Revolutionizing Disease Diagnosis: A

Microservices-Based Architecture for

Privacy-Preserving and Efficient IoT Data
Analytics Using Federated Learning

Safa Ben Atitallah1 , Maha Driss1,2 , and Henda Ben Ghezala1

arXiv:2308.14017v1 [cs.LG] 27 Aug 2023

1
RIADI Laboratory, University of Manouba, Manouba, Tunisia
2
Security Engineering Lab, CCIS, Prince Sultan University, Riyadh, Saudi Arabia

Abstract. Deep learning-based disease diagnosis applications are es-

sential for accurate diagnosis at various disease stages. However, using
personal data exposes traditional centralized learning systems to pri-
vacy concerns. On the other hand, by positioning processing resources
closer to the device and enabling more effective data analyses, a dis-
tributed computing paradigm has the potential to revolutionize disease
diagnosis. Scalable architectures for data analytics are also crucial in
healthcare, where data analytics results must have low latency and high
dependability and reliability. This study proposes a microservices-based
approach for IoT data analytics systems to satisfy privacy and perfor-
mance requirements by arranging entities into fine-grained, loosely con-
nected, and reusable collections. Our approach relies on federated learn-
ing, which can increase disease diagnosis accuracy while protecting data
privacy. Additionally, we employ transfer learning to obtain more effi-
cient models. Using more than 5800 chest X-ray images for pneumonia
detection from a publically available dataset, we ran experiments to as-
sess the effectiveness of our approach. Our experiments reveal that our
approach performs better in identifying pneumonia than other cutting-
edge technologies, demonstrating our approach’s promising potential de-
tection performance.

Keywords: Microservices architecture, federated learning; transfer learn-

ing, data analytics; microservices; disease detection; pneumonia, privacy.

The rapid emergence and evolution of the Internet of Things (IoT) have
facilitated the advancement of Machine Learning (ML) and Deep Learning (DL)
techniques for healthcare data analytics [1]. Recent DL and image processing
breakthroughs have enabled the development of state-of-the-art computer-aided
diagnosis systems, which assist pathologists in obtaining highly accurate and
reliable diagnosis results [2]. These cutting-edge systems harness the power of
DL and image processing to augment the diagnostic capabilities of healthcare
professionals, leading to improved patient care and outcomes [3].
In this context, we introduce a novel approach based on microservices and
Federated Learning (FL), which can potentially enhance disease detection accu-
racy while significantly safeguarding data privacy. Microservices are designed to
2 S. Ben Atitallah et al.

be autonomous and decoupled from each other, allowing for flexibility in terms
of development, deployment, and updates [4,5]. In addition, they are highly
responsive to user demands as they can be designed to be lightweight and opti-
mized for specific tasks [6,7]. This means that the system can quickly respond
to user requests, providing real-time or near-real-time results, which is crucial
in healthcare data analytics, where timely insights are often required. Moreover,
microservices can be deployed across multiple edge computing nodes, allowing
for distributed processing and reducing the need to transfer large amounts of
data to a central location. This can result in lower latency, reduced network
overhead, and improved overall system performance.
With FL, collaborative model training is conducted across multiple dis-
tributed devices or institutions without sharing raw data [8]. Instead, only model
updates are exchanged, ensuring the confidentiality of sensitive information [9].
The proposed approach harnesses the collective intelligence of multiple experts,
resulting in a robust and accurate disease detection model. FL effectively ad-
dresses privacy concerns by keeping data localized and minimizing data transfer,
ensuring patient data remains secure and protected.
Besides, Transfer Learning (TL) plays an important role in the healthcare
domain to improve the effectiveness and efficiency of various tasks [10]. TL lever-
ages pre-trained models trained on large-scale datasets, such as ImageNet, and
applies them to healthcare-specific tasks.
By leveraging the pre-trained model’s knowledge through TL and allowing
local fine-tuning on specific target domains, Federated Transfer Learning (FTL)
aims to improve the learning performance and generalization capability of models
trained in a federated setting [11]. It allows for the transfer of learning represen-
tations, which can capture relevant features and patterns while preserving local
data’s privacy.
Our proposed approach embodies a remarkable stride in disease detection,
harnessing the exceptional scalability and swift response time of microservices,
the invaluable privacy-preserving advantages of FL, and the augmentation of
model performance, generalization, and prediction accuracy through TL.
The paper outlines several key contributions, which can be summarized as
follows:

– The proposal of a novel microservices-based approach for disease diagnosis.

– The improvement of privacy in disease diagnosis through applying FL.
– The utilization of TL to improve model performance in disease diagnosis
case studies.
– The application and assessment of the proposed approach in a real healthcare
case study.

The structure of the paper is organized as follows. Section 2 provides an

overview of the relevant literature and related works in the field of disease detec-
tion. Section 3 presents the details of the proposed microservices-based approach
for disease detection, including the use of FL and TL. Section 4 covers the im-
plementation details of the proposed approach and presents the experimental
 Title Suppressed Due to Excessive Length 3

analysis conducted to evaluate its performance. Section 5 concludes the study

and outlines potential future research directions in disease diagnosis.

1 Related Works

This study’s literature review emphasizes the significant improvements in health-

care data analytics with the advent and expansion of IoT, ML, and DL tech-
niques. This study mainly deals with the emergence of computer-aided diagnos-
tics systems, which leverage the power of DL and image processing to improve
healthcare practitioners’ diagnostic capabilities.
[12] suggested altering the InceptionV3 model, adding five more layers, and
using TL to create a DL-based automated diagnostic tool for detecting pediatric
pneumonia. With the help of gamma correction and compression, the suggested
model’s accuracy reached 90.1%. The suggested model was tested on a standard
Women and Children’s Medical Center X-ray dataset. Its precision, recall, ac-
curacy, and ROC accuracy scores were compared to those of ResNet, ImageNet,
Xception, and Inception. Unlike our approach, which takes advantage of FL and
TL, the authors concentrated on preprocessing the dataset and training/testing
multiple CNN architectures.
In [13], a computer-aided diagnosis method for automatically identifying
pneumonia using images of chest X-rays was developed. Deep TL was used to
deal with the restricted amount of data provided, and an ensemble of three convo-
lutional neural network models—GoogLeNet, ResNet18, and DenseNet121—was
generated. The base learner weights were determined using the weighted average
ensemble methodology. A five-fold cross-validation approach was used to evalu-
ate the suggested method using two pneumonia X-rays available to the general
audience, the Kermany and RSNA datasets. The validation demonstrated that
the method outperformed commonly used ensemble approaches and produced
results superior to state-of-the-art techniques. This study focused on centralized
training and testing of DL models and did not address security and privacy
concerns related to handling patients’ sensitive data.
A novel approach based on evidence-based fusion theory was suggested in
[14], allowing a series of DL classifiers to be fused for more accurate illness detec-
tion findings. This study’s essential contribution was using the Dempster-Shafer
theory to merge five pre-trained convolutional neural networks for detecting
pneumonia from chest X-ray images, including VGG16, Xception, InceptionV3,
ResNet50, and DenseNet201. To test this method, experiments were carried out
utilizing a publically available dataset, including over 5800 chest X-ray pictures.
Compared to existing state-of-the-art approaches, the suggested methodology
demonstrated good detection performance. Contrary to our approach, this study
focused on a conventional methodology of applying and testing DL models.
The study presented in [15] aimed to provide a DL-based framework for
detecting pneumonia from X-ray images using Compressed Sensing (CS) data,
which could be useful for healthcare professionals working in distant areas. The
system was bandwidth-preserving and energy-efficient due to the use of CS,
4 S. Ben Atitallah et al.

which allowed for effective far-end pneumonia diagnosis. Through rigorous sim-
ulations, the suggested technique achieved a pneumonia detection accuracy of
96.48% even when only 30% of the samples were communicated. These results
demonstrated the effectiveness of the suggested strategy in detecting pneumonia,
even with limited information. This study addressed the challenge of applying
DL models to a restricted amount of data; it did not address important issues
related to the proposed approach’s privacy, scalability, and reusability.
A DL architecture based on EfficientNetB7 was proposed in [16] to clas-
sify medical X-ray and Computed Tomography (CT) images of lungs into three
categories: common pneumonia, coronavirus pneumonia, and normal cases. Re-
cent pneumonia detection methods were compared to the proposed architecture
in terms of accuracy. The proposed approach provided robust and consistent
pneumonia detection features with an accuracy rate of 99.81% for radiography
and 99.88% for CT. This study’s approach was confined to a single DL model,
EfficientNetB7. Furthermore, the suggested architecture was assessed without
using distributed or FL approaches, utilizing a traditional training and testing
methodology for applying DL models.
To overcome the abovementioned limitations, this work offers a unique mi-
croservices and FTL-based approach that utilizes the learned knowledge, re-
sulting in a robust and effective disease diagnosis model. The paper includes
several contributions, such as proposing a microservices-based approach for dis-
ease diagnosis, improving privacy through FL, utilizing TL to improve model
generalization,and application in a real-world healthcare case study.

2 Proposed Disease Detection Approach

This study focuses on developing an intelligent method for disease detection

through the utilization of distributed learning techniques. The proposed ap-
proach involves the implementation of a secure, flexible, scalable, and highly
responsive system consisting of a collection of microservices. This architectural
design enables efficient and effective disease detection utilizing distributed learn-
ing techniques while simultaneously ensuring the system’s security, scalability,
and responsiveness.
The proposed approach involves cloud-edge collaborative architecture.The
architecture design of the proposed approach is presented in Figure 1 and show-
cases the integration of cloud and edge nodes. The federated training process
is seamlessly executed across these nodes, enabling efficient collaboration and
leveraging the strengths of each component.
The first step in this process involves the collection of data from various
sources, such as IoT devices, sensors, or other data-generating entities. This raw
data is then transmitted to the edge computing resources, where preprocessing
and local training occur. Once the local training is completed, the model param-
eters are transmitted to the cloud FL server for aggregation. The cloud serves as
a central hub for receiving and consolidating model updates from edge clients.
Through the aggregation process, the cloud combines the knowledge learned
 Title Suppressed Due to Excessive Length 5

from the various edge FL clients, creating a comprehensive and refined global
model that benefits from the collective intelligence of the distributed network.

Fig. 1. The proposed distributed approach architecture for an efficient and secure
disease detection process.

Fig. 2. The architecture of the FL client.

The edge nodes in FL have a vital function by facilitating the training of DL

models on data situated at the network’s edge [17]. This innovative approach
utilizes an edge computer/server, eliminating the need to transmit data to a
centralized server. By employing this process, data privacy in healthcare is sig-
nificantly strengthened, and numerous benefits are achieved, including notable
reductions in network latency and substantial savings in bandwidth consump-
tion. In our proposed work, the edge node consists of a set of FL clients. Each
FL client is formed by a set of layers and composed of service-oriented microser-
6 S. Ben Atitallah et al.

vices that seamlessly collaborate to execute tasks, thereby empowering intelligent

decision-making capabilities.
The FL client consists of four essential layers as illustrated in Figure 2. The
first layer is responsible for data collection, acquiring data from various sources
located at the network’s edge. Once the data is collected, the second layer,
data preprocessing, performs necessary tasks to cleanse and prepare the data for
subsequent stages. The third layer, model learning, establishes communication
with the cloud FL server and initiates the training process based on predefined
settings. This layer collaborates with cloud resources to effectively train and
update the models. Finally, the application layer interprets the results generated
by the trained models, providing insights and predictions.

2.1 Federated Leaning Procedure

In the proposed distributed environment, multiple virtual instances at the net-

work’s edge are used to train models on their respective training data [18]. The
updated parameters are sent to a central cloud FL server for aggregation, cre-
ating a final global model incorporating insights from all individual instances.
This global model is deployed on edge for use in analytics and applications. The
FL model development procedure follows a three-step method, which is outlined
below.

2.1.1 Model Selection and System Initialization The model manager

selects the task, like pneumonia detection, and generates suitable DL models.
Multiple models can be chosen to be trained using the client’s data, and TL tech-
niques can be utilized to improve the performance of the models by leveraging
previously acquired knowledge [11]. The type of models to be used is determined
based on the specific use case, and important parameters such as learning rates
and communication rounds are selected to ensure effective and efficient training
of the models.

2.1.2 Distributed Local Training and Updates After the model manager
has set up the learning parameters and chosen the suitable models, the cloud FL
server initiates the global models. It sends them to the FL clients to commence
the learning process. Each client, labeled as k, trains its own local models us-
ing its own data, denoted as Dk , and calculates an update, denoted as wk , by
minimizing the loss function F (wk ):

wk = argminF (wk ), k ∈ K (1)

Afterward, each client k sends its computed update wk to the FL server for
aggregation. The FL clients only exchange the model updates with the FL server
throughout the training process, ensuring that data privacy is preserved.
 Title Suppressed Due to Excessive Length 7

2.1.3 Models Aggregation and Download The server aggregates locally

trained models from participating FL clients in each round, generating a new
version of the global model. The number of communication rounds in FL refers
to how often this aggregation technique is executed, involving the global model’s
collection, combination, and update. Each round typically includes multiple com-
munication iterations where FL clients send their local models to the FL server,
and the FL server sends back the updated global model. The number of com-
munication rounds in the FL procedure is influenced by model convergence rate,
dataset size, number of IoT devices, and communication bandwidth.
During each training cycle, the local updates provided by FL clients are
merged using the federated averaging function , where the global model is gen-
erated by averaging the weights contributed [19].

2.2 Hybrid Knowledge Transfer in Federated Learning

Traditional Centralized Learning (CL) offers immense computational power but

falls short of ensuring data privacy. To address this, FL has emerged as a privacy-
preserving paradigm [20]. However, FL may face challenges in terms of compu-
tational efficiency compared to CL. To further enhance model performance, we
propose the integration of hybrid TL within FL. This approach combines the
strengths of both TL and FL, enabling efficient knowledge transfer while safe-
guarding privacy. Hybrid TL leverages the power of pre-trained models, which
have learned representations from large-scale centralized datasets, and combines
them with FL to enhance model performance [21]. By employing TL techniques,
models can benefit from existing knowledge while adapting to the specific data
distributions present on distributed devices.
The hybrid TL approach in FL involves several steps. First, a base pre-trained
model, which captures high-level features, is selected. Next, the base model is
fine-tuned using FL on the local data of each client. Each FL client trains the
model on its data while preserving privacy by only sharing model updates rather
than raw data. The central FL server aggregates the model updates from all
devices, allowing the global model to benefit from the collective knowledge of
the devices.
In our proposed approach, we enhance the method of disease detection by
incorporating different TL models. We employ popular CNN architectures, in-
cluding VGG16, Xception, MobileNetV2, DenseNet201, and InceptionV3.

2.3 Analytical Process in Microservices

A collaborative ensemble of microservices is designed to work together, ana-

lyze data, and generate valuable insights for the data analytics process. These
microservices are specifically designed and optimized for the main analytical
tasks such as data preprocessing, model development, and interpretation. The
microservices designed for each task are presented in the following:
8 S. Ben Atitallah et al.

2.3.1 Data Preprocessing Stage The initial stage in data analysis is data
preprocessing, which involves the following components:
- Data integration: This component merges and arranges data from different
sources to achieve a uniform format.
- Data scaling/transformation: This component standardizes and modifies
the collected data to a predefined range, aiming to improve data consistency.
This step ensures that the data is reliable and suitable for subsequent analysis.

2.3.2 Models Development Stage The model development stage is a cru-

cial phase in the process of creating and refining models. In our proposed ap-
proach, we adopt a collaborative approach where the model development takes
place in conjunction with both the cloud server and the edge FL clients. The
components deployed in the cloud are the following:
- Model creator: The model creator microservice generates and compiles the
initial models before the training process.
- Model uploader: The model uploader microservice is responsible for securely
and efficiently transferring the trained model parameters from the edge nodes to
the central server. It ensures that the model parameters are transferred securely
and reliably, maintaining the integrity and confidentiality of the model data.
- Model aggregator: The model aggregator microservice in the cloud is a cru-
cial component in FL. It receives the updated model parameters from the edge
devices and combines them to generate a new global model.
In the edge, these microservices are employed:
- Model training: The model training microservice is responsible for conduct-
ing the training process according to the configuration parameters specified by
the global model in the FL server. This microservice is deployed on the edge
nodes.
- Model evaluator: The model evaluator microservice assesses the performance
of the model based on predefined metrics or criteria, such as accuracy, loss, or
other relevant indicators.

2.3.3 Interpretation Stage In our approach, the last step in data analytics
involves utilizing the learned models to make predictions or decisions using new
data.
- Application microservice: This microservice utilizes the trained models to
make predictions or classifications when provided with new input data.

3 Implementation and Experimental Analysis

This section provides information about the dataset used for conducting the
experiments. We also define the evaluation metrics employed to assess the per-
formance of the proposed approach. Lastly, we present the results obtained from
the experiments and conduct a detailed analysis of the findings in the last sub-
section.
 Title Suppressed Due to Excessive Length 9

3.1 Dataset
The dataset utilized for the experiments is named Pneumonia Chest X-ray Im-
ages and is publicly available [22]. It was collected by the Guangzhou Women and
Children’s Medical Center in China and can be found on Kaggle. The dataset
contains a total of 5855 images, which are organized into three folders: train,
validation, and test. The dataset is imbalanced, with a class distribution of ap-
proximately 30% normal images and 70% pneumonia images.

3.1.1 Experimental Setup The suggested solution was designed and vali-
dated using publicly accessible open-source components. It adopts a microservices-
based architectural style along with FL and TL approaches, ensuring scalability
and reliability. Docker containerization was used to create distributed images
across three virtual machines, where each virtual machine hosted microservices
of each data analytics stage using Docker instances [23]. Swarm orchestration
managed the container cluster, providing high availability for applications [24].
TensorFlow Federated (TFF) [25] and Keras [26], open-source libraries, are used
in the implementation of the models. These libraries offer powerful functionalities
and tools for developing and training ML/DL models with the FL setting.

3.2 Evaluation Metrics

The evaluation metrics used in the proposed approach are as follows:
Accuracy: It measures the model’s overall performance across all categories
and is calculated using Equation 2.
Precision: It assesses the model’s accuracy in classifying a sample as positive
(pneumonia) or negative (normal) and is calculated using Equation 3.
Recall: It measures the model’s ability to identify positive samples and is cal-
culated using Equation 4.
F1-score: It combines accuracy and recall to provide a balanced performance
measure and is calculated using Equation 5. It considers both precision and
recall.
TP + TN
Accuracy = (2)
TP + TN + FP + FN
TP
P recision = (3)
TP + FP
TP
Recall = (4)
TP + FN
2 ∗ P recision ∗ Recall
F 1 − score = (5)
P recision + Recall
These metrics provide quantitative measures to evaluate the performance of
the proposed approach in classifying normal and pneumonia chest X-ray images
from the dataset. It is important to interpret these metrics collectively to com-
prehensively understand the model’s performance and make informed decisions
about its effectiveness for the studied task.
10 S. Ben Atitallah et al.

3.3 Experimental Results and Discussion

In this subsection, the results of the proposed approach for pneumonia detec-
tion are presented and analyzed. The outcomes of the implemented solution are
discussed, and their implications are examined in detail.

3.3.1 Assessment of Approach: Analytics Findings for Pneumonia

Detection
1. Data Preprocessing Stage
The dataset images were preprocessed by resizing them to a resolution of
224x224 and normalizing the pixel values to ensure consistency and com-
patibility across the dataset. In addition, we employed various techniques
to augment the dataset, including re-scaling, horizontal flipping, random ro-
tation, width and height shift, and adjusting zoom and brightness levels.
These techniques were applied to enhance the diversity and variability of
the dataset, leading to potentially improved model performance and gener-
alization.
In order to enable autonomous learning, the dataset was split into training
and testing datasets, and these datasets were distributed among five separate
clients.
2. Models Development
The proposed approach was implemented using the TensorFlow Federated
framework on 5 virtual IoT devices with a distributed configuration. Five
pre-trained CNN architectures were utilized for TL to an efficient pneu-
monia detection, including VGG16, Xception, InceptionV3, ResNet50, and
DenseNet201. Models’ aggregation was performed over 15 communication
rounds, with each CNN trained for 20 epochs using local data. The Adam
optimizer with a learning rate of 1e-4, the batch size of 32, and the cross-
entropy loss function were used for the parameter setting of DL models.
3. Interpretation Stage
After training the five CNNs, their performance was evaluated. The classifi-
cation results of each model are shown in Table 2. The DenseNet201 model
achieved the best performance, therefore it was selected to be deployed in
the clients as the final global model for interpretation. The performance dif-
ference between the models is not significant, indicating relatively balanced
performance. The performance outcomes of the generated models are pre-
sented in Table 1.

3.3.2 Microservices Performance Assessment Moreover, the performance

of the proposed approach was assessed by evaluating the execution time of mi-
croservices based on their functionality. The end-to-end response time of the
pneumonia detection function f is defined as the sum of the time spent in data
pre-processing TP re and Interpretation TInter stages, as shown in Equation 6.

Tf = TP re + TInter (6)
 Title Suppressed Due to Excessive Length 11

Table 1. Performance results of the learned models using TL and FL.

Model//
VGG16 Xception ResNet50 InceptionV3 DenseNet201
Metric
Accuracy 95.6 97.2 95.1 96.4 98.1
Precision 95.4 97.4 94.6 96.8 97.9
Recall 96.2 97.9 95.3 96.5 98.3
F1-score 95.8 97.6 94.9 96.6 98.1

Figure 3 showcases the execution time of each microservice in the pneumonia

detection process using the DenseNet201 model, encompassing data preparation
and interpretation. The data preprocessing microservice took 198 ms, while the
interpretation microservice took 264 ms. As a result, the total time for the pneu-
monia detection process was calculated to be approximately 462 ms. This rapid
model responses enable healthcare professionals to make informed and data-
driven decisions promptly. This can aid in selecting appropriate treatment plans
and determining the urgency of interventions.

Fig. 3. The execution time of the data preprocessing and application microservices in
addition to the end-to-end response time for the pneumonia detection process.

3.3.3 Comparison To validate the proposed approach, we compared our pre-

viously published study, which focused on the same use case scenario and utilized
the same dataset but adopted a CL setting with monolithic architectures [14].
The performance results, as depicted in Table 2, clearly demonstrate that our
approach outperforms the previous study in terms of achieving the highest per-
formance results. This validates the effectiveness of our approach that utilizes
decentralized TL models in an FL architecture.

3.4 Discussion
To ensure accurate pneumonia diagnosis, it is essential to identify all relevant
features present in the chest X-ray images. The hybrid TL approach in FL offers
12 S. Ben Atitallah et al.

Table 2. Comparison of the proposed approach results with our previously published
study

Model Type of learning Architecture Accuracy Precision Recall F1-score

[14] Centralized Monotholic 97.5 97.5 98 97.8
Proposed approach Distributed Microservices 98.1 97.9 98.3 98.1

several benefits. By leveraging pre-trained models, it harnesses the power of CL

while preserving privacy through FL. The collective knowledge of the distributed
devices enhances the model’s performance.
Privacy concerns arise when dealing with sensitive medical data. For this, FL
was used to address these concerns by enabling the training of models on dis-
tributed data while keeping the data securely on their local devices. This allowed
for collaborative model training across multiple institutions without sharing the
raw data, preserving privacy and confidentiality.
In addition to privacy considerations, achieving high scalability and response
time is crucial in healthcare. Microservices architecture was leveraged to ad-
dress these requirements by breaking the data analytics function into smaller,
specialized microservices that can be deployed and scaled independently. This
allowed for efficient resource utilization and improved scalability. Moreover, mi-
croservices architecture enabled faster response times as each microservice can
be optimized for performance and response time, leading to quicker results for
end-users.
A robust and efficient pneumonia diagnosis solution was achieved by combin-
ing TL for high model performance, FL for privacy preservation, and microser-
vices architecture for scalability and response time optimization. This approach
ensures accurate predictions, protects patient privacy, and enables efficient and
scalable data analytics in the healthcare environment.

4 Conclusion
This research aims to provide a novel approach for high-performing disease di-
agnosis models while ensuring privacy compliance. In detecting pneumonia us-
ing X-ray images, this solution uses TL for model development, FL for privacy
preservation, and microservices architecture for scalability and response time
optimization.
Our approach’s experimental findings show that it can give accurate predic-
tions, preserve patient privacy, and facilitate efficient and scalable data analytics
in a healthcare setting.
In the future, we aim to examine our approach’s performance on various
datasets to confirm its wide applicability and robustness in real-world case
studies. In addition, we intend to develop and test fusion techniques, such as
Dempster-Shafer-based fusion, weighted averaging, and attention-based strate-
gies. These strategies can potentially improve the accuracy and robustness of
the fusion process, and their effectiveness will be evaluated in the context of our
 Title Suppressed Due to Excessive Length 13

approach. In addition, we intend to explore and analyze the possible influence

of various parameters on model performance, such as demographics, comorbidi-
ties, and disease severity, which might give useful insights into improving the
suggested approach’s overall performance.

References
1. A. Qayyum, J. Qadir, M. Bilal, and A. Al-Fuqaha, “Secure and robust machine
learning for healthcare: A survey,” IEEE Reviews in Biomedical Engineering,
vol. 14, pp. 156–180, 2020.
2. S. B. Atitallah, M. Driss, W. Boulila, and H. B. Ghézala, “Leveraging deep learning
and iot big data analytics to support the smart cities development: Review and
future directions,” Computer Science Review, vol. 38, p. 100303, 2020.
3. S. Ben Atitallah, M. Driss, W. Boulila, and H. Ben Ghezala, “Randomly initialized
convolutional neural network for the recognition of covid-19 using x-ray images,”
International journal of imaging systems and technology, vol. 32, no. 1, pp. 55–73,
2022.
4. S. B. Atitallah, M. Driss, and H. B. Ghzela, “Microservices for data analytics in iot
applications: Current solutions, open challenges, and future research directions,”
Procedia Computer Science, vol. 207, pp. 3938–3947, 2022.
5. M. Driss, S. Ben Atitallah, A. Albalawi, and W. Boulila, “Req-wscomposer: a novel
platform for requirements-driven composition of semantic web services,” Journal
of Ambient Intelligence and Humanized Computing, pp. 1–17, 2022.
6. D. Hasan and M. Driss, “Sublµme: Secure blockchain as a service and
microservices-based framework for iot environments,” in 2021 IEEE/ACS 18th In-
ternational Conference on Computer Systems and Applications (AICCSA). IEEE,
2021, pp. 1–9.
7. M. Driss, A. Aljehani, W. Boulila, H. Ghandorh, and M. Al-Sarem, “Servicing your
requirements: An fca and rca-driven approach for semantic web services composi-
tion,” IEEE Access, vol. 8, pp. 59 326–59 339, 2020.
8. J. Xu, B. S. Glicksberg, C. Su, P. Walker, J. Bian, and F. Wang, “Federated
learning for healthcare informatics,” Journal of Healthcare Informatics Research,
vol. 5, pp. 1–19, 2021.
9. M. Driss, I. Almomani, Z. e Huma, and J. Ahmad, “A federated learning framework
for cyberattack detection in vehicular sensor networks,” Complex & Intelligent
Systems, vol. 8, no. 5, pp. 4221–4235, 2022.
10. H. Malik, M. S. Farooq, A. Khelifi, A. Abid, J. N. Qureshi, and M. Hussain,
“A comparison of transfer learning performance versus health experts in disease
diagnosis from medical imaging,” IEEE Access, vol. 8, pp. 139 367–139 386, 2020.
11. Y. Chen, X. Qin, J. Wang, C. Yu, and W. Gao, “Fedhealth: A federated transfer
learning framework for wearable healthcare,” IEEE Intelligent Systems, vol. 35,
no. 4, pp. 83–93, 2020.
12. P. Chhikara, P. Singh, P. Gupta, and T. Bhatia, “Deep convolutional neural net-
work with transfer learning for detecting pneumonia on chest x-rays,” in Advances
in Bioinformatics, Multimedia, and Electronics Circuits and Signals: Proceedings
of GUCON 2019. Springer, 2020, pp. 155–168.
13. R. Kundu, R. Das, Z. W. Geem, G.-T. Han, and R. Sarkar, “Pneumonia detection
in chest x-ray images using an ensemble of deep learning models,” PloS one, vol. 16,
no. 9, p. e0256630, 2021.
14 S. Ben Atitallah et al.

14. S. Ben Atitallah, M. Driss, W. Boulila, A. Koubaa, and H. Ben Ghezala, “Fusion
of convolutional neural networks based on dempster–shafer theory for automatic
pneumonia detection from chest x-ray images,” International Journal of Imaging
Systems and Technology, vol. 32, no. 2, pp. 658–672, 2022.
15. S. R. Islam, S. P. Maity, A. K. Ray, and M. Mandal, “Deep learning on compressed
sensing measurements in pneumonia detection,” International Journal of Imaging
Systems and Technology, vol. 32, no. 1, pp. 41–54, 2022.
16. A. Moussaid, N. Zrira, I. Benmiloud, Z. Farahat, Y. Karmoun, Y. Benzidia,
S. Mouline, B. El Abdi, J. E. Bourkadi, and N. Ngote, “On the implementation
of a post-pandemic deep learning algorithm based on a hybrid ct-scan/x-ray im-
ages classification applied to pneumonia categories,” in Healthcare, vol. 11, no. 5.
MDPI, 2023, p. 662.
17. G. Bao and P. Guo, “Federated learning in cloud-edge collaborative architec-
ture: key technologies, applications and challenges,” Journal of Cloud Computing,
vol. 11, no. 1, p. 94, 2022.
18. T. Li, A. K. Sahu, A. Talwalkar, and V. Smith, “Federated learning: Challenges,
methods, and future directions,” IEEE signal processing magazine, vol. 37, no. 3,
pp. 50–60, 2020.
19. B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas,
“Communication-efficient learning of deep networks from decentralized data,” in
Artificial intelligence and statistics. PMLR, 2017, pp. 1273–1282.
20. S. AbdulRahman, H. Tout, H. Ould-Slimane, A. Mourad, C. Talhi, and M. Guizani,
“A survey on federated learning: The journey from centralized to distributed on-
site learning and beyond,” IEEE Internet of Things Journal, vol. 8, no. 7, pp.
5476–5497, 2020.
21. Y. Liu, Y. Kang, C. Xing, T. Chen, and Q. Yang, “A secure federated transfer
learning framework,” IEEE Intelligent Systems, vol. 35, no. 4, pp. 70–82, 2020.
22. P. MOONEY, “Chest-xray pneumonia dataset,” https://www.kaggle.com/
datasets/paultimothymooney/chest-xray-pneumonia, accessed: April 19, 2023.
23. Docker, “Documentation,” https://docs.docker.com/, 2022, accessed: April 6,
2023.
24. ——, “Orchestration,” https://docs.docker.com/get-started/orchestration/, 2023,
accessed: April 19, 2023.
25. TensorFlow, “Tensorflow federated: Machine learning on decentralized data,”
https://www.tensorflow.org/federated, 2023, accessed: April 19, 2023.
26. Keras, “Keras library,” https://keras.io/, 2023, accessed: May 19, 2023.