See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/378936665

Revolutionizing healthcare image analysis in pandemic-based fog-cloud

computing architectures

Article in Indonesian Journal of Electrical Engineering and Computer Science · April 2024
DOI: 10.11591/ijeecs.v34.i1.pp441-454

CITATIONS READS

0 15

3 authors, including:

Khalil Mohamed Hany M. Harb

Al-Azhar University Al-Azhar University
18 PUBLICATIONS 53 CITATIONS 94 PUBLICATIONS 749 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Khalil Mohamed on 13 March 2024.

The user has requested enhancement of the downloaded file.

Indonesian Journal of Electrical Engineering and Computer Science
Vol. 34, No. 1, April 2024, pp. 441~454
ISSN: 2502-4752, DOI: 10.11591/ijeecs.v34.i1.pp441-454  441

Revolutionizing healthcare image analysis in pandemic-based

fog-cloud computing architectures

Alzahraa Elsayed, Khalil Mohamed, Hany Harb

Department of Systems and Computers Engineering, Faculty of Engineering, Al-Azhar University, Nasr, Egypt

Article Info ABSTRACT

Article history: Healthcare data analysis has become essential after epidemic outbreaks. The
manual examination of medical images such as X-rays and computed
Received Jul 4, 2023 tomography (CT) scans became one of these challenges. This paper
Revised Dec 29, 2023 introduces a healthcare architecture that tackles the analysis efficiency and
Accepted Jan 3, 2024 accuracy challenges by harnessing artificial intelligence (AI) capabilities.
This architecture utilizes fog computing and presents a modified
convolutional neural network (CNN) designed specifically for image
Keywords: analysis. Different architectures of CNN layers are thoroughly explored and
evaluated to optimize overall performance. To demonstrate the effectiveness
Convolution neural network of the proposed approach, a dataset of X-ray images is utilized for analysis
COVID-19 and evaluation. Comparative assessments are conducted against recent
Deep learning models such as VGG16, VGG19, MobileNet, and related research papers.
Fog computing Notably, the proposed approach achieves an exceptional accuracy rate of
Healthcare system 99.88% in classifying normal cases, accompanied by a validation rate of
96.5%, precision and recall rates of 100%, and an F1 score of 100%. These
results highlight the immense potential of fog computing and modified
CNNs in revolutionizing healthcare image analysis and diagnosis, not only
during pandemics but also in the future. By leveraging these technologies,
healthcare professionals can improve the efficacy and accuracy of medical
image analysis, leading to improved patient care and outcomes.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Alzahraa Elsayed
Department of Systems and Computers Engineering, Faculty of Engineering
Al-Azhar University
11765, Nasr City, Cairo, Egypt
Email: [email protected]

1. INTRODUCTION
The field of healthcare data analysis has undergone a significant transformation, particularly in
response to pandemics, which have heightened the demand for efficient solutions. During such critical times,
healthcare professionals and researchers face a significant challenge in manually analyzing medical images,
including X-rays and computed tomography (CT) scans [1]. This task is time-consuming and complicated by
the logistical hurdles of transferring these large image datasets to centralized cloud computing servers.
Moreover, the speed and accuracy of image analysis are crucial factors in effective healthcare image
management.
Cloud computing, a technology which permits users to access computing resources which include
data storage and processing power via the internet, has the likelihood to enhance the safety, quality, and
efficiency of healthcare. One application of cloud computing in healthcare involves storing and analyzing
extensive patient data to identify trends and patterns which could aid in more efficient disease diagnosis and
management [1]. Additionally, cloud computing can facilitate the remote delivery of healthcare services,

Journal homepage: http://ijeecs.iaescore.com

442  ISSN: 2502-4752

particularly benefiting patients residing in rural areas or those facing challenges in visiting a doctor's office [2].
However, the traditional cloud computing architecture for medical and healthcare purposes relies on a
centralized approach for data transmission, which poses several challenges [3]:
i) Security: centralized data storage increases vulnerability to cyberattacks.
ii) Scalability: scaling centralized data storage to meet the needs of a growing number of users can be
challenging.
iii) Compliance: healthcare organizations must adhere to strict regulations governing the privacy and security
of patient data, which can be difficult with a centralized storage system.
iv) Latency: the time it takes for data to travel from the network's edge to the cloud and back may be too long
for time-sensitive healthcare applications, particularly those requiring rapid response in emergency
situations.
v) Cost: transferring huge amounts of data to the cloud can be expensive.
These challenges have led some healthcare organizations to adopt fog computing as an alternative to
cloud computing. Fog computing, a distributed cloud computing architecture, is better suited for healthcare
applications. Fog computing nodes are positioned closer to the network's edge, reducing latency and
improving security. Moreover, fog computing could be a more cost-effective solution for health care
applications that involve significant data transfer. Fog computing holds great promise for healthcare and is
expected to witness increased adoption in the future [4].
Deep learning (DL), a subset of artificial intelligence (AI), is being leveraged to enhance healthcare
in various ways. DL algorithms can analyse medical images and data to diagnose diseases, develop new
treatments, and deliver personalized care [5], [6]. It is also instrumental in developing healthcare applications
which include virtual assistants as well as chatbots. DL is a rapidly evolving field that is poised to have a
significant impact on healthcare in the years to come [7], [8].
The advancement of science and technology has historically been driven by medicine and healthcare
[9], [10]. Recent research has focused on integrating fog computing into internet of health technology (IoHT)
applications, yielding positive outcomes such as reduced service response time, improved system
performance, and increased energy efficacy. For instance, Xue et al. [11] developed the analytic network
technique to identify and rank fog computing-based internet of things (IoT) solutions for health system
supervising. Fog computing in health care includes establishing a distributed intermediate layer between the
cloud and sensor hubs using IoT frameworks.
Gia et al. [12] demonstrated the use of fog computing as a gateway to enhance health monitoring
systems. They created fog computing features, including interoperability, a distributed database, a real-time
notification mechanism, position awareness, and a graphical user interface with access management.
Additionally, they presented a lightweight, customizable framework for extracting electrocardiography
(ECG) features (such as pulse, P, and T waves). Elhadad et al. [13] have suggested a fog-based health
monitoring framework which utilizes fog gateways in the context of medical decision-making according to
data collected from sensors embedded in wearable devices. These sensors which include temperature sensors,
ECG sensors, and blood pressure (BP) sensors, measure a patient's temperature, pulse, and BP respectively.
Al-Khafajiy et al. [14] introduced the concept of IoT-fog computing in IoT-based healthcare
systems, suggesting a methodology for improving fog performance through collaborative policies among fog
nodes for optimal workload and job distribution. Similarly, El-Rashidy et al. [15] presented a detailed
strategy to monitor pregnant females by utilizing a data replacement and prediction framework (DRPF)
divided into three layers: (i) IoT, (ii) fog, and (iii) cloud. Their findings indicated strong associations between
patient age, body mass index (BMI), BP, lymphocyte vitamin E levels, and the diagnosis of gestational
diabetes.
Quy et al. [16] presented an all-in-one computer architectural framework and conducted a survey of
IoT applications according to fog computing in the health care industry. They explored the application
potential, challenges, and future research objectives in this field. Similarly, Shi et al. [17] evaluated the
vision and essential characteristics of fog computing, which aims to address the latency issue caused by IoT
by distributing processing, storage, and networking resources to the network edge, interacting with the cloud.
Arunkumar et al. [18] recommended HealthFog-CCNN, a fog-based smart healthcare system to
automatically diagnose cardiac disorders that combines DL and IoT. Their research focused on the medical
aspects of heart disease patients, utilizing DL in edge computing devices for real-time analysis of cardiac
problems. Lastly, Mutlag et al. [19] aimed to contribute to the existing knowledge by providing specific
examples categorized into four groups: fog computing approaches in healthcare applications, system
development in fog computing for healthcare applications, and evaluation and surveys of fog computing in
healthcare applications.
While fog computing and artificial intelligence (AI) have been effectively utilized in the healthcare
field, no effective framework has been used for heavy healthcare processing, particularly during a pandemic.

Indonesian J Elec Eng & Comp Sci, Vol. 34, No. 1, April 2024: 441-454
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752  443

Furthermore, most papers examine small datasets for efficiency testing. Therefore, the current paper suggests
a new framework according to a distributed data ingestion/collection layer, with distributed data processing
and integration layers to simplify the processing of similar data on specific devices. The collected data is then
forwarded to fog nodes for additional analysis. The next section offers a detailed description of the proposed
framework.
To address these formidable challenges, this research paper introduces an innovative healthcare
architecture that aims to achieve both analysis efficiency and accuracy. At its core, this architecture harnesses
the power of AI. Specifically, it presents a novel framework that combines the fog computing paradigm with
a meticulously designed modification of convolutional neural networks (CNNs), tailored explicitly for image
analysis. A comprehensive exploration of various CNN layer architectures is conducted and subjected to
rigorous evaluation to optimize performance.
To empirically validate the efficiency of the recommended approach, a curated dataset of COVID-
19-related X-ray images is used for analysis and evaluation. These images serve as a practical testbed to
enable comparative assessments against contemporary models such as VGG16, VGG19, and MobileNet. The
outcomes of these assessments unequivocally demonstrate the exceptional potential of the proposed
framework.
These compelling results underscore the transformative potential that emerges from the intersection
of fog computing and modified CNNs in the domain of healthcare image analysis. In a world grappling with
the challenges posed by pandemics, the convergence of cutting-edge technology and medical science holds
the promise of revolutionizing healthcare image analysis and diagnosis, not only during times of crisis but
also in the broader landscape of healthcare delivery. This research represents a critical step forward in
realizing this promise and provides a glimpse into a future where the fusion of AI and healthcare unlocks
boundless potential.
The remainder of the current paper is planned as follows: section two discusses previous work in the
field. Section three defines the proposed framework. Section four presents the CNNs architecture. Section 5
introduces mathematical formulas for the proposed model. Section 6 discusses database for chest X-rays
(CXR). Section 7 discusses the experimental results. Section 8 evaluation of three layers of the CNN model
with different epochs. Finally, section 9 concludes the paper by summarizing the outcomes and outlining
upcoming research directions.

2. METHOD
The integration of fog computing, cloud resources and AI technologies in the proposed healthcare
framework creates a comprehensive and efficient healthcare ecosystem [9], [10] as depicted in Figure 1. The
framework aims to improve patient care, streamline healthcare processes, and support research and
innovation. It consists of several interconnected layers, each with its unique roles and functionalities.
The first layer is the end layer, which has an essential role in collecting and assimilating diverse data
from sources which include medical devices, sensors, electronic health records, and patient-generated data.
Its main functions include data collection, transformation, quality assurance, and aggregation. This layer
ensures data compatibility, reliability, and security through encryption and authentication. It also prioritizes
scalability, reliability, adherence to healthcare regulations, and interoperability for efficient data sharing and
collaboration [11], [12].
The second layer is the edge/fog computing layer, that serves as a critical bridge between the data
sources and the centralized cloud infrastructure. Positioned closer to the data sources, this layer allows real-
time or near-real-time data processing, in particular essential for low-latency healthcare applications like
patient monitoring and emergency care. It leverages distributed edge or fog computing nodes to execute data
analytics and computational tasks at the source, reducing the burden on central cloud resources and
optimizing bandwidth usage. This layer ensures local data processing, secure data storage, and efficient
resource allocation to enhance system scalability, reliability, and rapid decision-making [13]−[16]. The
proposed model is in this layer in order to analyse images to detect disease.
The third layer is the cloud layer, which serves as the central component of the healthcare system,
providing hardware resources and high-capacity computer services as data centres for data computation and
storage. It encompasses data analysis and pre-processing procedures, supporting medical professionals in
making long-term treatment decisions. This layer involves various processing tasks, including normalization
and data preparation, before training machine learning algorithms such as CNN for disease diagnosis,
predictive modeling, anomaly detection, and data-driven decision support. It contributes to improved
healthcare outcomes and drives innovation within the healthcare ecosystem [18], [19]. This layer contains the
pre-trained CNN model to ensure the validity of the results.

Revolutionizing healthcare image analysis in pandemic-based fog-cloud … (Alzahraa Elsayed)

444  ISSN: 2502-4752

In summary, the proposed healthcare framework integrates fog computing, cloud resources, and AI
technologies to create a holistic and efficient healthcare ecosystem. The framework consists of the data
ingestion layer, data processing and integration layer, edge/fog computing layer, and cloud layer, each
playing an essential role in collecting, processing, analyzing, and storing healthcare data to support improved
patient care and decision-making.

Figure 1. The framework based on IoT fog-cloud computing architecture

2.1. The CNN architecture

Like other domains, the healthcare field has successfully implemented several DL applications that
have revealed significant results in various medical scenarios. This success can be attributed to two main
factors: i) the ability of DL models to learn from labeled or unlabeled datasets and ii) the inherent risk of
human error in diagnosing cases, regardless of the doctors' expertise level. Consequently, the medical
research community has developed numerous healthcare systems based on DL methods.
One specific DL technique, called CNN, has been designed to excel in image identification,
classification, and prediction tasks. CNN's ability to automatically extract features from images and perform
in-depth analysis makes it ideally suited for training in the recommended architecture. In this study, a CNN-
based model is proposed for detecting chest diseases in patients using chest radiography images. The model's
objective is to classify and identify chest diseases by distinguishing between normal X-rays and abnormal
CXR, as depicted in Figure 2.
For CNN to operate effectively, the input images must undergo processing to extract visual patterns.
This process involves a linear operation where two functions represented by matrices are multiplied to
produce an output. Initially, the images are transformed into a matrix format to facilitate information

Indonesian J Elec Eng & Comp Sci, Vol. 34, No. 1, April 2024: 441-454
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752  445

extraction. CNN comprises several layers that work together to perform these operations efficiently,
including as shown Figure 3:
i) Input data layer: this layer reads a pre-processed collection of images. In our case, the X-ray and CT
images are pre-processed separately.
ii) Convolutional layer: serving as the core of our proposed model, this layer is responsible for extracting
features from the image collection while preserving the spatial relation between pixels.
iii) Batch normalization layer: this layer is a crucial training strategy in deep neural networks as it ensures the
stability and proper training of convolutional features.
iv) Rectified linear unit (ReLU) layer: This layer replaces negative pixel values in the convolved features
with zero, generating a non-linearity map of the CNN network's features.
v) Fully connected layer: this layer categorizes the convolved features from the image datasets into the
desired classes.
vi) Softmax layer: interpreting the probability values of the activation function from the previous layer, this
layer is particularly relevant for illness diagnosis. The results can be interpreted as two classes: '0' for
negative (normal CXR or CT) and '1' for positive disease.
vii)Output layer: this final layer of the CNN model labels the results obtained from the previous layer
accordingly [20].

Figure 2. The proposed model is according to CNN architecture

Figure 3. The CNN architecture layers

In conclusion, the healthcare field has embraced DL applications, leveraging the remarkable
capabilities of DL models in various medical scenarios. The proposed architecture incorporates CNN, a
specialized DL technique for image identification, classification, and prediction, to detect chest diseases from

Revolutionizing healthcare image analysis in pandemic-based fog-cloud … (Alzahraa Elsayed)

446  ISSN: 2502-4752

chest radiography images. The CNN model consists of multiple layers, each playing a distinct role in
efficiently processing and analyzing the input data to classify and identify chest diseases [20].

3. RESULTS AND DISCUSSION

The experiments were conducted on an HP EliteBook operating on Windows 10 Pro 64-bit. The
system specifications included a 3.8 GHz Core i7 processor, 16 GB of RAM, an Intel HD Graphics 4600
GPU, and 1 TB of storage. The experiments were implemented using Python 3.9 with the Keras and
TensorFlow libraries within the PyCharm program.

3.1. Database for chest X-rays

A collaborative team of authors from Qatar University, the University of Dhaka in Bangladesh,
besides their partners from Pakistan and Malaysia, collaborated with medical professionals to develop a
comprehensive database of CXR image. This database includes cases of COVID-19-positive individuals, as
well as normal and viral pneumonitis cases. The dataset is being made available in stages, with the initial
release comprising 219 COVID-19, 1,341 normal, and 1,345 viral pneumonia CXR images. Subsequently,
the COVID-19 class was expanded to include 1,200 CXR images in the first update. In the second update, the
database was further expanded to encompass 1,345 viral pneumonia cases, 3,616 COVID-19 positive cases,
10,192 normal cases, 6,012 instances of lung opacity (non-COVID lung infections), and an additional 6,012
COVID-19 positive cases. For our research, we are specifically interested in two datasets: one containing
10,192 normal CXR and another containing 3616 COVID-19 positive CXR [21].
The dataset curated by Chowdhury et al. [22] consists of 13,808 chest images captured in the
posterior/anterior (PA) or anterior-posterior (AP) view. Each sample in the database has a resolution of
1,024×1,024 pixels and is stored in portable network graphics (PNG) format. To facilitate compatibility with
popular CNNs, the images were resized to standard dimensions of 224×224 or 227×227.
In this study, the dataset of 13,808 images was divided into training, validation, and test sets with a
ratio of 80:10:10. The corresponding counts for each dataset are as follows:
i) Training dataset: this dataset comprises a total of 13,808 images, with 2,894 being COVID-19 positive
CXR and 8,154 being normal CXR.
ii) Validation dataset: the validation dataset consists of 1,381 images, including 361 COVID-19 positive
CXR and 1,019 normal CXR.
iii) Testing dataset: the testing dataset contains 1,381 images, with 361 COVID-19 positive CXR and 1,019
normal CXR.
Table 1 provides an overview of the distribution of images across the training, validation, and test sets. In the
dataset, normal cases equal (0), and COVID-19 positive cases equal (1), as shown in Figure 4.

Table 1. Distribution of chest X-ray images after splitting

Classes Training set Validation set Testing set Total
COVID-19 2,894 361 361 3,616
NORMAL 8,154 1,019 1,019 10,192

Figure 4. The dataset of normal cases and COVID-19-positive cases

Indonesian J Elec Eng & Comp Sci, Vol. 34, No. 1, April 2024: 441-454
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752  447

3.2. Performance measures

In this paper, to evaluate our model, the cross-entropy log loss function between correct and
predicted labels is used as an evaluation criterion. The mathematical formula for calculating log loss is
written as (1):
1
Log Loss = − ∑𝐾 𝑆
𝑖=1 ∑𝑗=1 𝑔𝑖𝑗 𝑙𝑜𝑔 (𝐹𝑖𝑗 ) (1)
𝐾

here, K denotes the number of samples, and S denotes the number of classes. The true label of the class is
represented by g, and the probability of the given sample is represented by F. The natural logarithm is used in
the formula.
In terms of the current model assessment, we used the next classification metrics: true positive (TP),
true negative (TN), false positive (FP), and false negative (FN). We calculate recall, TP rate (TPR), false
positive rate (FPR), precision, specificity, sensitivity, F1 score, and accuracy using these measures.
Recall: it represents a model's ability to find all relevant cases within a dataset. Mathematically, recall is
defined as the number of TP divided by the total number of TP plus the number of FN.
TP
Recall = TPR = (2)
TP+FN

Precision: it indicates a classification model's ability to identify only relevant data points. Precision is defined
as the number of TP divided by the number of TP plus the number of false positives.
TP
Precision = (3)
TP+FP

Specificity refers to the number of correctly predicted negative records. It helps determine how well our
model predicts the class that we want to label as the negative class. In some ways, it is like Recall for the
negative class.
TN
Specificity = (4)
TN+FP

Sensitivity refers to the number of positive records correctly predicted. For the class that we want to declare
as the positive class, sensitivity is the same as recall.
TP
Sensitivity = (5)
TP+FN

Additionally, accuracy is another evaluation metric used to assess the performance of our model. It is
mathematically defined as (6).
Tp+Tn
Accuracy = (6)
Tp+TN+FP+FN

The F1-score is the harmonic mean of precision and recall, using the following equation to account for both
metrics as (7).

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛∗𝑅𝑒𝑐𝑎𝑙𝑙
𝐹1 − 𝑠𝑐𝑜𝑟𝑒 = 2 ∗ ( (7)
𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 + 𝑅𝑒𝑐𝑎𝑙𝑙

The results of our proposed framework showcased exceptional performance in classifying COVID-
19 and normal cases, accomplishing an accuracy of 99.88% and a validation rate of 96.50%. A summary of
the classification results is presented in Table 2. According to Table 2, our proposed approach achieves the
highest accuracy of 99.88% in classifying COVID-19 and normal cases. It is accompanied by a validation
rate of 96.5%, precision and recall rates of 100%, and an F1-score of 100% based on the training rate.
However, for the validation rate, the precision and recall rates are 98%, and the F1-score is 98.45% with a
precision of 98.88% and recall rate of 98%.
Figure 5 provides a visual representation comparing the pooling of a 3-by-3 image with a stride of 2
and the pooling of a 2-by-2 image with a stride of 3. The illustration concludes that while reducing the image
features, minimal information loss occurs. Before the completely linked CNN layer, feature maps from the
three sequential layers are concatenated. Weights are calculated using the Glorot technique [23], the Adam

Revolutionizing healthcare image analysis in pandemic-based fog-cloud … (Alzahraa Elsayed)

448  ISSN: 2502-4752

optimizer [24], and a learning rate of 0.001, 400 epochs, and 32 mini-batch sizes. Table 3 depicts the
proposed model based on CNN's structure.

Table 2. Classification results

Metrics Precision Recall F1-score Accuracy
Training 100% 100% 100% 99.88%
validation 98.88% 98% 98.45% 96.50%

Figure 5. Effect of stride and pooling on image resolution

Table 3. Proposed model based on CNN

Layer (type) Output shape Parameter
Conv2d (Conv2D) (None, 200, 200, 64) 6,976
Max_pooling2d (Maxpooling2D) (None, 100, 100, 64) 0
Dropout (Dropout) (None, 100, 100, 64) 0
Conv2d_1(Conv2D) (None, 100, 100, 128) 295,040
Max_pooling2d_1(Maxpooling2D) (None, 50, 50, 128) 0
Dropout_1 (Dropout) (None, 50, 50, 128) 0
Conv2d_2(Conv2D) (None, 50, 50, 256) 1,179,904
Max_pooling2d_2(Maxpooling2D) (None, 25, 25, 256) 0
Dropout_2 (Dropout) (None, 25, 25, 256) 0
Flatten (Flatten) (None, 160,000) 0
Dense (Dense) (None, 512) 81,920,512
Dropout_3 (Dropout) (None, 512) 0
Dense_1 (Dense) (None, 1) 513

Indonesian J Elec Eng & Comp Sci, Vol. 34, No. 1, April 2024: 441-454
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752  449

3.3. Evaluation of three layers of the CNN model with 400 epochs
To validate the output of the presented approach, we implemented and tested the approach using
(400) epochs with three layers of the CNN model. A series of experiments were conducted to assess the
performance of the proposed model. The results encompass different performance metrics, comprising
accuracy, precision, recall, and F1-score.

3.3.1. Using 400 epochs

Deep-COVID-19, our proposed model according to CNN, was trained for 400 epochs with early
stopping. It was tested on the remaining 20% of the dataset after all models were trained on 80% of it. Figure 6
demonstrates the relation between the number of epochs and the loss value. At the first epoch, the loss values
for the training and validation sets are 0.3852 and 0.4419 for the three-layer model. The loss values for the
three-layer model decrease dramatically to 0 0.3229 at the fifth epoch. The training loss for the three-layer
model gradually decreases to 0.0011 at 361 epochs. The validation losses for the three-layer model gradually
decrease to 0.1418 at 75 epochs.
Figure 7 shows the relation between the number of epochs and the accuracy value. At the first
epoch, the accuracy values for the training and validation sets are 0.845 and 0.785 for the three-layer model.
The accuracy values for three-layer model increase dramatically to 0.9013 at the fifth epoch. The training the
training accuracy for the three-layer model gradually increases to 0.9988 at 373 epochs. The validation
accuracy for the three-layer model gradually increases to 0.965 at 159 epochs.

Figure 6. model loss of our Proposed model based on CNN using three layers

Figure 7. model accuracy of our Proposed model according to CNN using three layers

Revolutionizing healthcare image analysis in pandemic-based fog-cloud … (Alzahraa Elsayed)

450  ISSN: 2502-4752

Figure 8 shows the relation between the number of epochs and the precision value. At the first
epoch, the precision values for the training and validation sets are 0.8485 and 0.728 for the three-layer model.
The precision values for the three-layer model increase dramatically to 0.9303 at the fifth epoch. The training
precision for the three-layer model gradually increases to 1 at 188 epochs. The validation precision for the
three-layer model gradually increases to 0.9888 at 324 epochs.

Figure 8. Model precision of our proposed model according to CNN using three layers

Figure 9 shows the relation between the number of epochs and the recall value. At the first epoch,
the recall values for the training and validation sets are 0.84 and 0.91 for the three-layer model. The recall
values for the three-layer model increase dramatically to 0.8675 at the fifth epoch. The training recall for the
three-layer model gradually increases to 0.98 at 360 epochs.

Figure 9. model recall of our proposed model according to CNN using three layers

Figure 10 shows the relation between the number of epochs and the F1 score value for the three-
layer CNN model. At the first epoch, the F1 scores for the training and validation sets are 3.36 and 3.64 for
the three-layer model. The F1 scores for the three-layer model increase dramatically to 3.47 at the fifth
epoch. The training F1 score for the three-layer model gradually increases to 4 at 204 epochs.

Indonesian J Elec Eng & Comp Sci, Vol. 34, No. 1, April 2024: 441-454
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752  451

The validation F1 score for the three-layer model gradually increases to 3.92 at 360 epochs. By
comparing the results of our proposed model based on CNN with other models in the literature, it was clear
that our model has higher accuracy in classifying COVID-19 and normal cases. Our model achieved an
accuracy of 99.88%, a validation rate of 96.5%, a precision of 100%, a recall of 100%, and an F1 score of
100%.

Figure 10. Model F1-score of our proposed model according to CNN using three layers

3.4. Comparison of deep learning models for the diagnosis of COVID-19

In this section we compared our proposed model according to CNN with three pre-trained models:
VGG16, VGG19, and MobileNet. We used the same dataset for all models. Transfer learning has been
considered as a popular practice in computer vision. A pre-trained model is one that has been trained on a
major benchmark dataset for problem solving comparable to the one we want to solve. Because of the
computational charge of training these models, it is frequent practice to import and utilize models from
documented literatures (e.g., VGG16, VGG19, MobileNet).
Transfer learning is a powerful technique in computer vision as it permits us to build precise models
in a timely manner. With transfer learning, we do not start from scratch; instead, we start from patterns that
have been established when solving dissimilar problems. This allows us to build on existing knowledge
rather than starting from scratch [25], [26].
Figure 11 shows the relation between the number of epochs and the accuracy value for the three pre-
trained models. At the first epoch, the training and validation accuracy for all three models are low. However,
the accuracy for all models gradually increases over time. Figure 11(a) to 11(c) shows that for 400 epochs,
the training accuracy for VGG16, VGG19, and MobileNet are 0.925, 0.946, and 0.957, respectively. The
validation accuracy for VGG16, VGG19, and MobileNet are 0.895, 0.927, and 0.935, respectively.
Therefore, the proposed model outperforms the three pre-trained models. Our model achieved a
higher accuracy at 400 epochs than the three pre-trained models. Due to the limited scope of the current
study, we have summarized the results of the comparison in the Table 4.

Table 4. Classification results

Models Dataset Precision Recall F1-score Accuracy
VGG16 Train 71% 70% 70.5% 92.5%
Valid 68.9% 67% 67.9% 89.5%
VGG19 Train 82% 82% 82% 94.6%
Valid 79.3% 78% 78.6% 92.7%
MobileNet Train 93% 93% 93% 95.7%
Valid 89.6% 87% 88.3% 93.5%
Proposed model based on CNN Train 100% 100% 100% 99.88%
Valid 98.88% 98% 98.45% 96.50%

Revolutionizing healthcare image analysis in pandemic-based fog-cloud … (Alzahraa Elsayed)

452  ISSN: 2502-4752

(a)

(b)

(c)

Figure 11. Model accuracy of three pre-trained models: (a) VGG16, (b) VGG19 and (c) MobileNet

4. CONCLUSION
This paper presents a framework according to an IoT fog-cloud computing architecture for
identifying COVID-19. We also propose a model based on CNN that is deployed and implemented on a fog
computing layer to detect COVID-19 from CXR images. We evaluate the performance of the proposed
model by studying its categorization accuracy. The proposed model was experimented by utilizing three

Indonesian J Elec Eng & Comp Sci, Vol. 34, No. 1, April 2024: 441-454
Indonesian J Elec Eng & Comp Sci ISSN: 2502-4752  453

layers of CNN, and the results demonstrated that the training and validation accuracy gradually increased to
99.87% and 95.50%, correspondingly.
The proposed model was also experimented with three layers of CNN. The results revealed that the
training and validation accuracy increased to 99.88% and 96.50%, correspondingly. The proposed model was
then compared with other studies and with three pre-trained models: VGG16, VGG19, and MobileNet. The
results showed that the accuracy of the proposed model was higher than the other models. The proposed
model attained an accuracy of 99.88% in classifying COVID-19 and normal cases, along with a validation
rate of 96.5%, precision of 100%, recall of 100%, and F1 score of 100%. In the future, we plan to: i) secure
user multimedia data in the cloud using fog computing; ii) focus on authentication and key agreement using
different authentication algorithms; iii) use ECG to detect COVID-19, as recent forms of COVID-19 can
affect the cardiovascular system; and iv) investigate the use of empirical wavelet transform (EWT) and
principal component analysis (PCA) for data filtering.

REFERENCES
[1] A. M.-H. Kuo, “Opportunities and challenges of cloud computing to improve health care services,” Journal of Medical Internet
Research, vol. 13, no. 3, p. e67, Sep. 2011, doi: 10.2196/jmir.1867.
[2] T. Alam, “Cloud computing and its role in the information technology,” IAIC Transactions on Sustainable Digital Innovation
(ITSDI), vol. 1, no. 2, pp. 108–115, Feb. 2020, doi: 10.34306/itsdi.v1i2.103.
[3] S. Dash, S. K. Shakyawar, M. Sharma, and S. Kaushik, “Big data in healthcare: management, analysis and future prospects,”
Journal of Big Data, vol. 6, no. 1, p. 54, Dec. 2019, doi: 10.1186/s40537-019-0217-0.
[4] F. Emmert-Streib, Z. Yang, H. Feng, S. Tripathi, and M. Dehmer, “An introductory review of deep learning for prediction models
with big data,” Frontiers in Artificial Intelligence, vol. 3, p. 4, Feb. 2020, doi: 10.3389/frai.2020.00004.
[5] Q. Althebyan, Q. Yaseen, Y. Jararweh, and M. Al-Ayyoub, “Cloud support for large scale e-healthcare systems,” Annals of
Telecommunications, vol. 71, no. 9–10, pp. 503–515, Oct. 2016, doi: 10.1007/s12243-016-0496-9.
[6] K. Bhargava and S. Ivanov, “A fog computing approach for localization in WSN,” in 2017 IEEE 28th Annual International
Symposium on Personal, Indoor, and Mobile Radio Communications (PIMRC), Oct. 2017, pp. 1–7, doi:
10.1109/PIMRC.2017.8292245.
[7] A. Canziani, A. Paszke, and E. Culurciello, “An analysis of deep neural network models for practical applications,” arXiv
preprint, 2016, [Online]. Available: http://arxiv.org/abs/1605.07678.
[8] Y. LeCun, Y. Bengio, and G. Hinton, “Deep learning,” Nature, vol. 521, no. 7553, pp. 436–444, May 2015, doi:
10.1038/nature14539.
[9] V. Chamola, V. Hassija, S. Gupta, A. Goyal, M. Guizani, and B. Sikdar, “Disaster and pandemic management using machine
learning: a survey,” IEEE Internet of Things Journal, vol. 8, no. 21, pp. 16047–16071, Nov. 2021, doi:
10.1109/JIOT.2020.3044966.
[10] A. Feriani, A. Refaey, and E. Hossain, “Tracking pandemics: a MEC-enabled IoT ecosystem with learning capability,” IEEE
Internet of Things Magazine, vol. 3, no. 3, pp. 40–45, Sep. 2020, doi: 10.1109/IOTM.0001.2000142.
[11] D. Xue, S. Nazir, Z. Peng, and H. Khattak, “Selection and ranking of fog computing-based IoT for monitoring of health using the
analytic network approach,” Complexity, vol. 2021, pp. 1–11, Aug. 2021, doi: 10.1155/2021/9964303.
[12] T. N. Gia, M. Jiang, A.-M. Rahmani, T. Westerlund, P. Liljeberg, and H. Tenhunen, “Fog Computing in healthcare internet of
things: a case study on ECG feature extraction,” in 2015 IEEE International Conference on Computer and Information
Technology; Ubiquitous Computing and Communications; Dependable, Autonomic and Secure Computing; Pervasive
Intelligence and Computing, Oct. 2015, pp. 356–363, doi: 10.1109/CIT/IUCC/DASC/PICOM.2015.51.
[13] A. Elhadad, F. Alanazi, A. I. Taloba, and A. Abozeid, “Fog computing service in the healthcare monitoring system for managing
the real-time notification,” Journal of Healthcare Engineering, vol. 2022, pp. 1–11, Mar. 2022, doi: 10.1155/2022/5337733.
[14] M. Al-khafajiy, L. Webster, T. Baker, and A. Waraich, “Towards fog driven IoT healthcare,” in Proceedings of the 2nd
International Conference on Future Networks and Distributed Systems, Jun. 2018, pp. 1–7, doi: 10.1145/3231053.3231062.
[15] N. El-Rashidy, N. E. ElSayed, A. El-Ghamry, and F. M. Talaat, “Utilizing fog computing and explainable deep learning
techniques for gestational diabetes prediction,” Neural Computing and Applications, vol. 35, no. 10, pp. 7423–7442, Apr. 2023,
doi: 10.1007/s00521-022-08007-5.
[16] V. K. Quy, N. Van Hau, D. Van Anh, and L. A. Ngoc, “Smart healthcare IoT applications based on fog computing: architecture,
applications and challenges,” Complex & Intelligent Systems, vol. 8, no. 5, pp. 3805–3815, Oct. 2022, doi: 10.1007/s40747-021-
00582-9.
[17] Y. Shi, G. Ding, H. Wang, H. Eduardo Roman, and S. Lu, “The fog computing service for healthcare,” in 2015 2nd International
Symposium on Future Information and Communication Technologies for Ubiquitous HealthCare, Ubi-HealthTech 2015, 2015,
pp. 70–74, doi: 10.1109/Ubi-HealthTech.2015.7203325.
[18] P. M. Arunkumar, M. Masud, S. Aljahdali, and M. Abouhawwash, “Healthcare monitoring using ensemble classifiers in fog
computing framework,” Computer Systems Science and Engineering, vol. 45, no. 2, pp. 2265–2280, 2023, doi:
10.32604/csse.2023.032571.
[19] A. A. Mutlag, M. K. Abd Ghani, N. Arunkumar, M. A. Mohammed, and O. Mohd, “Enabling technologies for fog computing in
healthcare IoT systems,” Future Generation Computer Systems, vol. 90, pp. 62–78, Jan. 2019, doi: 10.1016/j.future.2018.07.049.
[20] T. Bezdan and N. Bačanin Džakula, “Convolutional neural network layers and architectures,” in Proceedings of the International
Scientific Conference - Sinteza 2019, 2019, pp. 445–451, doi: 10.15308/Sinteza-2019-445-451.
[21] T. Rahman, M. Chowdhury, and A. Khandakar, “COVID-19 radiography database,” www.kaggle.com, 2021.
https://www.kaggle.com/datasets/tawsifurrahman/covid19-radiography-database.
[22] M. E. H. Chowdhury et al., “Can AI help in screening viral and COVID-19 pneumonia?,” IEEE Access, vol. 8, pp. 132665–
132676, 2020, doi: 10.1109/ACCESS.2020.3010287.
[23] X. Glorot and Y. Bengio, “Understanding the difficulty of training deep feedforward neural networks,” Journal of Machine
Learning Research, vol. 9, pp. 249–256, 2010.
[24] D. P. Kingma and J. L. Ba, “Adam: a method for stochastic optimization,” in 3rd International Conference on Learning
Representations, ICLR 2015 - Conference Track Proceedings, 2015, pp. 1–15.

Revolutionizing healthcare image analysis in pandemic-based fog-cloud … (Alzahraa Elsayed)

 454  ISSN: 2502-4752

[25] H. Chen et al., “Pre-trained image processing transformer,” in Proceedings of the IEEE Computer Society Conference on
Computer Vision and Pattern Recognition, 2021, pp. 12294–12305, doi: 10.1109/CVPR46437.2021.01212.
[26] X. Han et al., “Pre-trained models: past, present and future,” AI Open, vol. 2, pp. 225–250, 2021, doi:
10.1016/j.aiopen.2021.08.002.

BIOGRAPHIES OF AUTHORS

Alzahraa Elsayed received an M.Sc. in Communications Engineering and

Computers Engineering from the University of Al Azhar, Egypt (2018), where she is currently
pursuing a Ph.D. in Communications Engineering from 2019 to 2022, her research interests
include fog computing, cloud computing, and internet of things (IoT) technologies. She can be
contacted at email: [email protected].

Khalil Mohamed received a Ph.D. in robotics and control engineering from Al-
Azhar University, Egypt in 2019. He is currently an assistant professor at Systems and
Computers Engineering Department, at Al-Azhar University, Egypt. His research interests
include AI, machine learning, deep learning, reinforcement learning, robotics, control theory,
intelligent control systems, automotive control systems, robust control, stochastic control,
motion and navigation control, traffic and transport control, predictive control, optimal
control, mathematics, optimization, task assignment in multi-robot systems, task
decomposition. He can be contacted at email: [email protected].

Hany Harb received a B.Sc. degree in computers and control engineering from
the Faculty of Engineering, Ain Shams University, Egypt in 1978, and an M.Sc. degree in
computers and systems engineering from the Faculty of Engineering, Al-Azhar University,
Egypt in 1981. He also received a Ph.D. degree in computer science and an M.Sc. degree in
operations research (MSOR) from the Institute of Technology (IIT), USA in 1986 and 1987,
respectively. He is a professor of Software Engineering in the System Engineering
Department, Faculty of Engineering, Al- Azhar University, Egypt. His research interests
include artificial intelligence, cloud computing, and distributed systems. He can be contacted
at email: [email protected].

Indonesian J Elec Eng & Comp Sci, Vol. 34, No. 1, April 2024: 441-454

View publication stats