Advancements and Applications of Using Federated Learning in

Diagnosing and Analyzing Brain Tumor Images

Yusong Yang

Data Science and Big Data Technology, Xi'an Jiaotong-Liverpool University, Suzhou, Jiangsu, China

Keywords: Federated Learning, Brain Tumor, Mathematical Model.

Abstract: Brain tumors represent a major health concern, and conventional diagnostic approaches can be intricate and

prone to errors, especially given the diversity of tumor types. The swift progress in Artificial Intelligence (AI)

has emerged as a promising avenue for enhancing brain tumor diagnosis. Federated Learning (FL), which is

a distributed machine learning approach, facilitates collaborative model training among various institutions,

improving diagnostic precision while safeguarding data privacy. This study introduces a federated learning

framework for classifying brain tumors using Convolutional Neural Networks (CNNs), specifically

leveraging an optimized Visual Geometry Group 16 (VGG16) architecture alongside transfer learning

methods. The model was trained across several clients, achieving an outstanding classification accuracy of

98%. Furthermore, U-Net was utilized for segmenting brain tumors, demonstrating notable performance

enhancements with an increasing number of participating clients. Despite the evident advantages of FL

regarding privacy preservation and model efficacy, challenges such as differences in institutional equipment

and the Non-independent and Identically Distributed (non-IID) characteristics of data impede generalization

and convergence of models. To address these challenges, this paper suggests employing adaptive algorithms

and data augmentation strategies to improve model flexibility and effectiveness. Additionally, effectively

merging multimodal data remains a significant technical challenge that needs resolution in future work.

1 INTRODUCTION

A brain tumor is characterized as an abnormal

proliferation or mass of cells located in or around the

human brain (DeAngelis et al., 2001). In this country,

both the incidence and mortality rates associated with

brain tumors are notably high, with an annual

occurrence rate of 7 per 100,000 individuals. Brain

tumors can impact people of all ages, and genetic

conditions such as neurofibromatosis and Turner

syndrome may elevate the risk. Patients often present

symptoms including paralysis and psychological

disturbances. The five-year relative survival rate for

those diagnosed with malignant brain tumors is

approximately 36%, while glioblastoma—the most

common primary malignant type—has a significantly

lower five-year survival rate of just 7.2%

(International Neurosurgery Group, INC., 2021).

Traditional diagnostic methods for brain tumors tend

to be complex and inefficient; due to the variety of

tumor types, there exists a notable misdiagnosis

frequency. Consequently, there is a pressing need for

https://orcid.org/0009-0001-3780-9837

improved diagnostic approaches and adjunct

treatments.

Artificial Intelligence (AI), known for its robust

feature extraction capabilities and predictive power,

presents a potential solution to enhance diagnosis

rates. AI has been utilized across various fields—

including chemical biology—and particularly within

medicine where numerous studies have emerged.

Reham noted that processing large volumes of images

can be time-consuming and labor-intensive; thus,

automated segmentation and classification techniques

are essential to expedite the diagnosis process for

brain tumors. Imaging modalities such as Computed

Tomography (CT) and Magnetic Resonance Imaging

(MRI) facilitate rapid detection.

For instance, Akmalbek et al. developed a model

aimed at automating the detection of brain tumors

effectively from MRI scans by leveraging advanced

object detection features inherent in You Only Look

Once Version 5 (YOLOv5) architecture; their model

demonstrates exceptional accuracy in identifying

tumor regions critical for early diagnosis and timely

Yang and Y.

Advancements and Applications of Using Federated Learning in Diagnosing and Analyzing Brain Tumor Images.

DOI: 10.5220/0013528300004619

In Proceedings of the 2nd International Conference on Data Analysis and Machine Learning (DAML 2024), pages 557-561

ISBN: 978-989-758-754-2

557

medical intervention (Abdusalomov et al., 2024).

Furthermore, Liu et al. indicated that while AI

systems in medical imaging show significant promise

in aiding diagnoses and treatment processes,

challenges like data silos related to medical images,

privacy concerns regarding data security, along with

inconsistent industry standards severely limit their

effectiveness. By integrating federated learning

principles alongside FAIR scientific data

management guidelines, these issues could be

addressed technically, maximizing multi-center data

utility to develop more efficient disease diagnosis

systems while guiding AI technology implementation

within clinical settings (Liu et al., 2022).

Given the significance surrounding brain tumors

coupled with recent advancements in artificial

intelligence necessitates a thorough review on how AI

applications can aid in diagnosing these conditions

effectively. The subsequent sections will outline

previous research efforts focused on techniques

designed to improve recognition of brain tumor

images before discussing existing challenges faced

when employing AI methodologies within this

domain. Finally, the paper will conclude by

summarizing key findings derived from

investigations into using artificial intelligence for

diagnosing brain tumors.

2 METHOD

2.1 Introduction of Federated Learning

Federated Learning (FL) is a novel method in

distributed machine learning that allows various

devices or organizations to collaboratively develop a

model while maintaining the confidentiality of their

raw data. In this approach, each participant performs

training on their local datasets and subsequently

transmits only the updated model parameters to a

central server. The server then integrates these

parameters to create an overarching global model.

This strategy is especially beneficial in sensitive

fields such as healthcare and finance, where privacy

and security are paramount.

The primary characteristics of federated learning

include: protecting data privacy (since data remains

stored on local devices), accommodating

heterogeneity (recognizing that different nodes may

possess varying data distributions and processing

capabilities), improving communication efficiency

(by sending only model parameters), and ensuring

fault tolerance (considering potential node instability

or dropout). Essential elements of this framework

involve local model training, aggregation of the

global model, effective communication strategies,

and privacy-preserving techniques like differential

privacy and secure multiparty computation.

2.2 Federated Learning-based CNN

MRI imaging plays a crucial role in identifying brain

tumors, a task essential for developing targeted

treatment strategies. The complexity of tumor shapes

and the differences in imaging make accurate

diagnosis challenging. Typically, this process

depends on manually analyzing MRI scans and using

basic machine learning methods. Yet, these

traditional techniques often fall short in consistently

and automatically categorizing tumors due to several

drawbacks. These include the need for extensive

manual input, the risk of human mistakes, difficulties

in managing vast amounts of data, and limited

flexibility in accommodating different kinds of

tumors and imaging scenarios.

The authors introduced a deep learning model that

combines federated learning with advanced

Convolutional Neural Networks (CNNs) to achieve

precise and automated classification of brain tumors.

This model improves upon the Visual Geometry

Group 16 (VGG16) architecture, tailoring it

specifically for brain MRI images. It emphasizes the

dual benefits of federated learning and transfer

learning in medical imaging. Federated learning

allows for the decentralized training of models across

various clients while maintaining the privacy of

medical data, which is vital in handling sensitive

health information.

The architecture utilizes transfer learning by

employing pre-trained CNNs, greatly enhancing its

accuracy in brain tumor classification by leveraging

knowledge from diverse datasets. The model

underwent training with a mixed dataset that included

Figshare, SARTAJ, and Br35H, using federated

learning to facilitate decentralized training while

maintaining privacy. Additionally, the application of

transfer learning improved the model's ability to

handle the complex variations in MRI images

associated with different types of brain tumors,

boosting its overall performance.

The model achieved high precision rates (glioma:

0.99; meningioma: 0.95; no tumor: 1.00; pituitary:

0.98), along with impressive recall rates and F1

scores during classification—outperforming existing

methodologies overall with an accuracy rate reaching

98%. This demonstrates the model’s ability to

accurately classify various tumor types while

emphasizing how federated learning combined with

transfer learning can transform brain tumor

classification using MRI imagery (Albalawi et al.,

2024).

DAML 2024 - International Conference on Data Analysis and Machine Learning

558

2.3 Federated Learning-based U-net

Segmenting brain tumors in medical imaging is

essential for precise diagnosis and treatment

planning, but it presents notable challenges in

maintaining patient data privacy and security.

Traditional centralized approaches struggle with data

sharing constraints imposed by privacy laws and

security risks, which can stymie the progress of

advanced AI technologies in this area. This study

addresses these challenges by introducing a federated

learning framework that supports collaborative

training of models on distributed datasets across

multiple medical institutions, without needing direct

access to the original data. This method ensures

privacy and enhances security while advancing AI

applications in brain tumor segmentation.

The research utilizes the Universal Network (U-

Net) architecture, known for its effectiveness in

semantic segmentation tasks, emphasizing its

scalability for applications within medical imaging.

Experimental findings indicate that federated

learning significantly enhances performance metrics:

specificity improved from 0.92 to 0.96, while the

Dice coefficient rose from 0.84 to 0.89 as the number

of participating clients increased from 50 to 100.

Furthermore, this proposed method outperforms

existing techniques based on CNNs and Recurrent

Neural Networks (RNNs), achieving greater accuracy

and efficiency overall. This work contributes to

advancing segmentation techniques in medical

imaging while maintaining essential standards of data

privacy and security (Ullah et al., 2023).

2.4 FL-PedBrain

Lee et al. introduced an FL system named FL-

PedBrain, specifically aimed at addressing pediatric

pilocytic astrocytomas (PF tumors) in children. While

brain tumors are the most common solid tumors

found in childhood, they remain relatively rare and

scattered across various pediatric and subspecialty

centers. Thus, creating a collaborative platform that

enables large-scale AI training among different

institutions can greatly benefit this patient

demographic.

In their research, the authors utilized an extensive

and diverse dataset of brain MRI scans gathered from

19 institutions globally, concentrating on pediatric PF

tumors. They designed and assessed an FL

framework for joint tumor pathology prediction and

segmentation tailored to this data-limited population.

The results revealed that FL-PedBrain exhibited

strong generalization capabilities across all

participating sites, including three external validation

cohorts.

When comparing federated learning with

Centralized Data Sharing (CDS), which pooled data

from all locations, FL showed a classification bias of

less than 1.5% and only a 3% bias in segmentation

performance. Although there was no statistically

significant difference in classification outcomes

between CDS and FL, the latter demonstrated slightly

lower segmentation performance for two out of four

tumor categories. In contrast, models trained

exclusively on local datasets—referred to as isolated

training—performed approximately 20% worse than

both FL and CDS approaches. This finding highlights

the challenges associated with AI generalization as

well as the vulnerabilities of models developed in

isolation (Lee et al., 2024).

2.5 MQTT

Intelligent healthcare utilizes AI to analyze and learn

from patient data. Due to the rarity of large and

diverse datasets for training Machine Learning (ML)

models within a single medical facility, traditional

centralized AI methods necessitate the transfer of

privacy-sensitive information from healthcare

institutions to data centers for processing and

integration. This movement of data not only requires

substantial communication resources and energy but

also raises significant privacy concerns, creating

challenges for international clinical research

collaborations.

FL has emerged as a distributed AI strategy that

enables collaborative training of ML models without

requiring the sharing of patient information. This

paper provides an in-depth analysis of various

federated learning techniques and introduces a real-

time distributed network framework utilizing the

Message Queuing Telemetry Transport (MQTT)

protocol. The authors specifically developed several

network-based ML solutions using FL tools, which

depend on parameter servers (PS) as well as fully

decentralized paradigms driven by consensus

mechanisms.

The proposed approach was tested on brain tumor

segmentation tasks using a modified U-Net model,

which incorporated a clinical dataset typical of

routine clinical workflows. The FL process was

executed on physically separate machines situated in

various countries, with these machines

communicating over the Internet. This setup allowed

for effective validation while adhering to the

requirements of data privacy and security in a

distributed computing environment. A real-time

testbed was utilized to evaluate the balance between

training accuracy and latency, emphasizing critical

operational conditions that influence performance

during actual deployment (Tedeschini et al., 2022).

Advancements and Applications of Using Federated Learning in Diagnosing and Analyzing Brain Tumor Images

559

2.6 LEAF

FL presents an opportunity to create customized AI

services for hospital networks, aiding in the reduction

of overfitting and improving the robustness of

models. However, integrating FL introduces

significant challenges, particularly regarding user

privacy, which remains a major barrier to its practical

implementation. Many current solutions utilize

blockchain technology to mitigate these privacy

issues. While blockchain can prevent external

systems from interfering with decision-making

processes, it still permits network devices to access

shared data. Additionally, adopting blockchain

requires a new framework and infrastructure,

resulting in further indirect costs.

To tackle these challenges, a Limited Access

Encryption Algorithm Framework (LEAF)

framework has been proposed that merges Federated

Learning with a limited access encryption algorithm.

This cryptographic approach employs edge AI

models to effectively address privacy concerns while

aiming to safeguard user confidentiality and reduce

indirect expenses. The authors assessed the

performance of the LEAF framework through

comprehensive simulations and obtained encouraging

results. Notably, the precision achieved by this

framework is 3% higher than that of traditional

centralized systems as well as FL-based approaches

without compromising user privacy. In optimal

conditions, the encryption process within this

proposed framework can decrease data size by four to

five times (Patel et al., 2024).

3 DISCUSSIONS

3.1 Limitations and Challenges

3.1.1 Generalizability

Model adaptability: Different hospitals or research

centers may use different types of equipment,

imaging techniques, and data formats. Even if

federated learning can share models across locations,

these device or imaging differences can cause the

model to perform poorly in certain scenarios. For

example, the quality, resolution, or imaging

conditions of MRI scans may affect the model's

generalizability, leading to some hospitals' models

performing better than others.

Data heterogeneity: Brain tumor types, sizes, and

shapes vary, and different patient populations may

have significant differences. This heterogeneity

presents a challenge to the model's generalization,

and the federated learning global model may have

difficulty simultaneously adapting to all different

brain tumor features.

3.1.2 Data Discrepancies

Non-IID data: One of the major challenges of

federated learning is that the data contributed by

participants is often not Independent and Identically

Distributed (Non-IID). In brain tumor prediction,

some hospitals may receive more rare case types,

leading to differences in data distribution. These

differences can affect the convergence of the global

model, causing it to overfit certain datasets or perform

poorly on certain datasets.

Data quality and annotation: Medical imaging

data typically requires professional annotation, and

the accuracy and consistency of annotation may vary

between hospitals. Federated learning relies on local

annotation quality, and if certain institutions'

annotations are incorrect or inconsistent, it may affect

the performance of the global model.

3.1.3 Multimodal Learning

Complexity of Fusing Information from Different

Data Modalities: Brain tumor prediction may not only

depend on MRI or CT images, but also requires

combining patient's clinical information (such as age,

genetic data, etc.). In federated learning of multi-

modal data, how to effectively fuse information from

different modalities is a critical issue. The feature

dimensions and distribution of data from different

modalities are quite different, how to fully utilize this

information without exchanging data remains a

technical challenge.

Synchronization Problem of Multi-modal Data:

The data sources, storage systems, and collection

frequencies in different hospitals may all differ. To

achieve multi-modal federated learning, it is

necessary to ensure that the data from each hospital

can be synchronized in time and content, which will

face considerable technical difficulties in practice.

3.2 Future Prospects

3.2.1 Improving Scalability

To improve the scalability of federated learning

models and address the challenge of data

heterogeneity, strategies such as multi-center

collaboration and transfer learning can be employed.

By promoting collaboration among multiple medical

centers, the diversity and scale of the data increase,

enhancing model robustness and generalizability

across different populations. In domains with limited

labeled data, transfer learning can transfer knowledge

from data-rich areas to improve model performance

DAML 2024 - International Conference on Data Analysis and Machine Learning

560

in new environments. Additionally, personalized

models can be developed to adapt to the data

characteristics of individual users, with ensemble

learning methods combining predictions from

multiple models to boost overall performance.

Aligning data distributions through augmentation and

preprocessing techniques is also crucial for reducing

bias caused by non-IID data, thus improving the

model's generalization and fairness.

3.2.2 Addressing Data Heterogeneity (non-

IID) Issues

Adaptive algorithms: Develop adaptive algorithms

that dynamically adjust model parameters based on

the distribution of data, improving the model's ability

to adapt to non-IID data.

Label balancing and sampling strategies:

Introduce label balancing and effective sampling

strategies to ensure that the model has access to data

samples from different categories during training,

thereby improving overall performance.

3.2.3 Integration of Multimodal Learning

Multimodal Data Fusion: Establish cross-modal

learning frameworks to utilize complementary

information from different modalities (such as

imaging and genomic data) to improve prediction

accuracy.

Cross-modal Knowledge Distillation: Extract

knowledge from one modality and transfer it to

another modality to enhance the model's

generalization ability.

4 CONCLUSIONS

This article provides a comprehensive overview of

how federated learning can be used in the field of

brain tumor detection and analysis. It describes how

methods such as CNN and U-NET can be used to

detect and analyze brain tumors, as well as the

accuracy and adaptability of these methods. The

article also highlights the limitations of federated

learning in terms of its lack of scientific rigor and

limited applicability, and suggests that efforts should

be made in the future to improve its generalizability

and scientific validity in order to better integrate it

into the detection, analysis, and treatment of brain

tumors.

REFERENCES

Abdusalomov, A., Rakhimov, M., Karimberdiyev, J.,

Belalova, G., & Cho, Y. I. 2024. Enhancing automated

brain tumor detection accuracy using artificial

intelligence approaches for healthcare environments.

Bioengineering, 11(6), 627.

Albalawi, E. T. R., Thakur, M., Kumar, A., Gupta, V. V.,

Khan, M., Bhatia, S., & Almusharraf, A. 2024.

Integrated approach of federated learning with transfer

learning for classification and diagnosis of brain tumor.

BMC Medical Imaging, 24(1).

DeAngelis, L. M. 2001. Brain tumors. New England

journal of medicine, 344(2), 114-123.

International Neurosurgery Group, INC. 2021. How many

brain tumors are there? Here’s a quick primer on brain

tumors. https://m.incsg.com/naoliu/1333.html

Lee, E. H., Han, M., Wright, J., Kuwabara, M., Mevorach,

J., Fu, G., ... & Yeom, K. W. 2024. An international

study presenting a federated learning AI platform for

pediatric brain tumors. Nature communications, 15(1),

7615.

Liu, Z., Shi, Z., & Liang, C. 2022. Advance the application

of federated learning technology in medical imaging

artificial intelligence. National Medical Journal of

China, 5, 318-320.

Patel, N. P., Parekh, R., Amin, S. A., Gupta, R., Tanwar, S.,

Kumar, N., Iqbal, R., & Sharma, R. 2024. LEAF: A

federated learning-aware privacy-preserving

framework for healthcare ecosystem. IEEE

Transactions on Network and Service Management,

21(1), 1129-1141.

Reham, K. 2023. A review of recent advances in brain

tumor diagnosis based on AI-based classification.

Diagnostics, 13(18), 3007.

Tedeschini, B. C., Savazzi, S., Stoklasa, R., Barbieri, L.,

Stathopoulos, I., Nicoli, M., & Serio, L. 2022.

Decentralized federated learning for healthcare

networks: A case study on tumor segmentation. IEEE

Access, 10, 8693-8708.

Ullah, F., Nadeem, M., Abrar, M., Amin, M., Salam, F.,

Khan, A., & Salabat. 2023. Enhancing brain tumor

segmentation accuracy through scalable federated

learning with advanced data privacy and security

measures. Mathematics, 11(19), 4189.

Advancements and Applications of Using Federated Learning in Diagnosing and Analyzing Brain Tumor Images

561