserve data privacy.
Unlike fully automated DL systems, the IoMT still relies on the medical expert’s
intervention to validate the results generated by a DL model or assess the accuracy of the
DL model. However, DL models have shown a remarkable and accurate performance in
many medical applications where they can help in decision-making and the early detection
of infectious diseases from big data. Thus, developing a robust DL model to perform
a specific task is vital to provide the patients with the best medicament and control the
disease in its early stages [41].
2.2. Transfer Learning on Medical Images
In recent years, pretrained models for different applications have outperformed the
regular learning process and training models from scratch. Thus, the performance on
various applications has increased, and the learning time has been reduced [
42
]. The
transfer learning process aims to transfer the knowledge learned while solving specific
tasks to a new related task. For instance, Cheplygina et al. [
43
] addressed the use of
transfer learning and different learning approaches in the medical field to perform medical
image analysis. Transfer learning can be applied while fine-tuning all or specific layers to
adapt the previously learned knowledge to the new related task. For instance, Ayan and
Ünver [
44
] fine-tuned two pretrained models, including Xception and VGG16, trained on a
large set of images from the ImageNet dataset. The fine-tuned models were used to detect
pneumonia in chest X-ray images where the VGG16 exceeded the Xception model in terms
of detection accuracy.
The ability to extract features from the VGG and ResNet models using bilinear classifi-
cation techniques combined with SVM classifiers yielded the best results on several test
sets [
45
]. A combination of data-driven approaches and InceptionV3 was used to train
roughly 13W dermatology images, with findings on the testing set comparable to those
of physicians [
46
]. Skin lesion segmentation was utilized to categorize melanoma in the
ISBI-2016 skin lesion analysis towards cancer diagnosis [
47
]. As a result of this, the final
classification had to be performed step by step. Multiple CNNs employing dynamic pattern
training were used to simulate cancer intraclass conflict and associated noise interference
in [
48
]. Kawahara et al. [
49
] decided to employ a pretrained CNN to identify skin images
throughout their entire dataset rather than starting from scratch with randomly initialized
parameters. After that pretraining, the CNN’s number of training rounds was considerably
decreased, and the accuracy percentage for five classes was 84.8%.
Lopez et al. [50]
applied
a deep learning method for early detection. It was developed using an adapted VGGNet
design and a transfer learning technique. A sensitivity value of 78.56% was achieved using
the ISIC archive dataset using the developed model. The performance of a CNN model
for detecting lesions was tested using a dataset that was both extended and unaugmented
in [
51
]. The researchers noted that deep learning approaches could be practical, and more
data had to be collected. In addition, the network performed better on the additional
dataset than other models.
Yu et al. [47]
implemented a very deep residual network-based
multistage model for automatically detecting melanomas in dermoscopy images. They

Diagnostics 2023,13, 834 5 of 26
merged VGG and ResNet networks with the SVM classifier to improve the model’s detec-
tion performance.
Zhang et al. [52]
developed a deep synergic learning (SDL) model based
on multiple deep CNNs in parallel with a sharing strategy for mutual learning. The authors
validated the model’s performance on the ImageCLEF and ISIC datasets for medical image
classification tasks.
Most of the well-known pretrained models in computer vision are based on con-
volution blocks, such as Inception, MobileNet, ResNet, DenseNet, and EfficientNet [
53
].
Furthermore, Transformer-based pretrained models were first established for language
modelling and have been widely adopted for computer vision tasks. Transformer-based
pretrained models benefit from the attention mechanism to learn contextual feature rep-
resentation. For instance, ResViT [
54
] is a residual Vision-Transformer-based model for
medical image tasks. ResViT synthesizes multimodal MRI and CT images in an adversarial
learning process that relies on residual convolutional and transformer building blocks.
2.3. Medical Images Classification Using FS Optimizers
Currently, metaheuristic (MH) optimization techniques are applied to find solutions
for different optimization problems. Those MH techniques provide a set of solutions rather
than a single answer, supporting them in efficiently exploring the search space. Thus, they
provide better results than traditional optimization approaches [55].
In the same context, Ravi K Samala et al. [
56
] presented an approach to the multi-
layered pathway used to predict breast cancer. They developed a two-stage approach
consisting of transfer learning and determining features, respectively. To train pretrained
CNNs, ROIs from large lesions were used. A random forest classification was developed
based on the learned CNN. A genetic algorithm (GA) was used to select the relevant
features.
Silva et al. [20]
optimized the hyperparameters of a CNN using PSO for the
false-positive reduction in CT lung images.
Shankar K. et al. [
57
] developed the grey wolf optimization (GWO) technique for
Alzheimer’s disease using brain imaging analysis. Then, a CNN was used to extract
the features from the retrieved images. Goel et al. [
58
] developed an OptCoNet as an
optimized CNN architecture for recognizing COVID-19 patients as having pneumonia
or not. The GWO was used to determine the parameters of the convolution layer. To
improve architectures for denoising images, Mohamed et al. [
59
] developed an enhanced
version of firefly algorithms (FFA) to categorize the images as abnormal and normal.
There was a significant enhancement in performance as a result of this adjustment. The
diagnosis of melanoma was improved using the whale optimization algorithm (WOA) and
levy
flight [60]
. These methods have some limitations, such as premature convergence,
primarily when worked in a large search space [
61
]. These limitations have a negative
impact on the prediction performance, especially in the IoMT environment. Therefore, the
main objective of this paper was to determine the best solutions to improve the convergence
rate by reducing the number of selected features.
As part of our developed study, to overcome these problems, transfer learning is
integrated with metaheuristic optimization to build the IoMT framework. The qualities
of this framework enable excellent performance and affordable computing expenses and
address the financial concerns discussed earlier. Treating and detecting infections in or
out of the clinic is essential. In order to use the IoMT system, all we need is an internet-
connected device and a digital copy of the examination. The service’s quick reply allows
for meaningful data throughout a session.

Diagnostics 2023,13, 834 6 of 26
Table 1. The literature review on selected datasets.
DS Model / Source Methodology
ISIC-2016
CUMED [47]
Integrating a fully convolutional residual network (FCRN) and other very deep residual
networks for classification.
BL-CNN [45]
Combining two different types of deep CNN (DCNN) features as local and global features,
using deep ResNet for the global features and a bilinear (BL) pooling technique to extract
local features.
DCNN-FV [62]
Integrating a ResNet method and a local descriptor encoding strategy. The local descriptors
were based on a Fisher vector (FV) encoding to build a global image representation.
MC-CNN [52]
Using multiple DCNNs simultaneously and enabling them to mutually learn from
each other
.
MFA [63]
Cross-net-based combination of several fully convolutional were suggested. Used multiple
CNNs for selecting semantic regions, local color and patterns in skin images. The FV was
used to encode the selected features.
FUSION [64]
MobileNet and DenseNet were coupled to boost feature selectivity, computation complexity,
and parameter settings.
PH2
ANN [65]
A decision support system mad a doctor’s decision easier utilizing four distinct ML algo-
rithms, where the artificial neural network (ANN) achieved the best performance.
DenseNet201-SVM [66]
U-Net was used with spatial dropout to solve the problem of overfitting, and different
augmentation effects were applied on the training images to increase the data samples.
DenseNet201-KNN [37]
Combined twelve CNN models as resource extractors with seven different classifier con-
figurations, which the greatest results obtained using the DenseNet201 model with a
KNN classifier.
ResNet50-NB [67]
A ResNet model was applied to map images and learn features through TL. The ex-
tracted features were optimized using a grasshopper optimization algorithm with a naïve
Bayes classification.
Blood-Cell
CNN-SVM [68]
A CNN with SVM-based classifiers with features derived by a kernel principal component
analysis of the intensity and histogram data was able to classify images.
CNN [69]
An SVM and a granularity feature were used to detect and classify blood cells independently.
CNNs were utilized to automatically extract high-level features from blood cells, and these
features were then used to identify the other 3 types of blood cells using a random forest.
CNN-Augmentation [
70
]
The extraction and selection of features, as well as the classification of white blood cells, were
all automated. A DL approach was used to automate the entire procedure with CNNs for
binary and multiclass classification.
3. Background
Improved deep learning for extracting features and two feature selection algorithms,
the arithmetic optimization algorithm and the hunger games search are all presented in
the following.
3.1. Enhanced Deep Learning
It has been shown that DL methods are effective in various tasks, such as the catego-
rization [
71
] and segmentation of images and object identification [
72
]. There is still much
to know about the difficulties of these activities, particularly regarding the quality and
effect of the acquired representations. Many DL architectures and learning methods have
been developed during the last decade. With its many topologies, layouts, settings, and
training procedures, the CNN is among the most studied DL models. Instead of conven-
tional convolution operation, depthwise separable convolutions may be used on embedded
devices or edge apps since they replace the existing convolutions. In order to overcome
the drawbacks of conventional convolution operation, numerous DL models have adopted
the idea of depthwise separable convolutions, such as EfficientNet [
73
]. The depthwise
separable convolutions differ from conventional convolution operations in that they are
applied individually to each input port. As a result, the models are operationally affordable

Diagnostics 2023,13, 834 7 of 26
and can be learned using lower parameters and little training time. MobileNetV3 [
74
] is
available in two structures based on the model size: MobileNetV3-large and MobileNetV3-
small. Compared to MobileNetV2, the MobileNetV3 structure is intended to reduce delay
and improve accuracy. MobileNetV3-large, for example, increased accuracy by 3.2% over
MobileNetV2 while decreasing latency by 20%. The NetAdapt method was used to find the
best network topology and kernel dimensions for the convolution layer on MobileNetV3.
The MobileNetV3 structure comprises the following fundamental components: a depth-
separable convolution operation with a specific convolution kernel, a batch normalization,
and an activation function. Next, the depthwise-separable, fully connected layer’s mutual
information calculations and the retrieval of hidden units use 1
×
1 convolutions. Third,
a global average pooling makes feature maps more manageable regarding their spatial
dimension. Furthermore, by using an inverted residual block [
75
], we may avoid the
bottlenecks caused by the residual skipped-connection method. These blocks make up
the inverted residual one: (a) using the 1
×
1 extension and convolutions, as well as the
depthwise convolution kernels of size 1
×
1, for more complicated representations and to
reduce model computations; (b) a convolutional layer with depth separation; (c) a method
for retaining a skip connection. In addition, the squeeze-and-excite (SE) block [
74
] can be
used to choose the appropriate features channel by channel. Finally, a rectified linear unit
(ReLU) and the h-swish activation function are interchangeable terms for the same thing,
the activation function.
3.2. Arithmetic Optimization Algorithm
The arithmetic optimization algorithm (AOA) [
76
] is an MH technique that depends
on essential functions to find the optimal solution. Like other MH techniques, it begins
with a randomized number of candidate alternatives (
X
) and the best-obtained or nearly
optimal solution. For the AOA to begin functioning, the search stage should be selected
first (i.e., exploration or exploitation). In the following search stages, the maths optimizer
accelerated (MO A) is used and defined as in Equation (1).
MOA(t) = Min +t×Max −Min
T(1)
The variable
t
denotes the current repetition and ranges from one to the maximum
allowable number of epochs (
T
). The terms indicate the accelerating function’s lowest and
greatest values, Min and Max.
To discover an ideal option, AOA’s exploration agents examine the research scope at
random locations across multiple areas, using two primary search methods (the divide
technique and the multiply technique described in Equation (2).
xi,j(t+1) = Xbj÷(MOP)×(U Lj×µ+LBj),r2>0.5
Xbj×MOP ×(ULj×µ+LBj),Otherwise (2)
where
ULj=UBj−LBj
. In this scenario,
xi(t+
1
)
represents the
i
th solution during the
next repetition,
xi,j(t)
represents the
j
th location of the
i
th solution in the latest iteration, and
Xbj
represents the
j
th place in the optimal method thus far.
e
is a tiny integer number. The
j
th location’s minimum and maximum bounds are denoted by
UBj
and
LBj
, respectively.
The µ=0.5 process parameters regulate the search behaviour.
MOP(t) = 1−+t1/α
T1/α(3)
where
MOP(t)
in Equation
(3)
represents the probability of the maths optimizer (
MOP
). The
current iteration is represented by
t
, while the total number of iterations is represented
by (
T
). The exploitation accuracy across iterations is defined by the sensitivity parameter
α=5.
It is necessary to do this stage of exploitation by only researching if
r
1 is less than
the existing
MOA(t)
quantity (see Equation
(1)
). In AOA, the exploitation operators

Diagnostics 2023,13, 834 8 of 26
(subtraction and addition) discover the research scope intensely across several populated
areas and methods to produce a solution based on two primary search techniques (i.e.,
subtraction and addition) that are modelled in Equation (4).
xi,j(t+1) = Xbj−MOP ×(ULj×µ+LBj),r3>0.5
Xbj+MOP ×(ULj×µ+LBj),Otherwise (4)
3.3. Hunger Games Search
The hunger games search (HGS) algorithm was developed by [
77
] as an optimization
technique that resembles organismal biology. As a result of the HGS, a creature’s capacity
to use hunger as a physiological incentive for all of these things is one of its most distin-
guishing features. HGS mathematical modelling begins with a population of
N
alternatives
X
before obtaining
Fiti
estimates for each alternative’s fitness function. The modernization
step is instead carried out using the given formula in Equation (5).
X=


X(t)×(1+rand),r1<l
W1×Xb +R×W2×Xbi,r1>l,r2>E
W1×Xb −R×W2×Xbi,r1>l,r2<E
(5)
where
Xbi =|Xb −X(t)|
. The two variables
r1
and
r2
represent random numbers, and the
parameter
rand
produces random numbers from a normally distributed set. The parameter
R
determines the search area and may be dependent on the number of rounds as defined
in Equation (6).
R=2×s×rand −s,s=2×1−t
T(6)
where Eindicates the parameter, which is specified as in Equation (7).
E=sech(|Fiti−Fitb|)(7)
Fitb
indicates the fitness function’s highest value, whereas
Sech
denotes the hyperbolic
value, defined as in Equation (8).
sech(x)=2
ex−e−x. (8)
Additionally, W1and W2are the hunger weights from Equations (9) and (10).
W1=Hi×N
SH ×r4,r3<l
1, r3>l(9)
W2=21−e(−|Hi−SH |)×r5(10)
SH
represents the solution’s hunger-experiencing accumulation, and the parameters
SH
correlate to
r3
,
r4
and
r5
being random integers with ranges in the interval [0, 1],
as follows:
SH =∑
i
Hi(11)
Hi=0, Fiti=Fitb
Hi+Hn,otherwise (12)
where Hnrepresents the new hunger, and it is formulated as:
Hn=LH ×(1+r),T H <LH
TH,otherwise (13)
TH =2Fiti−Fitb
Fitw−Fitb
×r6×(UB −LB)(14)
Moreover, there is a lower value provided by
Fitw
for the fitness function; in addition,
r6∈[
0, 1
]
is a randomised number that indicates if hunger has positive or negative effects
based on various variables.

Diagnostics 2023,13, 834 9 of 26
4. Developed Approach
To accomplish our approach, we created a 6G-enabled IoMT framework. It is capable
of transmitting data quicker than a 5G-enabled system. In bandwidth, 6G may reach
microseconds, significantly improving its speed over 5G’s milliseconds [
78
]. Furthermore,
6G enables real-time broadcast and processing better quality images and assists artificial
intelligence in achieving real-time broadcast and execution. Nevertheless, only low-latency
and high-bandwidth wireless communication technologies can satisfy the developing
requirements of DL and IoMT. Therefore, based on the 6G network and DL model concepts,
we suggest incorporating the combined DL and FS optimizer algorithms presented in the
following subsections into our 6G-enabled IoMT framework.
4.1. Feature-Extraction-Based Deep Learning
To identify and extract feature information, we utilized a transfer learning approach.
Pretrained models for image recognition tasks are helpful because they speed up training
and implication. Instead of building models from scratch, it is possible to fine-tune a few
layers while the model’s weights are fine-tuned. We replaced the model’s top part with
new layers for classification and feature extraction. MobileNetV3 was used as a core block
for extracting features after fine-tuning its weights on different task-specific datasets.
MobileNetV3 was adjusted and trained to retrieve feature representations from input
with a size equal to 224
×
224. The ImageNet data [
75
] were used to train the MobileNetV3
model and produce pretrained versions based on the model size (large or small). We
used the dataset representing images of skin cancer, blood cells, and optical tomography
to fine-tune the MobileNetV3-Large pretrained model. In our experiments, we replaced
the MobileNetV3 model’s classification layer with two layers represented as 1
×
1 point-
wise convolutions to extract the image representations and fine-tune the model for the
classification task.
The 1
×
1 pointwise convolution is often used to categorize and extract features that
have similar applications to those of multilayer perceptrons (MLPs). After fine-tuning the
MobileNetV3 layers, we fed the extracted features to a 1
×
1 pointwise convolution which
learned task-specific features. The MobileNet3 core layers are a combination of inverted
residual blocks stacked sequentially. Each inverted residual block consists of several com-
ponents derived from the MobileNetV2 structure, including a 1
×
1 expansion convolution,
a depthwise-separable convolution, a squeeze-and-excite block, a 1
×
1 projection convolu-
tion, and a skip-connection mechanism. Furthermore, a kernel of size 3
×
3 is used in the
depthwise-separable convolution with an activation function which can be placed in the
following order
(
3
×
3
Conv)→(BN)→(ReLU/h−swish)→(
1
×
1
Conv)→(BN)→
(ReLU/h−swish)
. A depth-separable fully connected layer with various nonlinearity
variables, including hard swish (h-swish) or ReLU, may be included in each construction
block. These functions are described in Equations (15) and (16).
ReLU(x) = max(0, x)(15)
h−swish(x) = x×σ(x)(16)
where
σ(x)
specifies the piecewise linear difficult analogue functional, where
σ(x) = ReLU6(x+3)
6
.
The output of the 1
×
1 pointwise convolution placed before the classification layer (1
×
1
pointwise convolution) is the feature extraction block that generates the learned image
embeddings during the network training and fine-tuning. Each extracted image embedding
is represented with a 128-feature vector. The developed model was trained on each dataset
for 100 epochs with a batch size of 32 with an early stopping strategy (20 epochs). The
RMSprop algorithm, with the learning rate of 1
×
10
−4
was applied to modify the model’s
weight and bias values. For this reason, we employed a dropout layer and data augmen-
tation with randomized horizontal flips, randomized crops, colour jitters, and periodic
vertical flips to counteract the model’s overfitting problem. The Pytorch framework was
used to implement the model, and the training was conducted on an Nvidia RTX1080 GPU.

Diagnostics 2023,13, 834 10 of 26
4.2. The Developed FS Algorithm
This article aims to provide a novel technique for enhancing the efficiency of the
arithmetic optimization algorithm (AOA). This was achieved by using the operators of the
hunger games search (HGS) algorithm. Whenever the AOA could not discover the optimal
solution within a specified iteration, a much more effective searching focused on the HGS
was applied to enhance the exploration ability. The HGS enhanced the capacity to do global
and regional searches concurrently.
The basic steps of the FS technique, called AOAHG, are shown in Figure 1. The initial
stage in the developed AOAHG was to create the set of
N
agents
X
, reflecting the FS
challenge solutions. This procedure was obtained by using the following formula:
Xi=rand ∗(U−L) + L,i=1, 2, . . . , N,j=1, 2, . . . , Dim (17)
Figure 1. Flowchart showing the developed FS algorithm.
The term
Dim
denotes the number of features. As a result, the available dimensionality
was restricted to values between
U
and
L
. We used the following equation to obtain the
binary form of each Xi:
BXi j =1i f Xij >0.5
0otherwise (18)
As a further step, we calculated the fitness value for
Xi
as in Equation (19), based on
its binary form BXi.
Fiti=λ×γi+ (1−λ)×|BXi|
Dim , (19)
In this case, the proportion of features associated is denoted as
(|BXi|
Dim )
.
γi
is the
validation loss of the SVM. In general, the SVM is often applied because it is more reliable
and has fewer parameters than other classifiers. The value of the parameter
λ
balances the
ratio between the accuracy of a classifier’s predictions and the selection of features.
The following procedure was used to adjust a solution
Xi
by using either the HGS or
AOA operators. This was accomplished through the use of the probability
Pi
associated
with each
Xi
. While the HGS may take longer, it was used if the probability of
Pi
was less
than the MO A, as defined by the given equations:

Diagnostics 2023,13, 834 11 of 26
Xij =(XA
ij i f Pi>MOA
XHG
ij otherwise (20)
where
MOA
is specified in Equation
(1)
. The value of
XA
ij
is updated using the operators of
the AOA as described in Equation (2).
XA
ij =T he f i rst r ule o f Equ ation(2)i f PA>0.5
The second rule o f Equation(2)otherwise (21)
where
PA∈[
0, 1
]
is a random sample variable that is utilised to keep the AOA operators
comparable throughout solution updates.
The HGS operators were then applied on the upgraded population of
X
. The following
formula yielded XHG
ij :
XHG
ij =W1×Xij −R×W2|Xij −X(t)|i f PH>0.5
W1×Xij +R×W2|Xij −X(t)|otherwise (22)
where
W1
and
W2
are defined in Equations
(9)
and
(10)
, respectively. If
PH
was greater than
0.5, the first HGS rule was applied; otherwise, the second HGS rule was applied.
Additionally, the search space
[U
,
L]
was dynamically changed throughout the finding
process as follows: Lj=min(Xi j)(23)
Uj=max(Xij )(24)
The next stage was to determine if the closure conditions were met, and if so, the
optimum solution was given. If this happened, the upgrade procedure was repeated from
the beginning. The suggested AOAHG’s pseudocode is given in Algorithm 1.
Algorithm 1 Pseudocode of the developed AOAHG algorithm
1: Initialize the parameters.
2: Split the dataset into training and testing sets after extracting the features.
3: Initialize the number of solutions (N).
4: repeat
5: Determine the value of the fitness function.
6: Find the best solution.
7: Update the MO A value using Equation (1).
8: Update the MOP value using Equation (3).
9: Calculate the hunger weight of each position using Equations (9) and (10).
10: Enhance Hi using Equation (12).
11: for i=1 to Ndo
12: for j=1 to Positions do
13: Generate a random values in [0, 1] (Pi,PA, and PH).
14: if Pi>MOA then
15: Position limitations can be adjusted for new seeds.
16: if PA> 0.5 then
17: Update ith solutions’ positions by the first rule in Equation (2).
18: else
19: Update ith solutions’ positions by the second rule in Equation (2).
20: else
21: if PH> 0.5 then
22: Update ith solutions’ positions by the first rule in Equation (22).
23: else
24: Update ith solutions’ positions by the second rule in Equation (22).
25: until The iteration (t) criterion has been met.
26: Return the best solution.

Diagnostics 2023,13, 834 12 of 26
4.3. Sixth-Generation-Enabled IoMT Framework
The suggested 6G-enabled IoMT architecture is shown in Figure 2. The terminal
intelligence of the IoT first collected diagnostic images, and if the expert’s aim was to learn
the framework, the input images could be transmitted through a 6G network. Then, the
data collected from the multiaccess edge-computing servers could be uploaded to a cloud
computing service.
Figure 2. The suggested 6G-enabled IoMT framework diagram.
The three primary processes were still in place in cloud computing. In the first stage,
the DL design’s features were retrieved, as discussed in Section 4.1. As a second stage, we
used the modified AOA depending on an HGS (AOAHG) to select the significant features,
as illustrated in Section 4.2. Finally, once the classifier had been learned, it could be
distributed across several API forecasting/prediction nodes, saving on transmission fees.
On the other hand, if the user’s goal was to test the case/disease of the collected image,
the test pattern in the API prediction/forecasting tools was employed. API forecasting
enabled the system’s authorized training product to forecast anything without retraining,
saving time and reducing internet traffic. Finally, the sender/specialist was given the last
diagnostic and several evaluation metrics such as accuracy, F1-score, and others to back up
the system’s forecasts.
The time complexity of the developed method depended on the AOAHG and Mo-
bileNetV3. The complexity of the developed AOAHG method was represented as
O(N×(T×D+N+
1
)
where
N
,
T
, and
D
are the number of solutions, iterations, and di-
mensions, respectively. In addition, MobileNetV3 had around 3 million
trainable parameters
.
5. Experimental Studies and Results
5.1. Dataset
Four medical datasets were employed for our experimental assessment, including a
white blood cells (WBC) dataset, retinal optical coherence tomography (OCT) images, and
skin images to identify malignant ones. To perform skin cancer classification, two datasets
of dermatoscopic images were used: PH2 [
79
] and ISIC-2016 [
80
]. Figure 3depicts a sample
of images from the tested datasets.

Diagnostics 2023,13, 834 13 of 26
Figure 3. Sample of images from: ISIC, PH2, WBC, and OCT datasets.
5.1.1. WBC Dataset
The accessible data utilized in this research were classified into four types, as described
in [
81
], and included the following: eosinophil, lymphocyte, monocyte, and neutrophil.
The WBC dataset contains microscopic images of 3120 eosinophils, 3103 lymphocytes,
3098 monocytes, and 3123 neutrophils. Each picture has a resolution of 320
×
240 pixels
and a depth of 24 bits. Furthermore, the dataset was divided into two parts: 80% for
training and 20% for testing. To be more specific, the training set had 2496 eosinophils,
2484 lymphocytes, 2477 monocytes, and 2498 neutrophils, while the testing set contained
620 from monocytes, 624 from neutrophils, also, 623 from eosinophils and lymphocytes.
5.1.2. OCT Dataset
In this section, we introduce the description of the OCT dataset, which consists of
84,484 OCT B-scans obtained using 4686 patients (collected at the Shiley Eye Institute of
the University of California, San Diego (UCSD)). These images are categorized into four
classes, DME, CNV, drusen, and normal, which contain 8866, 37,455, 11,598, and 26,565 im-
ages, respectively. Additionally, this dataset includes 968 test images and 83,516 training
images. For the training, 37,213, 11,356, 8624, and 26,323 images were used from each class,
respectively, and for the testing set, we used all images from each class.
5.1.3. PH2 Dataset
A sample size of 200 dermoscopy images was included in the PH2 dataset, compris-
ing 80 common nevus, 80 atypical nevus, and 40 melanoma. The data were split into
85% training and 15% testing sets.
5.1.4. ISIC Dataset
In all, 1179 samples were included from the ISIC-2016 dataset, which was divided into
two categories including benign and cancerous. The ISIC-2016 dataset contains 248 images
of malignant tumours and 1031 images of benign tumours. Furthermore, the data were
divided into 70% and 30% training and testing sets, respectively. For the training, we used
173 malignant and 727 benign images, whereas for the testing, we used 75 malignant and
304 benign images.

Diagnostics 2023,13, 834 14 of 26
To assess the efficiency of the developed method for classifying medical images, the
recall, precision, balanced accuracy, accuracy, and F1-score were used.
5.2. Experimental Results and Discussion
This section summarises the results of the experiments conducted to evaluate the
efficiency of the developed 6G-IoMT approach. We assessed our developed FS method
against other FS based on MH approaches, including the Aquila optimizer (AO) [
82
], PSO,
GWO, moth-flame optimization (MFO) [
83
], bat algorithm (BAT) [
84
], Archimedes opti-
mization algorithm (ArchOA) [
85
], chaos game optimization (CGO) [
86
], hunger games
search (HGS) [
77
], and arithmetic optimization algorithm (AOA) [
76
]. After that, extreme
gradient boosting (XGB), the K-nearest neighbours (KNN), random forest (RF), and sup-
port vector machine (SVM) classifiers were assessed against each other. All tests used a
population size of 50 and 20 iterations. The other parameters were set according to the
original implementation.
5.2.1. Results of FS Methods
Results from the ISIC-2016 dataset and PH2 dataset can be found in Table 2. Table 3
contains the findings from the WBC dataset and OCT dataset.
From Table 2, the SVM-based AOAHG provided better results than other approaches
on the ISIC-2016 dataset. The accuracy of the AOAHG algorithm using the SVM was
87.34%, representing the best efficiency, followed by MFO, which achieved the second rank
with 86.54%. ArchOA, AOA, and HGS followed the previous two algorithms (AOAHG and
MFO). The BAT and CGO algorithms, with 86.02%, followed the preceding methods. The
algorithms that followed were the PSO (85.75%), AO (85.49%), HGS (84.96%), and GWO
(84.43%) algorithms. For the value of precision, the developed AOAHG method achieved
86.53%, followed by the MFO with an 85.60% accuracy.
The results of the recall of the AOAHG method were better than others. The developed
algorithm was followed by MFO, which had a success rate of 85.54%. The ArchOA and
AOA all had the same value of recall, that is, 86.28%. For the vote, the PSO obtained 86.02%,
and both CGO and BAT achieved 85.75%. Finally, the GWO algorithm had the worse
outcome at 84.43%. The presented AOAHG method outscored the other methods with an
F1-score of 86.47%. The AOA and ArchOA obtained the exact value of 85.73%. Next, MFO,
CGO, BAT, and AO had an F1-score of 85.57%, 85.50%, 85.00%, and 84.86%, respectively.
For the PH2 dataset, Table 2illustrates that the AOAHG method significantly improved
the determination of features when using an SVM classification method; this was evident
across all metrics. According to the accuracy measure, AOAHG correctly classified 96.43%
of the testing samples when an SVM was used. In addition, these results were significantly
different from the accuracy of other FS approaches. Moreover, the AOAHG had the
better precision value of any SVM methods at 96.44%, the highest of any other optimizer
algorithm. AOA, CGO, BAT, ArchOA, MFO, GWO, and PSO placed second. Following
these optimizers were the AO and the HGS methods, which achieved an 96.70% accuracy.
As a further analysis, the recall metric for the SVM classification model was 96.43% for
AOAHG, which indicated that the developed method had the maximum effectiveness. The
developed AOAHG method was the best optimizer based on the F1-score with 96.43%.
The PSO, GWO, MFO, ArchOA, AO, BAT, HGS, CGO, and AOA approaches had an F1-
score of nearly 96.07%. Furthermore, the presented AOAHG method had the highest
balanced accuracy value, nearly 97.02%. These algorithms came in second place for all the
other optimization algorithms with 96.73%. Nevertheless, when these ten optimizers were
combined with the KNN, XGB, and RF classifiers, the outcomes had the poorest overall
performance measures compared to those of the SVM classification algorithm.

Diagnostics 2023,13, 834 15 of 26
Table 2.
Classification results (%) of each FS algorithm using two skin datasets (ISIC and PH2). ET is
the execution time.
Alg. Cls. ISIC PH2
AC BA F R P ET AC BA F R P ET
PSO
SVM 85.75 72.04 84.78 85.75 84.68 0.13 96.07 96.73 96.07 96.07 96.10 0.12
XGB 84.43 73.72 84.14 84.43 83.92 0.26 92.86 93.45 92.88 92.86 93.40 3.61
KNN 84.17 71.55 83.53 84.17 83.21 0.08 95.71 96.13 95.72 95.71 95.77 0.24
RF 85.22 74.72 84.90 85.22 84.68 0.33 91.79 91.67 91.85 91.79 92.60 0.56
GWO
SVM 84.43 71.21 83.65 84.43 83.34 0.16 96.07 96.73 96.07 96.07 96.10 0.09
XGB 82.85 71.73 82.62 82.85 82.43 0.22 92.86 93.75 92.85 92.86 93.43 2.57
KNN 84.17 72.55 83.73 84.17 83.45 0.06 95.71 96.13 95.72 95.71 95.77 0.17
RF 85.22 73.71 84.72 85.22 84.46 0.30 91.79 91.67 91.85 91.79 92.60 0.48
MFO
SVM 86.54 73.03 85.57 86.54 85.60 0.15 96.07 96.73 96.07 96.07 96.10 0.09
XGB 85.49 76.39 85.38 85.49 85.28 0.23 93.21 94.05 93.22 93.21 93.58 2.63
KNN 82.59 71.07 82.31 82.59 82.08 0.06 95.71 96.13 95.72 95.71 95.77 0.17
RF 85.22 73.71 84.72 85.22 84.46 0.31 92.14 91.96 92.21 92.14 92.87 0.50
ArchOA
SVM 86.28 74.88 85.73 86.28 85.51 0.09 96.07 96.73 96.07 96.07 96.10 0.06
XGB 83.64 73.23 83.47 83.64 83.32 0.19 91.79 92.56 91.80 91.79 92.60 1.69
KNN 84.96 74.05 84.59 84.96 84.35 0.05 95.36 95.83 95.37 95.36 95.40 0.11
RF 85.75 74.55 85.27 85.75 85.03 0.29 93.57 93.75 93.61 93.57 94.00 0.45
OA
SVM 85.49 73.38 84.86 85.49 84.60 0.14 96.07 96.73 96.07 96.07 96.07 0.06
XGB 84.70 73.39 84.27 84.70 84.01 0.27 93.57 94.35 93.58 93.57 93.97 1.81
KNN 85.49 74.38 85.04 85.49 84.79 0.07 95.71 96.13 95.72 95.71 95.77 0.12
RF 86.28 75.88 85.90 86.28 85.69 0.33 92.14 91.96 92.21 92.14 92.87 0.47
BAT
SVM 86.02 72.20 85.00 86.02 84.97 0.10 96.07 96.73 96.07 96.07 96.10 0.08
XGB 85.75 75.05 85.36 85.75 85.13 0.19 92.86 93.75 92.86 92.86 93.28 2.48
KNN 82.32 68.39 81.55 82.32 81.12 0.05 95.71 96.13 95.72 95.71 95.77 0.16
RF 85.75 75.05 85.36 85.75 85.13 0.28 92.50 92.26 92.56 92.50 93.15 0.46
HGS
SVM 84.96 73.05 84.40 84.96 84.12 0.12 96.07 96.73 96.07 96.07 96.07 0.07
XGB 85.49 75.39 85.22 85.49 85.02 0.23 92.50 93.15 92.52 92.50 93.15 2.07
KNN 84.96 73.05 84.40 84.96 84.12 0.07 95.71 96.13 95.72 95.71 95.77 0.15
RF 84.96 74.05 84.59 84.96 84.35 0.30 92.14 92.56 92.18 92.14 92.85 0.47
CGO
SVM 86.02 74.71 85.50 86.02 85.27 0.14 96.07 96.73 96.07 96.07 96.10 0.07
XGB 84.96 73.05 84.40 84.96 84.12 0.22 93.57 94.35 93.58 93.57 93.97 2.15
KNN 84.96 73.55 84.50 84.96 84.23 0.06 95.71 96.13 95.72 95.71 95.77 0.14
RF 85.22 73.21 84.63 85.22 84.36 0.30 92.86 92.56 92.91 92.86 93.44 0.46
AOA
SVM 86.28 74.88 85.73 86.28 85.51 0.16 96.07 96.73 96.07 96.07 96.10 0.11
XGB 85.75 73.54 85.08 85.75 84.85 0.25 93.93 94.64 93.94 93.93 94.27 2.86
KNN 84.96 71.04 83.99 84.96 83.78 0.07 95.71 96.13 95.72 95.71 95.77 0.18
RF 85.49 73.88 84.95 85.49 84.69 0.31 91.43 91.37 91.50 91.43 92.33 0.50
AOAHG
SVM 87.34 74.53 86.47 87.34 86.53 0.06 96.43 97.02 96.43 96.43 96.44 0.10
XGB 84.43 72.72 83.96 84.43 83.67 0.09 93.57 94.35 93.58 93.57 93.97 3.01
KNN 84.17 72.55 83.73 84.17 83.45 0.04 95.71 96.13 95.72 95.71 95.77 0.19
RF 85.75 75.05 85.36 85.75 85.13 0.26 91.79 91.67 91.85 91.79 92.60 0.53

Diagnostics 2023,13, 834 16 of 26
Table 3.
Classification Results (%) of each FS algorithm using WBC dataset and OCT dataset. ET is
the execution time.
Alg. Cls. WBC OCT
AC BA F R P ET AC BA F R P ET
PSO
SVM 88.46 88.46 88.65 88.46 90.49 1.1 99.28 99.28 99.28 99.28 99.30 16
XGB 88.42 88.41 88.64 88.42 90.60 58.0 99.17 99.17 99.18 99.17 99.20 178
KNN 88.42 88.42 88.61 88.42 90.44 8.4 99.28 99.28 99.28 99.28 99.30 3
RF 88.46 88.46 88.66 88.46 90.53 6.4 99.28 99.28 99.28 99.28 99.30 104
GWO
SVM 88.50 88.50 88.69 88.50 90.51 0.9 99.38 99.38 99.38 99.38 99.40 11
XGB 88.42 88.42 88.65 88.42 90.63 44.3 99.38 99.38 99.38 99.38 99.40 137
KNN 88.50 88.50 88.69 88.50 90.55 6.5 99.17 99.17 99.18 99.17 99.20 2
RF 88.42 88.41 88.61 88.42 90.43 5.5 99.38 99.38 99.38 99.38 99.40 90
MFO
SVM 88.54 88.54 88.74 88.54 90.59 0.8 99.38 99.38 99.38 99.38 99.40 9
XGB 88.50 88.50 88.71 88.50 90.58 42.5 99.38 99.38 99.38 99.38 99.40 119
KNN 88.50 88.50 88.70 88.50 90.60 6.1 99.17 99.17 99.18 99.17 99.20 2
RF 88.46 88.46 88.66 88.46 90.55 5.4 99.17 99.17 99.18 99.17 99.20 89
ArchOA
SVM 88.26 88.25 88.44 88.26 90.32 0.4 99.38 99.38 99.38 99.38 99.40 7
XGB 88.30 88.30 88.52 88.30 90.45 15.7 99.28 99.28 99.28 99.28 99.30 90
KNN 88.58 88.58 88.75 88.58 90.55 1.9 99.17 99.17 99.18 99.17 99.20 1
RF 88.46 88.46 88.67 88.46 90.58 3.3 99.17 99.17 99.18 99.17 99.20 74
AO
SVM 88.26 88.25 88.47 88.26 90.44 0.6 99.28 99.28 99.28 99.28 99.30 9
XGB 88.34 88.33 88.56 88.34 90.48 29.9 99.38 99.38 99.38 99.38 99.40 125
KNN 88.46 88.46 88.65 88.46 90.51 4.0 99.28 99.28 99.28 99.28 99.30 2
RF 88.34 88.33 88.55 88.34 90.48 4.4 99.48 99.48 99.48 99.48 99.49 76
BAT
SVM 88.34 88.34 88.51 88.34 90.23 0.8 99.38 99.38 99.38 99.38 99.40 12
XGB 88.06 88.05 88.31 88.06 90.42 43.0 99.28 99.28 99.28 99.28 99.30 141
KNN 88.46 88.46 88.63 88.46 90.41 6.5 99.48 99.48 99.48 99.48 99.49 2
RF 88.30 88.29 88.52 88.30 90.44 5.5 99.28 99.28 99.28 99.28 99.30 88
HGS
SVM 88.54 88.54 88.74 88.54 90.62 0.4 99.59 99.59 99.59 99.59 99.59 9
XGB 88.26 88.25 88.47 88.26 90.39 20.5 99.38 99.38 99.38 99.38 99.40 122
KNN 88.38 88.38 88.58 88.38 90.46 2.7 99.17 99.17 99.18 99.17 99.20 2
RF 88.46 88.46 88.67 88.46 90.62 4.0 99.38 99.38 99.38 99.38 99.40 86
CGO
SVM 88.50 88.50 88.68 88.50 90.49 1.0 99.59 99.59 99.59 99.59 99.59 10
XGB 87.94 87.93 88.21 87.94 90.41 36.6 99.17 99.17 99.18 99.17 99.20 125
KNN 88.22 88.21 88.43 88.22 90.32 5.4 99.17 99.17 99.18 99.17 99.20 2
RF 88.22 88.21 88.44 88.22 90.41 5.2 99.17 99.17 99.18 99.17 99.20 86
AOA
SVM 88.58 88.58 88.76 88.58 90.57 0.8 99.38 99.38 99.38 99.38 99.40 12
XGB 88.42 88.41 88.62 88.42 90.54 47.6 99.17 99.17 99.18 99.17 99.20 142
KNN 88.42 88.42 88.61 88.42 90.44 7.4 99.28 99.28 99.28 99.28 99.30 2
RF 88.42 88.41 88.62 88.42 90.51 5.9 99.07 99.07 99.07 99.07 99.10 86
AOAHG
SVM 88.62 88.62 88.80 88.62 90.59 1.0 99.69 99.69 99.69 99.69 99.69 6
XGB 88.30 88.30 88.51 88.30 90.40 48.7 99.38 99.38 99.38 99.38 99.40 68
KNN 88.46 88.46 88.63 88.46 90.42 6.7 99.59 99.59 99.59 99.59 99.59 1
RF 88.38 88.37 88.59 88.38 90.52 5.8 99.38 99.38 99.38 99.38 99.40 62
Table 3shows the comparison between the AOAHG approach and other optimizers
using the white blood cell dataset. Based on the results, the AOAHG algorithm based
on an SVM provided a better accuracy (nearly 88.62%) than the other algorithms. The

Diagnostics 2023,13, 834 17 of 26
AOA was second with an accuracy of 88.58%. Then, the MFO and HGS techniques had the
same outcome (i.e., 88.54%), and both the CGO and GWO approaches obtained the same
accuracy value, 88.50%. In addition, from those results, it can be noticed that the ArchOA
and AOA had the worst score at 88.26%.
Moreover, the recall values of AOAHG were better than those of the other methods.
The AOA, HGS, and MFO methods all had similar recall values (i.e., around 88.54%).
Finally, the AO and ArchOA had a worse outcome of 88.26%. The developed AOAHG
technique outperformed other methods according to the F1-score, which was 88.80%. The
AOA was second, with 88.76%, and the ArchOA obtained the worst F1-score value of
88.44%. Based on the results of the balanced accuracy, the AOAHG algorithm provided
better results with an accuracy of 88.62%. Similarly, according to the balanced accuracy, the
AOA was ranked second (88.58%).
Table 3shows the results of the algorithms applied on the OCT dataset. From those
results, it can be noticed that the AOAHG method was better than the other optimization
methods. The best performance based on the accuracy measure was the AOAHG approach
using the SVM with a 99.69% accuracy. At the same time, the CGO and HGS methods
were ranked second with an accuracy 99.59%. Based on the precision value, the developed
AOAHG approach achieved a score of 99.69%. The CGO and HGS algorithms followed
with a precision of 99.59%. The precision of the AOA, BAT, ArchOA, MFO, and GWO
algorithms was 99.40%. On the other side, the AOAHG approach had the highest recall
measure performance of any SVM classifiers at 99.69%. Coming in second were HGS and
CGO, which both had a recall of 99.59%. Five optimizers had a common recall value of
almost 99.38%: GWO, MFO, ArchOA, BAT, and AOA. On the other hand, PSO and AO
performed the worst at 99.28%. Our new algorithm (i.e., AOAHG) outperformed earlier
ones with 99.69% on the F1-score metric. Following that, CGO and HGS both obtained
99.59% for the F1-score. Nevertheless, that was not the end of it. With a result of 99.28%,
AO and PSO came in dead last in the competition. The AOAHG algorithm had the most
excellent actual quality, with a balanced accuracy of 99.69%. HGS and CGO obtained
99.59% and came in second. Following that, AOA, BAT, ArchOA, MFO, and GWO had a
balanced accuracy of 99.38%. However, PSO and AO had the worst results, with a balanced
accuracy of just 99.28%.
From a different viewpoint, as shown in Figure 4, the average outcomes of the ten
feature selection optimizers investigated on the four classifiers (i.e., SVM, KNN, RF, and
XGB) on the four chosen datasets, PH2, ISIC, WBC, and OCT, are given in Figure 4. From
Figure 4a it can be noticed that the overall average accuracy on the PH2 dataset was nearly
96.11% and 95.68% for the SVM and KNN, respectively. In addition, the overall balanced
accuracy of the SVM classifier was the best (96.76%). It was followed by the KNN (96.10%),
the XGB (93.84%), and the RF (92.14%) classifiers. Moreover, the best F1-score from the
ten optimization techniques was obtained by the SVM at about 96.11%; the KNN was
second with 95.69%. Furthermore, the XGB outperformed the RF algorithm, with a success
rate of 93.08% for XGB and 92.27% for RF, whereas the SVM was better than the other
classifiers based on the recall value. Furthermore, the SVM achieved 96.11%, while the
KNN classifier achieved 95.68%. Finally, the XGB and RF algorithms obtained 93.07%
and 92.22%, respectively. In terms of precision measure, the SVM classification algorithm
delivered superior results compared to the KNN, XGB, and RF classifiers, with 95.73%,
93.56%, and 92.93%, respectively.

Diagnostics 2023,13, 834 18 of 26
(a) PH2 (b) ISIC
(c) WBC (d) OCT
Figure 4. Average results from the four classifiers on the selected datasets.
As shown in Figure 4b, the average accuracy of the ten optimization techniques on
the ISIC dataset using the SVM was 85.91%; the RF algorithm took second place with
85.49%. Furthermore, the XGB achieved 84.75%, outperforming the KNN, which achieved
84.28%. Moreover, the RF was better than the other classifiers in terms of balanced accuracy.
To be more specific, the RF achieved 74.38%, while the XGB achieved 73.82%. Finally,
the SVM and KNN algorithms obtained 73.39% and 72.22%, respectively. Regarding the
F1-score measure, the SVM classification algorithm delivered superior results compared to
those of the RF, XGB, and KNN classifiers, with 85.04%, 84.39%, and 83.74%, respectively.
Additionally, the overall average recall was approximately 85.91% for the SVM classifier,
whereas the RF classifier came in second with 85.49%. XGB obtained a higher percentage
(with 84.75%) compared to the KNN classification algorithm. In addition, the SVM classifier
precision was the highest at 85.01%. It was followed by the RF (84.08%), the XGB (84.18%),
and the KNN (83.46%).
As shown in Figure 4c, the SVM classifier achieved the highest accuracy on the WBC
dataset with 88.46% followed by the KNN (88.44%), RF (88.39%), and XGB (88.30%) classi-
fiers, respectively. Meanwhile, the BA metric of all optimization approaches outperformed
the SVM classifier by 88.46% and the KNN classifier by 88.44%, respectively. The RF classi-
fier surpassed the XGB classifier, with an accuracy equal to 88.39% for RF and 88.29% for
XGB. The SVM classifier scored an average F1-score equal to 88.65%, whereas the KNN
scored 88.63%. The RF obtained a higher F1-score (with 88.60%) than the XGB classifica-
tion algorithm. The SVM classification algorithm delivered better results than the KNN,
RF, and XGB classifiers, with a recall equal to 88.44%, 88.39%, and 88.30%, respectively.
Additionally, the RF was better than the other classifiers based on the precision score. For
instance, the RF achieved 90.51%, while the KNN and SVM achieved 90.49%. Finally, the
XGB algorithm obtained 90.47%.
As shown in Figure 4d, the SVM classification algorithm delivered a superior average
accuracy score on the OCT dataset compared to XGB, KNN, and RF classifiers, with 99.30%,
99.28%, and 99.28%, respectively. The average balanced accuracy was 99.43% for the SVM
classifier, whereas the XGB classifier came in second with 99.30%. KNN obtained a lower
average balanced accuracy (with 99.28%) than the RF classification algorithm. From a
different perspective, the overall F1-score of the SVM classifier was the best (99.43%). It
was preceded by the XGB (99.30%), the KNN (99.28%), and the RF (99.28%) classifiers.
Meanwhile, the SVM scored a better recall than the other classifiers, equal to 99.43%, while
the recall for XGB was 99.30%. The KNN and RF algorithms obtained 99.28%. Moreover,
the SVM classifier achieved an average precision equal to 99.45%, followed by the XGB

Diagnostics 2023,13, 834 19 of 26
classifier with 99.32%. In addition, the KNN and RF classifiers achieved the same precision
at 99.30%.
Figure 5presents the average accuracy of the experimented classifiers on the four
datasets using different optimization strategies. As shown in Figure 5, the SVM showed
a significantly better performance compared to the other classifiers in terms of accuracy.
To be more precise, the SVM had an accuracy of 92.48%, while the KNN method had an
accuracy of 91.92%. Finally, the XGB and RF algorithms achieved an accuracy of 91.35%
and 91.34%, respectively.
Figure 5. Average accuracy on the four classifiers.
The whole operation took less time to finish than it did for a consumer, as shown
in Figure 6, which displays the average execution time for the optimizers on the selected
datasets. According to the findings, the KNN classification model was the fastest (i.e., less
time), followed by the SVM classification algorithm, which required 2.7829 s to finish. In
all, 22.5073 s were needed to complete the RF classification. The XGB required the longest
duration at 41.4983 s.
Figure 6. Average execution time of the classifiers across the datasets.
For the four datasets, using the SVM classifier, the developed AOAHG and ArchOA
took an average of 1.7031 and 1.9405 s to execute, as shown in Figure 7. Overall, these
times were faster than those of other similar methods. The AO optimizer completed
in
2.3397 s
, while the HGS, MFO, CGO, GWO, BAT, and AOA optimizers completed in
2.3704 s
,
2.4415 s
, 2.7837 s, 3.1439 s, 3.3205 s, and 3.3388 s, respectively. For instance, the
PSO algorithm achieved the longest execution time (4.4469 s).

Diagnostics 2023,13, 834 20 of 26
Figure 7. Average execution times of the SVM across the datasets.
From a different viewpoint, Figure 8illustrates each feature selection strategy on four
datasets and the corresponding average accuracy. Using various optimizers, the SVM
classifier outperformed the AOAHG technique on average with a 93.02% accuracy. The
MFO method ranked second with an accuracy of 92.63%. The AOA outperformed the
CGO, ArchOA, and BAT algorithm, averaging accuracies of 92.55%, 92.50%, and 92.45%,
respectively. The PSO obtained an accuracy of 92.39%. These three optimizers produced
the worst results, with an average balanced accuracy of 92.29% (HGS), 92.28% (AO), and
92.1% (GWO).
Figure 8. Average accuracy of the SVM across the datasets.
To summarise, for the ISIC-2016, PH2, WBC, and OCT datasets, the AOAHG algorithm
alongside the SVM classifier obtained the best accuracy. In addition, our suggested method
also produced the quickest results (i.e., the least execution time).
5.2.2. Compared Methods
Other medical image categorization approaches are examined in this section to com-
pare our developed method. Table 4summarises the findings of many critical methods.
Medical image categorization requires the development of highly accurate technologies.
Comparing our approach to other models evaluated on the same datasets is critical. Table 4
compares the accuracy of various illness detection approaches using the ISIC, PH2, WBC,
and OCT datasets.

Diagnostics 2023,13, 834 21 of 26
Table 4. Accuracy (AC) results of the developed method and other existing methods.
DS Model AC (%) Year Ref.
ISIC
BL-CNN 85.00 2017 [45]
DCNN-FV 86.81 2018 [62]
MC-CNN 86.30 2019 [52]
MFA 86.81 2020 [63]
AOAHG + SVM 87.30 present Ours
PH2
ANN 92.50 2017 [65]
DenseNet + SVM 92.00 2020 [66]
DenseNet + KNN 93.16 2020 [37]
ResNet + NB 95.40 2021 [67]
AOAHG + SVM 96.40 present Ours
WBC
CNN + SVM 85.00 2013 [68]
CNN 87.08 2017 [69]
CNN + Augm 87.00 2019 [70]
AOAHG + SVM 88.60 present Ours
OCT
Transfer Learning 80.30 2018 [87]
IFCNN 87.30 2019 [88]
LGCNN 89.90 2019 [89]
IBDL 94.57 2018 [87]
ScSPM 97.75 2017 [90]
InceptionV3 98.86 2018 [91]
AOAHG + SVM 99.69 present Ours
The ISIC dataset was used to evaluate various skin cancer detection techniques, in-
cluding integrated feature fusion [
45
], corroborated by Fisher coding and extensive deep
networks [
62
], interactive model of multi-CNN learning [
52
], and fusing Fisher vectors
with CNN data [63].
In [
65
], the authors built a decision framework using a convolutional neural network
to evaluate the PH2 dataset for skin cancer detection. A U-Net could automatically identify
malignant tumours, according to [
66
]. Rodrigues et al. [
37
] used transfer learning and
a CNN as components of their IoT architecture. A hierarchical structure based on two-
dimensional elements in the picture and ResNet were presented in [
67
] for improved deep
learning. The following identification techniques were utilized to recognize and estimate
essential blood cells in the WBC dataset. A CNN was used to perform classification, as
described in [
68
]. Additionally, Ref. [
69
] took advantage of a selectivity feature and an SVM.
CNNs were offered as a deep learning strategy in [
70
] for automating the whole operation.
Six well-known classification algorithms were implemented and tested in [
87
] to vali-
date their performance on the OCT dataset. The algorithms included transfer learning [
87
]
and IFCNN [
88
]. IFCNN [
88
] used numerous convolutional features inside of a CNN and a
recurrent fusion method to identify OCT images. Huang et al. [
89
] devised a special layer
guided convolutional neural network (LGCNN) to discriminate between the typical retina
and three common macular disorders. Kermany et al. [
87
] introduced an image-based deep
learning (IBDL) technique that adjusted the channel’s parameters and was utilized as a
feature representation. Sun et al. [
90
] used sparse coding and dictionary learning based on
the scale-invariant feature transform (SIFT) metaphor to identify AMD, DME, and routine
images. Ji et al. [
91
] used Inception V3 via transfer learning as a feature extractor where a

Diagnostics 2023,13, 834 22 of 26
CNN was added on top of the pretrained Inception V3 network after eliminating the top
layers to detect feature-space alterations.
Overall, our technique allows us to eliminate unnecessary features from high-dimensional
representations of the input medical image extracted from a CNN network. Nevertheless,
our framework’s primary flaw is that it is time- and memory-intensive. The next stage is to
simplify the framework and make it more efficient. Other techniques of augmentation may be
investigated in the future to further enhance our current system.
6. Conclusions and Future Works
The attractive characteristics of 6G compared to earlier generations of wireless net-
works have lately generated a considerable attention in business and academic fields. In
our study, the developed framework depended on the cloud centre’s classification models
being trained before they could be put to work. They were then transmitted to the cloud
centre after being analysed at the cloud centre using learned representations from medical
images obtained from edge devices on IoT/fog computing nodes. MobileNetV3 was modi-
fied and fine-tuned using medical pictures to determine more complex and informative
representations and extract image embedding. Furthermore, a novel metaheuristic algo-
rithm that relied on the arithmetic optimization algorithm (AOA) and hunger games search
was developed as a feature selection method to filter only relevant features from image
embedding. Convergence was accelerated, and feature vectors were improved as a result.
In order to determine how well the developed framework model performed, it was sent
to a simulated medical imaging cloud centre or assessed using fog computing and a copy
of the developed algorithm. The developed framework was tested on the ISIC-2016, PH2,
WBC, and OCT datasets. The results showed that the presented technique outperformed
other methods already used for feature selection. In addition, the results of the assessments
with other new medical image categorization technologies showed that the IoMT technique
developed could enhance the overall performance and services. An increased amount of
medical information, as well as its use in medical treatment, will be assessed as part of
future research. Combining multiple classification techniques is also an intriguing research
topic since it may enable practitioners to improve the performance of current approaches.
In addition, the hyperparameters optimization of deep learning models can be investigated
as using the wrong hyperparameters can limit the performance of the model.
Author Contributions:
Conceptualization, M.A.E., A.D., A.M., R.A.I. and A.O.A.; methodology,
M.A.E., A.D., A.M., R.A.I. and A.O.A.; software, M.A.E., A.D., A.M. and R.A.I.; validation, M.A.E.,
A.D., A.M. and R.A.I.; formal analysis, M.A.E., A.D., A.M., R.A.I. and A.O.A.; investigation, M.A.E.,
A.D., A.M., R.A.I. and A.O.A.; writing—original draft preparation, M.A.E., A.D., A.M. and R.A.I.;
writing—review and editing, M.A.E., A.D., A.M., R.A.I. and A.O.A.; visualization, M.A.E., A.D.,
A.M., R.A.I. and A.O.A.; supervision, M.A.E., A.D., A.M. and R.A.I.; project administration, M.A.E.,
A.D., A.M., R.A.I. and A.O.A.; and funding acquisition, A.O.A. All authors have read and agreed to
the published version of the manuscript.
Funding:
The authors extend their appreciation to the Deputyship for Research & Innovation,
Ministry of Education in Saudi Arabia for funding this research work through the project number
(IF-PSAU-2022/01/19574).
Data Availability Statement: The data are available from the authors upon request.
Conflicts of Interest:
The authors declare that there are no conflicts of interest regarding the publica-
tion of this paper.
References
1.
Dao, N.N. Internet of wearable things: Advancements and benefits from 6G technologies. Future Gener. Comput. Syst.
2022
,
138, 172184. [CrossRef]
2.
Koundal, D.; Sharma, B.; Guo, Y. Intuitionistic based segmentation of thyroid nodules in ultrasound images. Comput. Biol. Med.
2020,121, 103776. [CrossRef]

Diagnostics 2023,13, 834 23 of 26
3.
Singh, K.; Sharma, B.; Singh, J.; Srivastava, G.; Sharma, S.; Aggarwal, A.; Cheng, X. Local statistics-based speckle reducing
bilateral filter for medical ultrasound images. Mob. Netw. Appl. 2020,25, 2367–2389. [CrossRef]
4.
Zhang, G.; Navimipour, N.J. A comprehensive and systematic review of the IoT-based medical management systems: Applica-
tions, techniques, trends and open issues. Sustain. Cities Soc. 2022,2022, 103914. [CrossRef]
5.
Krishnadas, P.; Chadaga, K.; Sampathila, N.; Rao, S.; Prabhu, S. Classification of Malaria Using Object Detection Models.
Informatics 2022,9, 76. [CrossRef]
6.
Sampathila, N.; Chadaga, K.; Goswami, N.; Chadaga, R.P.; Pandya, M.; Prabhu, S.; Bairy, M.G.; Katta, S.S.; Bhat, D.; Upadya, S.P.
Customized Deep Learning Classifier for Detection of Acute Lymphoblastic Leukemia Using Blood Smear Images. Healthcare
2022,10, 1812. [CrossRef]
7.
Acharya, V.; Dhiman, G.; Prakasha, K.; Bahadur, P.; Choraria, A.; Prabhu, S.; Chadaga, K.; Viriyasitavat, W.; Kautish, S.;
Sushobhitha , M.; et al. AI-assisted tuberculosis detection and classification from chest X-rays using a deep learning normalization-
free network model. Comput. Intell. Neurosci. 2022,2022, 2399428. [CrossRef]
8.
Faruqui, N.; Yousuf, M.A.; Whaiduzzaman, M.; Azad, A.; Barros, A.; Moni, M.A. LungNet: A hybrid deep-CNN model for lung
cancer diagnosis using CT and wearable sensor-based medical IoT data. Comput. Biol. Med. 2021,139, 104961. [CrossRef]
9.
Hu, M.; Zhong, Y.; Xie, S.; Lv, H.; Lv, Z. Fuzzy system based medical image processing for brain disease prediction. Front.
Neurosci. 2021,15, 714318. [CrossRef]
10.
Aurna, N.F.; Yousuf, M.A.; Taher, K.A.; Azad, A.; Moni, M.A. A classification of MRI brain tumor based on two stage feature level
ensemble of deep CNN models. Comput. Biol. Med. 2022,146, 105539. [CrossRef]
11.
Yang, J.; Shi, R.; Wei, D.; Liu, Z.; Zhao, L.; Ke, B.; Pfister, H.; Ni, B. MedMNIST v2-A large-scale lightweight benchmark for 2D
and 3D biomedical image classification. Sci. Data 2023,10, 41. [CrossRef]
12.
Nayak, S.; Patgiri, R. 6G communication technology: A vision on intelligent healthcare. In Health Informatics: A Computational
Perspective in Healthcare; Springer: Berlin/Heidelberg, Germany, 2021; pp. 1–18.
13.
Eid, M.M.; Rashed, A.N.Z.; Bulbul, A.A.M.; Podder, E. Mono-rectangular core photonic crystal fiber (MRC-PCF) for skin and
blood cancer detection. Plasmonics 2021,16, 717–727. [CrossRef]
14.
Jin, B.; Zhao, Y.; Liang, Y. Internet of things medical image detection and pediatric renal failure dialysis complicated with
respiratory tract infection. Microprocess. Microsyst. 2021,83, 104016. [CrossRef]
15.
Wang, W.; Liu, F.; Zhi, X.; Zhang, T.; Huang, C. An Integrated deep learning algorithm for detecting lung nodules with low-dose
CT and its application in 6G-enabled internet of medical things. IEEE Internet Things J. 2020,8, 5274–5284. [CrossRef]
16.
Abd Elaziz, M.; Mabrouk, A.; Dahou, A.; Chelloug, S.A. Medical Image Classification Utilizing Ensemble Learning and Levy
Flight-Based Honey Badger Algorithm on 6G-Enabled Internet of Things. Comput. Intell. Neurosci.
2022
,2022, 5830766. [CrossRef]
17.
Mabrouk, A.; Dahou, A.; Elaziz, M.A.; Díaz Redondo, R.P.; Kayed, M. Medical Image Classification Using Transfer Learning and
Chaos Game Optimization on the Internet of Medical Things. Comput. Intell. Neurosci. 2022,2022, 9112634. [CrossRef]
18.
Liu, Y.; Wang, J.; Li, J.; Niu, S.; Song, H. Machine learning for the detection and identification of internet of things (iot) devices: A
survey. arXiv 2021, arXiv:2101.10181.
19.
Karimi, D.; Warfield, S.K.; Gholipour, A. Transfer learning in medical image segmentation: New insights from analysis of the
dynamics of model parameters and learned representations. Artif. Intell. Med. 2021,116, 102078. [CrossRef]
20.
da Silva, G.L.F.; Valente, T.L.A.; Silva, A.C.; de Paiva, A.C.; Gattass, M. Convolutional neural network-based PSO for lung nodule
false positive reduction on CT images. Comput. Methods Progr. Biomed. 2018,162, 109–118. [CrossRef]
21.
Vijh, S.; Sharma, S.; Gaurav, P. Brain tumor segmentation using OTSU embedded adaptive particle swarm optimization method
and convolutional neural network. In Data Visualization and Knowledge Engineering; Springer: Berlin/Heidelberg, Germany, 2020;
pp. 171–194.
22.
Onay, F.K.; Aydemir, S.B. Chaotic hunger games search optimization algorithm for global optimization and engineering problems.
Math. Comput. Simul. 2022,192, 514–536. [CrossRef]
23.
Adel, H.; Dahou, A.; Mabrouk, A.; Abd Elaziz, M.; Kayed, M.; El-Henawy, I.M.; Alshathri, S.; Amin Ali, A. Improving crisis
events detection using distilbert with hunger games search algorithm. Mathematics 2022,10, 447. [CrossRef]
24.
Yang, Y.; Wu, Y.; Yuan, H.; Khishe, M.; Mohammadi, M. Nodes clustering and multi-hop routing protocol optimization using
hybrid chimp optimization and hunger games search algorithms for sustainable energy efficient underwater wireless sensor
networks. Sustain. Comput. Inform. Syst. 2022,35, 100731. [CrossRef]
25.
Devi, R.M.; Premkumar, M.; Jangir, P.; Kumar, B.S.; Alrowaili, D.; Nisar, K.S. BHGSO: Binary hunger games search optimization
algorithm for feature selection problem. CMC-Comput. Mater. Contin. 2022,70, 557–579.
26.
Fahim, S.R.; Hasanien, H.M.; Turky, R.A.; Alkuhayli, A.; Al-Shamma’a, A.A.; Noman, A.M.; Tostado-Véliz, M.; Jurado, F.
Parameter identification of proton exchange membrane fuel cell based on hunger games search algorithm. Energies
2021
,14, 5022.
[CrossRef]
27.
Kaveh, A.; Hamedani, K.B. Improved arithmetic optimization algorithm and its application to discrete structural optimization.
Structures 2022,35, 748–764. [CrossRef]
28.
Khatir, S.; Tiachacht, S.; Le Thanh, C.; Ghandourah, E.; Mirjalili, S.; Wahab, M.A. An improved Artificial Neural Network
using Arithmetic Optimization Algorithm for damage assessment in FGM composite plates. Compos. Struct.
2021
,273, 114287.
[CrossRef]

Diagnostics 2023,13, 834 24 of 26
29.
Wang, R.B.; Wang, W.F.; Xu, L.; Pan, J.S.; Chu, S.C. An adaptive parallel arithmetic optimization algorithm for robot path planning.
J. Adv. Transp. 2021,2021, 3606895. [CrossRef]
30.
Dahou, A.; Al-qaness, M.A.; Abd Elaziz, M.; Helmi, A. Human activity recognition in IoHT applications using arithmetic
optimization algorithm and deep learning. Measurement 2022,199, 111445. [CrossRef]
31.
Khodadadi, N.; Snasel, V.; Mirjalili, S. Dynamic arithmetic optimization algorithm for truss optimization under natural frequency
constraints. IEEE Access 2022,10, 16188–16208. [CrossRef]
32.
Gupta, K.D.; Sharma, D.K.; Ahmed, S.; Gupta, H.; Gupta, D.; Hsu, C.H. A Novel Lightweight Deep Learning-Based Histopatho-
logical Image Classification Model for IoMT. Neural Process. Lett. 2021,2021, 1–24.
33.
Sekhar, A.; Biswas, S.; Hazra, R.; Sunaniya, A.K.; Mukherjee, A.; Yang, L. Brain tumor classification using fine-tuned GoogLeNet
features and machine learning algorithms: IoMT enabled CAD system. IEEE J. Biomed. Health Inform.
2021
,26, 983–991. [CrossRef]
34.
Abeltino, A.; Bianchetti, G.; Serantoni, C.; Ardito, C.F.; Malta, D.; De Spirito, M.; Maulucci, G. Personalized Metabolic Avatar: A
Data Driven Model of Metabolism for Weight Variation Forecasting and Diet Plan Evaluation. Nutrients
2022
,14, 3520. [CrossRef]
35.
Bianchetti, G.; Abeltino, A.; Serantoni, C.; Ardito, F.; Malta, D.; De Spirito, M.; Maulucci, G. Personalized self-monitoring of
energy balance through integration in a web-application of dietary, anthropometric, and physical activity data. J. Pers. Med.
2022
,
12, 568. [CrossRef]
36.
Serantoni, C.; Zimatore, G.; Bianchetti, G.; Abeltino, A.; De Spirito, M.; Maulucci, G. Unsupervised clustering of heartbeat
dynamics allows for real time and personalized improvement in cardiovascular fitness. Sensors 2022,22, 3974. [CrossRef]
37.
Rodrigues, D.D.A.; Ivo, R.F.; Satapathy, S.C.; Wang, S.; Hemanth, J.; Reboucas Filho, P.P. A new approach for classification skin
lesion based on transfer learning, deep learning, and IoT system. Pattern Recognit. Lett. 2020,136, 8–15. [CrossRef]
38.
Han, T.; Nunes, V.X.; Souza, L.F.D.F.; Marques, A.G.; Silva, I.C.L.; Junior, M.A.A.F.; Sun, J.; Reboucas Filho, P.P. Internet of Medical
Things—Based on Deep Learning Techniques for Segmentation of Lung and Stroke Regions in CT Scans. IEEE Access
2020
,
8, 71117–71135. [CrossRef]
39.
Bianchetti, G.; Taralli, S.; Vaccaro, M.; Indovina, L.; Mattoli, M.; Capotosti, A.; Scolozzi, V.; Calcagni, M.L.; Giordano, A.;
De Spirito, M.; et al.
Automated detection and classification of tumor histotypes on dynamic PET imaging data through machine-
learning driven voxel classification. Comput. Biol. Med. 2022,145, 105423. [CrossRef]
40.
Hossen, M.N.; Panneerselvam, V.; Koundal, D.; Ahmed, K.; Bui, F.M.; Ibrahim, S.M. Federated machine learning for detection of
skin diseases and enhancement of internet of medical things (IoMT) security. IEEE J. Biomed. Health Inform.
2022
,27, 835–841.
[CrossRef]
41.
Jain, S.; Nehra, M.; Kumar, R.; Dilbaghi, N.; Hu, T.; Kumar, S.; Kaushik, A.; Li, C.Z. Internet of medical things (IoMT)-integrated
biosensors for point-of-care testing of infectious diseases. Biosens. Bioelectron. 2021,179, 113074. [CrossRef]
42.
Han, X.; Zhang, Z.; Ding, N.; Gu, Y.; Liu, X.; Huo, Y.; Qiu, J.; Yao, Y.; Zhang, A.; Zhang, L.; et al. Pre-trained models: Past, present
and future. AI Open 2021,2, 225–250. [CrossRef]
43.
Cheplygina, V.; de Bruijne, M.; Pluim, J.P. Not-so-supervised: A survey of semi-supervised, multi-instance, and transfer learning
in medical image analysis. Med. Image Anal. 2019,54, 280–296. [CrossRef]
44.
Ayan, E.; Ünver, H.M. Diagnosis of pneumonia from chest X-ray images using deep learning. In Proceedings of the 2019 Scientific
Meeting on Electrical-Electronics & Biomedical Engineering and Computer Science (EBBT), Istanbul, Turkey, 24–26 April 2019;
pp. 1–5.
45.
Ge, Z.; Demyanov, S.; Bozorgtabar, B.; Abedini, M.; Chakravorty, R.; Bowling, A.; Garnavi, R. Exploiting local and generic features
for accurate skin lesions classification using clinical and dermoscopy imaging. In Proceedings of the 2017 IEEE 14th International
Symposium on Biomedical Imaging (ISBI 2017), Melbourne, Australia, 18–21 April 2017; pp. 986–990.
46.
Esteva, A.; Kuprel, B.; Novoa, R.A.; Ko, J.; Swetter, S.M.; Blau, H.M.; Thrun, S. Dermatologist-level classification of skin cancer
with deep neural networks. Nature 2017,542, 115–118. [CrossRef]
47.
Yu, L.; Chen, H.; Dou, Q.; Qin, J.; Heng, P.A. Automated melanoma recognition in dermoscopy images via very deep residual
networks. IEEE Trans. Med. Imaging 2016,36, 994–1004. [CrossRef]
48.
Guo, Y.; Ashour, A.S.; Si, L.; Mandalaywala, D.P. Multiple convolutional neural network for skin dermoscopic image classification.
In Proceedings of the 2018 IEEE International Symposium on Signal Processing and Information Technology (ISSPIT), Louisville,
KY, USA, 6–8 December 2018; pp. 365–369.
49.
Kawahara, J.; BenTaieb, A.; Hamarneh, G. Deep features to classify skin lesions. In Proceedings of the 2016 IEEE 13th International
Symposium on Biomedical Imaging (ISBI), Prague, Czech Republic, 13–16 April 2016; pp. 1397–1400.
50.
Lopez, A.R.; Giro-i Nieto, X.; Burdick, J.; Marques, O. Skin lesion classification from dermoscopic images using deep learning
techniques. In Proceedings of the 2017 13th IASTED International Conference on Biomedical Engineering (BioMed), Innsbruck,
Austria, 20–21 February 2017; pp. 49–54.
51.
Ayan, E.; Ünver, H.M. Data augmentation importance for classification of skin lesions via deep learning. In Proceedings of
the 2018 Electric Electronics, Computer Science, Biomedical Engineerings’ Meeting (EBBT), Istanbul, Turkey, 18–19 April 2018;
pp. 1–4.
52.
Zhang, J.; Xie, Y.; Wu, Q.; Xia, Y. Medical image classification using synergic deep learning. Med. Image Anal.
2019
,54, 10–19.
[CrossRef]

Diagnostics 2023,13, 834 25 of 26
53.
Hosseinzadeh Taher, M.R.; Haghighi, F.; Feng, R.; Gotway, M.B.; Liang, J. A systematic benchmarking analysis of transfer learning
for medical image analysis. In Domain Adaptation and Representation Transfer, and Affordable Healthcare and AI for Resource Diverse
Global Health; Springer: Berlin/Heidelberg, Germany, 2021; pp. 3–13.
54.
Dalmaz, O.; Yurt, M.; Çukur, T. ResViT: Residual vision transformers for multimodal medical image synthesis. IEEE Trans. Med.
Imaging 2022,41, 2598–2614. [CrossRef]
55.
Anwar, S.M.; Majid, M.; Qayyum, A.; Awais, M.; Alnowami, M.; Khan, M.K. Medical image analysis using convolutional neural
networks: A review. J. Med. Syst. 2018,42, 226. [CrossRef]
56.
Samala, R.K.; Chan, H.P.; Hadjiiski, L.M.; Helvie, M.A.; Richter, C.; Cha, K. Evolutionary pruning of transfer learned deep
convolutional neural network for breast cancer diagnosis in digital breast tomosynthesis. Phys. Med. Biol.
2018
,63, 095005.
[CrossRef]
57.
Shankar, K.; Lakshmanaprabu, S.; Khanna, A.; Tanwar, S.; Rodrigues, J.J.; Roy, N.R. Alzheimer detection using Group Grey Wolf
Optimization based features with convolutional classifier. Comput. Electr. Eng. 2019,77, 230–243.
58.
Goel, T.; Murugan, R.; Mirjalili, S.; Chakrabartty, D.K. OptCoNet: An optimized convolutional neural network for an automatic
diagnosis of COVID-19. Appl. Intell. 2021,51, 1351–1366. [CrossRef]
59.
Elhoseny, M.; Shankar, K. Optimal bilateral filter and convolutional neural network based denoising method of medical image
measurements. Measurement 2019,143, 125–135. [CrossRef]
60.
Zhang, N.; Cai, Y.X.; Wang, Y.Y.; Tian, Y.T.; Wang, X.L.; Badami, B. Skin cancer diagnosis based on optimized convolutional neural
network. Artif. Intell. Med. 2020,102, 101756. [CrossRef]
61.
El-Shafeiy, E.; Sallam, K.M.; Chakrabortty, R.K.; Abohany, A.A. A clustering based Swarm Intelligence optimization technique for
the Internet of Medical Things. Expert Syst. Appl. 2021,173, 114648. [CrossRef]
62.
Yu, Z.; Jiang, X.; Zhou, F.; Qin, J.; Ni, D.; Chen, S.; Lei, B.; Wang, T. Melanoma recognition in dermoscopy images via aggregated
deep convolutional features. IEEE Trans. Biomed. Eng. 2018,66, 1006–1016. [CrossRef]
63.
Yu, Z.; Jiang, F.; Zhou, F.; He, X.; Ni, D.; Chen, S.; Wang, T.; Lei, B. Convolutional descriptors aggregation via cross-net for skin
lesion recognition. Appl. Soft Comput. 2020,92, 106281. [CrossRef]
64.
Wei, L.; Ding, K.; Hu, H. Automatic skin cancer detection in dermoscopy images based on ensemble lightweight deep learning
network. IEEE Access 2020,8, 99633–99647. [CrossRef]
65.
Ozkan, I.A.; Koklu, M. Skin lesion classification using machine learning algorithms. Int. J. Intell. Syst. Appl. Eng.
2017
,5, 285–289.
[CrossRef]
66.
Al Nazi, Z.; Abir, T.A. Automatic skin lesion segmentation and melanoma detection: Transfer learning approach with u-net
and dcnn-svm. In Proceedings of the International Joint Conference on Computational Intelligence, Budapest, Hungary,
2–4 November 2020; pp. 371–381.
67.
Afza, F.; Sharif, M.; Mittal, M.; Khan, M.A.; Hemanth, D.J. A hierarchical three-step superpixels and deep learning framework for
skin lesion classification. Methods 2021,202, 88–102. [CrossRef]
68.
Habibzadeh, M.; Krzy˙
zak, A.; Fevens, T. White blood cell differential counts using convolutional neural networks for low
resolution images. In Proceedings of the International Conference on Artificial Intelligence and Soft Computing, Zakopane,
Poland, 9–13 June 2013; pp. 263–274.
69.
Zhao, J.; Zhang, M.; Zhou, Z.; Chu, J.; Cao, F. Automatic detection and classification of leukocytes using convolutional neural
networks. Med. Biol. Eng. Comput. 2017,55, 1287–1301. [CrossRef]
70.
Sharma, M.; Bhave, A.; Janghel, R.R. White blood cell classification using convolutional neural network. In Soft Computing and
Signal Processing; Springer: Berlin/Heidelberg, Germany, 2019; pp. 135–143.
71.
Mabrouk, A.; Díaz Redondo, R.P.; Dahou, A.; Abd Elaziz, M.; Kayed, M. Pneumonia Detection on Chest X-ray Images Using
Ensemble of Deep Convolutional Neural Networks. Appl. Sci. 2022,12, 6448. [CrossRef]
72.
Ignatov, A.; Romero, A.; Kim, H.; Timofte, R. Real-time video super-resolution on smartphones with deep learning, mobile ai
2021 challenge: Report. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, Nashville,
TN, USA, 20–25 June 2021; pp. 2535–2544.
73.
Tan, M.; Le, Q. Efficientnet: Rethinking model scaling for convolutional neural networks. In Proceedings of the International
Conference on Machine Learning, PMLR, Long Beach, CA, USA, 10–15 June 2019; pp. 6105–6114.
74.
Howard, A.; Sandler, M.; Chu, G.; Chen, L.C.; Chen, B.; Tan, M.; Wang, W.; Zhu, Y.; Pang, R.; Vasudevan, V.; et al. Searching
for mobilenetv3. In Proceedings of the IEEE/CVF International Conference on Computer Vision, Seoul, Republic of Korea,
27 October–2 November 2019; pp. 1314–1324.
75.
He, K.; Zhang, X.; Ren, S.; Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE conference on
computer vision and pattern recognition, Honolulu, HI, USA, 1–26 July 2016; pp. 770–778.
76.
Abualigah, L.; Diabat, A.; Mirjalili, S.; Abd Elaziz, M.; Gandomi, A.H. The arithmetic optimization algorithm. Comput. Methods
Appl. Mech. Eng. 2021,376, 113609. [CrossRef]
77.
Yang, Y.; Chen, H.; Heidari, A.A.; Gandomi, A.H. Hunger games search: Visions, conception, implementation, deep analysis,
perspectives, and towards performance shifts. Expert Syst. Appl. 2021,177, 114864. [CrossRef]
78.
Giordani, M.; Polese, M.; Mezzavilla, M.; Rangan, S.; Zorzi, M. Toward 6G networks: Use cases and technologies. IEEE Commun.
Mag. 2020,58, 55–61. [CrossRef]

Diagnostics 2023,13, 834 26 of 26
79.
Mendonça, T.; Ferreira, P.M.; Marques, J.S.; Marcal, A.R.; Rozeira, J. PH 2-A dermoscopic image database for research and
benchmarking. In Proceedings of the 2013 35th Annual International Conference of the IEEE Engineering in Medicine and
Biology Society (EMBC), Osaka, Japan, 3–7 July 2013; pp. 5437–5440.
80.
Gutman, D.; Codella, N.C.; Celebi, E.; Helba, B.; Marchetti, M.; Mishra, N.; Halpern, A. Skin lesion analysis toward melanoma
detection: A challenge at the international symposium on biomedical imaging (ISBI) 2016, hosted by the international skin
imaging collaboration (ISIC). arXiv 2016, arXiv:1605.01397.
81.
Liang, G.; Hong, H.; Xie, W.; Zheng, L. Combining convolutional neural network with recursive neural network for blood cell
image classification. IEEE Access 2018,6, 36188–36197. [CrossRef]
82.
Abualigah, L.; Yousri, D.; Abd Elaziz, M.; Ewees, A.A.; Al-qaness, M.A.; Gandomi, A.H. Aquila Optimizer: A novel meta-heuristic
optimization Algorithm. Comput. Ind. Eng. 2021,157, 107250. [CrossRef]
83.
Mirjalili, S. Moth-flame optimization algorithm: A novel nature-inspired heuristic paradigm. Knowl.-Based Syst.
2015
,89, 228–249.
[CrossRef]
84.
Yang, X.S. A new metaheuristic bat-inspired algorithm. In Nature inspired Cooperative Strategies for Optimization (NICSO 2010);
Springer: Berlin/Heidelberg, Germany, 2010; pp. 65–74.
85.
Hashim, F.A.; Hussain, K.; Houssein, E.H.; Mabrouk, M.S.; Al-Atabany, W. Archimedes optimization algorithm: A new
metaheuristic algorithm for solving optimization problems. Appl. Intell. 2021,51, 1531–1551. [CrossRef]
86.
Talatahari, S.; Azizi, M. Chaos Game Optimization: A novel metaheuristic algorithm. Artif. Intell. Rev.
2021
,54, 917–1004.
[CrossRef]
87.
Kermany, D.S.; Goldbaum, M.; Cai, W.; Valentim, C.C.; Liang, H.; Baxter, S.L.; McKeown, A.; Yang, G.; Wu, X.; Yan, F.; et al.
Identifying medical diagnoses and treatable diseases by image-based deep learning. Cell 2018,172, 1122–1131. [CrossRef]
88.
Fang, L.; Jin, Y.; Huang, L.; Guo, S.; Zhao, G.; Chen, X. Iterative fusion convolutional neural networks for classification of optical
coherence tomography images. J. Vis. Commun. Image Represent. 2019,59, 327–333. [CrossRef]
89.
Huang, L.; He, X.; Fang, L.; Rabbani, H.; Chen, X. Automatic classification of retinal optical coherence tomography images with
layer guided convolutional neural network. IEEE Signal Process. Lett. 2019,26, 1026–1030. [CrossRef]
90.
Sun, Y.; Li, S.; Sun, Z. Fully automated macular pathology detection in retina optical coherence tomography images using sparse
coding and dictionary learning. J. Biomed. Opt. 2017,22, 016012. [CrossRef]
91.
Ji, Q.; He, W.; Huang, J.; Sun, Y. Efficient deep learning-based automated pathology identification in retinal optical coherence
tomography images. Algorithms 2018,11, 88. [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.