Advances in Pneumonia Detection: A Comprehensive Investigation of

Federated Learning and Deep Learning-Based Approaches

Bingchen Duan

Computer Science and Technology, Northeast Forestry University, Harbin, China

Keywords: Federated Learning, Pneumonia Detection, Convolutional Neural Networks, Ensemble Learning.

Abstract: In the realm of healthcare, federated learning (FL) emerges as a promising solution to address the challenges

of data silos and privacy concerns in medical diagnosis. This paper delves into the application of FL in the

context of pneumonia detection, with a focus on leveraging convolutional neural networks (CNNs) within a

federated learning framework. The study provides a comprehensive overview of the potential of FL in

processing sensitive medical data, particularly in enhancing the accuracy of pneumonia detection. By

employing deep learning models such as Convolutional Neural Networks, VGG-16, ResNet50, and

DenseNet121, the research demonstrates significant improvements in detection accuracy. Furthermore, the

paper explores the integration of ensemble learning with federated learning, highlighting its potential to

augment the generalization capabilities of models while bolstering data privacy protection. Despite the

promising results, the study also identifies several key challenges that need to be addressed, including issues

related to data quality, communication overhead, evolving healthcare regulations, and the need for

standardization in the application of federated learning in healthcare settings. Overall, this paper underscores

the potential of federated learning in revolutionizing the diagnosis of pneumonia while ensuring the protection

of patient privacy and data security.

1 INTRODUCTION

In today's healthcare sector, accurate and timely

detection of pneumonia is crucial for enhancing

patient recovery rates and reducing disease

transmission. As a principal cause of mortality

globally among various age groups, especially for

young and old (Hespanhol & Bárbara, 2019; Ngari et

al., 2017), the urgency for effective pneumonia

detection mechanisms has never been more

pronounced. With the increasing development of the

healthcare industry, a large number of multi-

structured patient data from clinical reports, doctor's

notes, wearable devices, etc., are being generated

every day. The advent of voluminous medical data

has led to a significant shift towards leveraging

machine learning and Artificial Intelligence (AI)

technologies to boost the efficiency and accuracy of

pneumonia detection (Ni et al. ,2020). Despite the

promise shown by these technologies, their

deployment faces substantial hurdles due to data

isolation and privacy concerns.

https://orcid.org/0009-0007-0730-5084

The proliferation of AI in medical diagnostics has

showcased the potential of Deep Learning (DL)

techniques in interpreting complex medical images

(Qiu, 2019; Qiu, 2022), including chest CT scans and

X-rays, for pneumonia detection. The COVID-19

pandemic has underscored the critical role of

advanced diagnostic tools, with several studies

demonstrating the efficacy of DL in enhancing

pneumonia detection rates from imaging data (Chen

et al., 2020). These advancements highlight the

transformative impact of AI on medical diagnostics,

offering a glimpse into the potential for more

accurate, efficient, and early detection of diseases

(Zheng et al., 2021).

However, the majority of this research is confined

to datasets within single healthcare institutions,

largely due to the sensitive nature of medical data and

stringent privacy laws. This limitation restricts the

generalizability of AI models, as datasets from a

single source may not adequately represent the

diverse manifestations of diseases across different

populations. Federated Learning (FL) emerges as a

714

Duan, B.

Advances in Pneumonia Detection: A Comprehensive Investigation of Federated Learning and Deep Learning-Based Approaches.

DOI: 10.5220/0012969400004508

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Engineering Management, Information Technology and Intelligence (EMITI 2024), pages 714-718

ISBN: 978-989-758-713-9

promising solution to this challenge (Bonawitz et al.,

2019). By enabling multiple entities to

collaboratively train models without exchanging raw

data, FL addresses privacy concerns while tapping

into a broader data pool. This approach not only

ensures patient privacy but also enhances the model's

ability to learn from a diverse set of data

representations, potentially increasing its diagnostic

accuracy and generalizability.

While the theoretical promise of Federated

Learning (FL) in medical AI is well-recognized,

practical applications remain relatively unexplored,

especially in the context of cross-institutional

pneumonia detection. This study aims to analyze the

deployment of various FL frameworks for pneumonia

detection, with the goal of providing valuable insights

and guidance for future applications of FL in this

domain.

2 METHOD

2.1 Federated Learning

Federated learning shown in Figure 1 is a paradigm

in machine learning that enables multiple participants

(which could be institutions or devices) to

collaboratively train an algorithmic model without

the need for direct data exchange between them

(Bonawitz et al., 2019). A global model in the

aggregation server is distributed to different local

train nodes for training on their respective data.

Rather than transferring the data, the weights are

exchanged between local nodes and aggregation

server. After the local training, local nodes send their

trained weights back to the aggregation server. Then

the server updates the global model by integrating the

trained weight returned by each local training node.

Various aggregation algorithms, such as FedAvg

(McMahan et al., 2017), Robust Aggregation (Wan &

Chen, 2021), SEAR (Zhao et al., 2022), have been

widely implemented in practical applications. The

updated global model is then sent back to the local

train nodes, and the iterative training process

continues. This approach demonstrates the unique

advantages of federated learning, which effectively

protects patient privacy when dealing with data across

healthcare organizations (Farkaš, Ciobanu, & Dobre,

2023). During federated learning, different healthcare

organizations can jointly contribute to model updates

while avoiding the sharing of sensitive patient data.

In a task such as pneumonia detection, where data

quality and dataset size are critical, federated learning

is able to learn across different healthcare

organizations, increasing the sources of data. Thus,

federated learning can accelerate the development of

pneumonia diagnostic techniques while ensuring

diagnostic accuracy.

Figure 1: The general workflow of federated learning

(Photo/Picture credit: Original).

2.2 Convolutional Neural Networks

(CNNs)

Convolutional neural networks are the cornerstone of

modern image analysis, especially in the field of

medical imaging. Their ability to adaptively learn

spatially hierarchical features in images makes them

well suited for classification and object detection. In

federated learning for pneumonia detection, CNNs

play a pivotal role due to their high efficiency in

processing image data and extracting relevant

features critical for accurate diagnosis (Liu, 2023;

Zhang et al., 2021; An, Chen, & Shao, 2024).

2.3 Application in Pneumonia

Detection with CNNs in Federated

Learning

2.3.1 VGG16

Currently VGG16 have achieved excellent

performance in federated learning for pneumonia

detection, according to (Farkaş, Ciobanu, and Dobre,

2023), federated learning models of CNNs

demonstrated promising results, reaching an accuracy

of about 95% in pneumonia detection and effectively

protecting data privacy (Farkaš, Ciobanu, & Dobre,

2023). They implemented their model using Keras

based on the TensorFlow Federated (TFF) framework

and the VGG16 (Wang, 2020) architecture. VGG16

uses a 3x3 convolutional filter (kernel) instead of a

larger filter, effectively reducing the number of

trainable parameters and enabling efficient feature

detection. They also added variety to the dataset

through transformations such as horizontal rotation,

scaling ranges, and width/height shifts. Through the

Advances in Pneumonia Detection: A Comprehensive Investigation of Federated Learning and Deep Learning-Based Approaches

715

combination of multiple advanced techniques and

models for image preprocessing techniques, a

federated learning model based on VGG16 is

realized, and the effectiveness of CNN in the

federated learning process is verified. However, since

the dataset is only derived from children in one

hospital, this will limit the generalization ability of

the model in other age groups, and this study divides

a dataset into ten subsets for simulation of federated

learning, which may not be able to realistically

simulate the distribution of data in the real world.

This study directly addresses the challenge of using

machine learning to detect pneumonia while

overcoming the barriers of data privacy and

computational efficiency.

In the work of (Khan et al., 2021), a federated

learning model was proposed. Their dataset has 4,

684 training images and 1, 172 test images classified

as either pneumonia or normal. The raw data is

processed by the canny filter, which improves the

accuracy of the deep learning model by about 10%.

The processed data set was split into two parts, which

were used to simulate the local data sets of the two

hospitals, trained using a federated learning process,

The performance of CNN model (using 4

Convolution, Max-Pooling and 2 Dense), ResNet50

model, VGG16 model and AlexNet model is

compared. The study evaluated each model based on

its training and validation accuracy, global accuracy,

precision, recall, F1 score, Cohen's Kappa score, area

under the ROC curve, sensitivity, specificity, and

construction time.

2.3.2 ResNet-50 and DenseNet121

ResNet-50 and DenseNet are also state-of-the-art

model architectures based on CNNs. DenseNet

improves the traditional CNN architecture by

introducing a dense connectivity mechanism, which

enhances the transfer of features and reduces the

number of parameters to ensure that the information

is maximally utilized in the network (Nithya,

MohanaSundaram, & Santhosh, 2023; Huang, Liu,

Van Der Maaten, & Weinberger, 2017).

According to a study by (Kareem et al., 2023), a

federated learning framework was developed to

enhance pneumonia image detection using distributed

data across different institutions, it enables different

healthcare organizations and hospitals to participate

in the training of the model together. This study

deployed four virtual devices within the framework,

each representing a training entity, and the models

were trained independently on their respective

devices, ensuring localized data processing and

privacy protection. The study began with the

collection of a dataset containing both pneumonia and

non-pneumonia images, followed by a series of

standardized data preprocessing operations and deep

dives into the data features through exploratory data

analysis (EDA). Once the data preprocessing was

completed, the data were divided into 70% training

set, 20% validation set and 10% test set. In addition,

the training set data was equally distributed to four

virtual devices, where the model hyperparameters and

optimizer were tuned and added.

In this study, various CNN models including

Resnet-50, AlexNet, Densenet, Inception, and

VGG.19 were employed for experiments under a joint

learning framework, involving 20 training cycles

across four virtual clients. Despite CNNs' advantages

in scalability and overfitting prevention with large

datasets, the study observed a notable discrepancy in

performance when models were applied individually

versus in a joint learning scenario. Particularly, the

aggregated performance in joint learning did not meet

expectations, notably in disease classification tasks.

DenseNet and ResNet-50 showed comparable

performance in joint learning, with DenseNet slightly

underperforming alone, suggesting data dispersion's

negative impact. AlexNet and Inception faced

significant performance drops, indicating issues with

data adaptability and model aggregation,

respectively. VGG.19 maintained relative

performance, albeit lower than solo use. Conversely,

ResNet-50 exhibited minimal performance decline,

underscoring its robustness and suitability for

federated learning in pneumonia detection models.

2.3.3 Ensemble Learning

The ensemble learning approach improves the

predictive performance of a model by combining

multiple learning algorithms. Its core idea is to

overcome the limitations of a single model by

aggregating the prediction results of multiple models,

and to reduce the overall prediction error by utilizing

the strong points of multiple models (Sulistya,

Bangun, & Tyas, 2023). In convolutional neural

network modeling, using an integration strategy

provides results superior to any single model.

In convolutional neural network modeling, using

an ensemble strategy provides results superior to any

single model (Mabrouk et al., 2022). In the 2023

study, Mabrouk et al. explore an ensemble federated

learning approach designed to facilitate collaborative

pneumonia diagnosis. This methodology integrates

multiple learning algorithms and distributed data

sources, enabling various institutions to share

learning models instead of raw data. This approach

enhances pneumonia diagnosis accuracy and

efficiency while safeguarding privacy (Mabrouk,

Redondo, Abd Elaziz, & Kayed, 2023).

EMITI 2024 - International Conference on Engineering Management, Information Technology and Intelligence

716

The research adopts a two-phase strategy, starting

with independent learning at the local node (hospital)

level using multiple Convolutional Neural Network

(CNN) models, followed by coordinating and

ensembling these learnings at the global level through

a federated learning framework, aiming to build a

more robust and accurate globally integrated model

(GEL) while ensuring that the data privacy of

individual nodes is protected.

In a two-phase approach to improve pneumonia

diagnosis, local nodes (hospitals) first employ eight

pre-trained CNN models, including densenet169,

mobilenetv2, xception, inceptionv3, resnet50, vgg16,

densenet121, and resnet152v2, for Local Ensemble

Learning (LEL). These models, initially trained on

ImageNet for broad image recognition, are further

tailored to chest x-ray datasets at each node. Based on

performance, primarily accuracy, each node selects

its top two models to create an LEL model.

In the subsequent federated learning phase, nodes

transmit their LEL models to a central server, which

aggregates them into a Global Ensemble Learning

(GEL) model. This federated approach facilitates the

integration and sharing of learned knowledge across

nodes without exposing sensitive data. The GEL

model is then circulated back to nodes for further

refinement and validation, in an iterative process

aimed at optimizing the global model's performance,

thereby enhancing the accuracy and efficiency of

pneumonia diagnostics while maintaining data

privacy.

3 DISCUSSIONS

In exploring the application of FL in pneumonia

image detection, many studies not only reveal its

great potential in protecting patient privacy and

facilitating cross-institutional collaboration, but also

must confront the multiple challenges and limitations

in applying this technology to real-world healthcare

scenarios.

Data quality plays a crucial role in sharing models

across hospitals and healthcare organizations

(Mashoufi, Ayatollahi, Khorasani-Zavareh, & Boni,

2022). If there are data quality issues in the

participants, such as corrupted or inaccurate data, it

will directly affect the training effect and

performance of the machine learning models. This

data heterogeneity problem requires us to perform

strict data quality control before data preprocessing

and model training to ensure the effectiveness of

model training. At the same time, the establishment

of a unified standard Healthcare Data Warehouse may

be an effective solution (Berndt, Fisher, Hevner, &

Studnicki, 2001), the establishment of a standardized

data warehouse to improve data quality, ensure data

security and privacy, and ensure that data can

effectively support decision making, should be able

to help federated learning data quality.

Communication overhead is also a major

challenge for real-time implementation of FL

frameworks. The amount of communication required

for model updating and synchronization among

collaborating organizations with large-scale datasets

can be huge, especially in bandwidth-constrained

environments, which may negatively affect the

efficiency and performance of model training.

Therefore, finding efficient communication strategies

and compression techniques is key to optimizing the

efficiency of FL implementation (Oh, Lee, Jeon, &

Poor, 2021; Huang, Li, & Li, 2023).

In addition, there is a lack of uniform

standardization in real-world implementations of FL.

Different organizations may use different data

formats, protocols, and architectures, which increases

the difficulty of collaboration and the cost of

implementation. The development and adoption of

common standards is essential to facilitate the

widespread adoption of FL technologies (Zhan, Li,

Guo, & Qu, 2021). In this way, the creation of a

federated learning with incentives facilitates data

sharing between different organizations and enables

the development of common standards. In the future,

it may be possible to devise ways to punish malicious

actors through incentive mechanisms.

Currently, there is no clear regulatory compliance

framework for FL methods in the healthcare industry.

This means that in order to achieve compliance,

additional resources may need to be invested in

developing and implementing complex compliance

strategies, which is not only a cumbersome process,

but can also be quite costly.

4 CONCLUSIONS

This paper comprehensively reviews the application

of federated learning in pneumonia detection,

especially the great potential it shows in processing

sensitive medical data. By analyzing convolutional

neural networks, ResNet50 and DenseNet121, the

effectiveness of deep learning models has been

demonstrated in improving the accuracy of

pneumonia detection. In addition, this study explores

how integration learning and federated learning,

when used in combination, can effectively improve

the generalization ability and data privacy protection

Advances in Pneumonia Detection: A Comprehensive Investigation of Federated Learning and Deep Learning-Based Approaches

717

of the models. Although federated learning has

demonstrated many advantages in pneumonia

detection, several key issues and challenges were

identified, including data quality issues,

communication overhead, changing healthcare

regulations, and uniform standardization of federated

learning. Future research could also explore how

federated learning can be combined with other

innovative technologies (e.g., quantum computing

and blockchain) to further improve the efficiency and

safety of pneumonia detection.

REFERENCES

An, Q., Chen, W., et al. 2024. A Deep Convolutional Neural

Network for Pneumonia Detection in X-ray Images

with Attention Ensemble. Diagnostics, 14(4), 390.

Berndt, D. J., Fisher, J. W., et al. 2001. Healthcare data

warehousing and quality assurance. Computer, 34(12),

56-65.

Bonawitz, K., Eichner, H., et al. 2019. Towards Federated

Learning at Scale: System Design. ArXiv,

abs/1902.01046.

Chen, J., Wu, L., et al. 2020. Deep learning-based model

for detecting 2019 novel coronavirus pneumonia on

high-resolution computed tomography. Scientific

Reports, 10.

Farkaş, C. R., Ciobanu, R. I., et al. 2023. Pneumonia

Detection Using Federated Learning. In 2023 24th

International Conference on Control Systems and

Computer Science (CSCS) (pp. 160-165). IEEE.

Hespanhol, V., Bárbara, C. 2020. Pneumonia mortality,

comorbidities matter? Pulmonology, 26(3), 123-129.

Huang, G., Liu, Z., et al. 2017. Densely connected

convolutional networks. In Proceedings of the IEEE

conf. on computer vision and pattern recognition (pp.

4700-4708).

Huang, X., Li, P., et al. 2023. Stochastic controlled

averaging for federated learning with communication

compression. arXiv preprint arXiv:2308.08165.

Kareem, A., Liu, H., et al. 2023. A federated learning

framework for pneumonia image detection using

distributed data. Healthcare Analytics, 4, 100204.

Khan, S. H., Alam, M. G. R. 2021. A federated learning

approach to pneumonia detection. In 2021 international

conference on engineering and emerging technologies

(ICEET) (pp. 1-6). IEEE.

Li, T., Sahu, A. K., et al. 2020. Federated learning:

Challenges, methods, and future directions. IEEE

signal processing magazine, 37(3), 50-60.

Liu, Y., Yang, H., & Wu, C. 2023. Unveiling patterns: A

study on semi-supervised classification of strip surface

defects. IEEE Access, 11, 119933-119946.

Mabrouk, A., Redondo, R. P. D., et al. 2023. Ensemble

Federated Learning: An approach for collaborative

pneumonia diagnosis. Applied Soft Computing, 144,

110500.

Mabrouk, A., Diaz Redondo, R. P., et al. 2022. Pneumonia

detection on chest X-ray images using ensemble of

deep convolutional neural networks. Applied Sciences,

12(13), 6448.

Mashoufi, M., Ayatollahi, H., et al. 2023. Data quality in

health care: main concepts and assessment

methodologies. Methods of Information in Medicine,

62(01/02), 005-018.

McMahan, B., Moore, E., et al. 2017. Communication-

efficient learning of deep networks from decentralized

data. In Artificial intelligence and statistics (pp. 1273-

1282). PMLR.

Ngari, M. M., Fegan, G., et al. 2017. Mortality after

inpatient treatment for severe pneumonia in children: a

cohort study. Paediatric and perinatal epidemiology,

31(3), 233-242.

Ni, Q., Sun, Z. Y., et al. 2020. A deep learning approach to

characterize 2019 coronavirus disease (COVID-19)

pneumonia in chest CT images. European radiology,

30, 6517-6527.

Nithya, V. P., Mohanasundaram, N., et al. 2023. An Early

Detection and Classification of Alzheimer's Disease

Framework Based on ResNet-50. Current Medical

Imaging.

Oh, Y., Lee, N., et al. 2022. Communication-efficient

federated learning via quantized compressed sensing.

IEEE Transactions on Wireless Communications,

22(2), 1087-1100.

Qiu, Y., Chang, C. S., et al. 2019. Semantic segmentation

of intracranial hemorrhages in head CT scans. In 2019

IEEE 10th International Conference on Software

Engineering and Service Science (ICSESS) (pp. 112-

115). IEEE.

Qiu, Y., Wang, J., et al. 2022. Pose-guided matching based

on deep learning for assessing quality of action on

rehabilitation training. Biomedical Signal Processing

and Control, 72, 103323.

Sulistya, Y. I., Bangun, E. T. B., et al. 2023. CNN Ensemble

Learning Method for Transfer learning: A Review.

ILKOM Jurnal Ilmiah, 15(1), 45-63.

Wan, C. P., Chen, Q. 2021. Robust federated learning with

attack-adaptive aggregation. arXiv preprint

arXiv:2102.05257.

Wang, H. 2020. Garbage recognition and classification

system based on convolutional neural network vgg16.

In 2020 3rd International Conference on Advanced

Electronic Materials, Computers and Software

Engineering (AEMCSE) (pp. 252-255). IEEE.

Zhang, D., Ren, F., et al. 2021. Pneumonia detection from

chest X-ray images based on convolutional neural

network. Electronics, 10(13), 1512.

Zhao, L., Jiang, J., et al. 2021. Sear: Secure and efficient

aggregation for byzantine-robust federated learning.

IEEE Transactions on Dependable and Secure

Computing, 19(5), 3329-3342.

Zhan, Y., Li, P., et al. 2021. Incentive mechanism design

for federated learning: Challenges and opportunities.

IEEE network, 35(4), 310-317.

Zheng, F., et al. 2021. Accurately discriminating COVID-

19 from viral and bacterial pneumonia according to CT

images via deep learning. Interdisciplinary Sciences:

Computational Life Sciences, 13, 273-285

EMITI 2024 - International Conference on Engineering Management, Information Technology and Intelligence

718