JOURNAL OF LATEX CLASS FILES, VOL. 14, NO.

8, AUGUST 2015 1

Enhancing Collaborative Medical Image Diagnosis

Using Federated Learning: A Case Study from
Shenzhen’s Top Hospitals
Yuhang Li, Chenglong Wang, Lei Xu

Abstract—Federated Learning (FL) represents a transforma- but also fosters trust among healthcare professionals as they
tive approach in collaborative medical image diagnosis by en- engage in the training process. The study highlights the dual
abling the training of high-quality AI models while preserving objectives of improving diagnostic capabilities and ensuring
patient data privacy. This study explores the integration of FL
in healthcare, focusing on data governance, system performance, adherence to stringent data protection regulations, making it a
and challenges such as statistical heterogeneity, communication noteworthy advancement in the field of medical technology.
overhead, and ethical considerations. Key findings reveal FL’s po-
tential to improve diagnostic accuracy across institutions, though II. C HALLENGES IN F EDERATED L EARNING
issues like resource consumption and privacy vulnerabilities Despite its advantages, the deployment of FL in healthcare
require further attention. Recommendations emphasize optimiz-
is not without challenges. Key issues include managing model
ing model architectures, standardizing evaluation metrics, and
fostering international collaborations to enhance model general- and communication heterogeneity, ensuring secure commu-
izability and clinical applicability. Future research opportunities nication, and addressing the legal and ethical implications
include integrating multimodal data and addressing convergence associated with data sharing [5][6]. The case study underscores
dynamics in FL algorithms. By tackling these challenges, FL has the importance of developing robust frameworks that not only
the potential to revolutionize medical diagnostics, ensuring both
enhance model performance but also prioritize patient privacy
innovation and ethical compliance.
and regulatory compliance, paving the way for future research
Index Terms—Federated Learning, Medical Imaging, Data and implementation of FL in medical settings.
Privacy, AI in Healthcare, Statistical Heterogeneity, Model Gen-
eralizability, Ethical AI, Collaborative Diagnostics, Convergence III. T YPES OF F EDERATED L EARNING
Dynamics, Multimodal Data Integration
Federated Learning frameworks can be categorized into
three primary types, especially relevant in healthcare scenarios:
I. I NTRODUCTION
Federated Learning (FL) is a decentralized machine learning A. Cross-Silo Federated Learning
approach that allows multiple institutions to collaboratively In a Cross-Silo setting, a limited number of organizations
train models on sensitive data without sharing the underlying or data centers, such as medical institutions, collaborate to
information. This method is particularly significant in the train a global model using substantial datasets without sharing
healthcare sector, where patient privacy is paramount and sensitive information. This approach is particularly useful
compliance with regulations such as the General Data Pro- in scenarios where organizations face regulatory restrictions
tection Regulation (GDPR) and Health Insurance Portability against data sharing due to privacy and security concerns
and Accountability Act (HIPAA) is crucial [1][2]. The im- [7][8].
plementation of FL in medical image diagnosis has gained
notable attention, particularly through a case study involving B. Cross-Device Federated Learning
top hospitals in Shenzhen, China, where it demonstrated the
Cross-Device Federated Learning focuses on utilizing data
potential to enhance diagnostic accuracy while safeguarding
from numerous edge devices, such as smartphones and tablets,
patient data.
to improve model performance. This variant leverages the
The Shenzhen case study illustrates how FL enables hos-
vast amount of data generated by users while ensuring that
pitals to train artificial intelligence (AI) models using local
individual data points never leave the user’s device, thus
datasets, preserving the unique characteristics of their data
preserving privacy [5].
without the need for data centralization [3][4]. This collab-
orative effort not only enhances the robustness of AI models C. Hierarchical Federated Learning
Y. Li was with the School of Artificial Intelligence and Machine Learning, Hierarchical Federated Learning introduces a layered ap-
Nanjing Medical University, 140 Hanzhong Road, Gulou District, Nanjing, proach to model training, where local models can be trained
Jiangsu Province, 210029China(e-mail: [email protected])
C. Wang was with the School of Artificial Intelligence and Machine at different levels, such as individual devices, local servers,
Learning, Nanjing Medical University, 140 Hanzhong Road, Gulou District, and central servers. This structure allows for effective collab-
Nanjing, Jiangsu Province, 210029China(e-mail: [email protected]) oration among devices with varying computational resources,
L. Xu was with the school of Artificial Intelligence and Machine Learning,
Harbin Institute of Technology, Harbin, Heilongjiang, 130407 China(e-mail: improving the overall performance of the model without
[email protected]) compromising data privacy [2][5].
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 2

throughout the process, attending regular updates and partic-

ipating actively in model training, which fostered a sense of
ownership and trust among healthcare professionals [9].

C. Legal and Ethical Considerations

The study also aimed to ascertain the extent to which FL ap-
plications adhere to GDPR regulations, specifically concerning
the handling of medical data. By emphasizing the significance
of privacy-enhancing technologies, the research sought to
delineate how these innovations influence legal compliance
within medical research contexts [3][9]. A critical aspect of the
initiative was the acknowledgment of hospitals as key players
in AI development, necessitating their involvement in model
creation to ensure practical applicability and compliance with
regulatory frameworks [9].
Fig. 1. Key dimensions of Federated Learning in Healthcare: Data Gov-
ernance, Performance, Challenges, Findings, Recommendations, and Future
Opportunities. V. I MPLEMENTATION OF F EDERATED L EARNING IN THE
C ASE S TUDY
IV. C ASE S TUDY OVERVIEW A. Overview
This case study focuses on the implementation and impact The integration of Federated Learning (FL) into collabora-
of Federated Learning (FL) in collaborative medical image tive medical image diagnosis has been a significant advance-
diagnosis across several top hospitals in Shenzhen. The ini- ment, particularly in settings such as Shenzhen’s top hospitals.
tiative highlights the dual objectives of improving medical This approach enables the development of high-quality AI
diagnostic accuracy while ensuring compliance with strin- models while ensuring patient data privacy and compliance
gent data protection regulations, particularly the General Data with regulations like the General Data Protection Regulation
Protection Regulation (GDPR) and related privacy-enhancing (GDPR) [2][1]. FL allows participating hospitals to train
technologies such as Differential Privacy (DP) and Secure shared predictive models without the need to transfer sensitive
Multi-Party Computation (SMPC) [3][4]. patient data, thus addressing privacy concerns associated with
centralized data storage [10][11].
A. Study Design and Objectives
The study commenced with the identification of relevant
literature and the establishment of research questions aimed at B. Data Governance and Compliance
evaluating the legal implications of FL in medical settings [3]. In the context of the case study, robust data governance
A comprehensive search strategy yielded 6,498 publications, strategies are essential to maintain compliance and protect pa-
of which only 56 (0.86%) were selected for inclusion based tient privacy. Each hospital participating in the FL framework
on predefined criteria [3]. The primary goals were twofold: to must adhere to strict data minimization principles, ensuring
summarize current legal discussions surrounding FL systems only necessary data is utilized for model training [12][13].
in medical research and to identify unresolved legal challenges This decentralized model not only mitigates risks associated
while proposing actionable recommendations for stakeholders with data breaches but also promotes regulatory compliance,
in healthcare [3]. thereby fostering trust among patients and medical institutions
alike [2][14].
B. Implementation of Federated Learning
FL enables participating hospitals to train AI models collab-
C. Integration with Existing Workflows
oratively without sharing sensitive patient data. This ensures
the protection of individual privacy while still facilitating ad- For successful implementation, it is crucial that federated
vancements in diagnostic capabilities [9]. The model training learning techniques are seamlessly integrated into existing
process is initiated from various hospitals’ data, retaining each workflows within the hospitals. This requires the development
institution’s unique data characteristics, thus contributing to a of strategies that facilitate the smooth incorporation of FL
more robust and diverse AI model [4]. into medical practices, allowing healthcare professionals to
As a result of this collaborative approach, the models leverage AI-driven insights without disrupting their routine
developed through FL demonstrated improved reliability and operations [1][10]. By aligning FL initiatives with hospital ob-
robustness, essential qualities for their application in real- jectives, institutions can enhance their diagnostic capabilities,
world medical environments [4]. Hospitals were engaged ultimately leading to improved patient outcomes [2].
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 3

D. Challenges and Solutions E. Evaluation Metrics and Interpretability

The case study highlights several challenges associated with Most studies employed conventional machine learning met-
implementing FL in medical imaging, including high commu- rics for evaluation, such as accuracy, precision, and area under
nication costs and the potential for model biases when train- the receiver operating characteristic curve (AUC). However,
ing datasets lack diversity. Hospitals often face bureaucratic only a limited number specifically tackled unique aspects
hurdles in accessing necessary imaging data due to privacy of FL, including communication overhead and scalability
restrictions, which can impede the model training process [16]. Interpretability was also a concern, with a minority of
[13][15]. To address these issues, the case study advocates for studies exploring feature selection and attention mechanisms
a collaborative framework that allows hospitals to share model to enhance model transparency [6].
updates rather than raw data, thus enabling diverse training
without compromising data security [12][14]. VII. C HALLENGES AND L IMITATIONS
A. Secure Communication
VI. R ESULTS AND F INDINGS
The effectiveness of federated learning (FL) in healthcare is
A. Overview of Analyzed Studies heavily reliant on secure communication protocols. Employing
The analysis incorporated a comprehensive search strategy robust encryption is essential to maintain the integrity of the
that identified 6,498 publications, from which 56 studies FL system, ensuring that sensitive patient data is protected
(0.86%) were selected based on predefined criteria. Among during transmission [16]. However, achieving this level of
these, 55 were articles, and one was a book, with the majority security while maintaining efficiency can pose significant
published in 2020 (36%) and 2021 (43%) [3]. The studies challenges.
predominantly focused on the legal or technical dimensions
of data protection in federated learning (FL), highlighting a B. Flexible Deployment
limitation in the diversity of research perspectives, as only The deployment of FL frameworks in healthcare must be
11% addressed broader issues [3]. streamlined to support secure, reliable, and flexible methods.
A seamless integration with existing IT and data science in-
B. Key Indicators and Measurements frastructures is critical to facilitate a routine and uneventful de-
The research evaluated several critical indicators, including ployment experience. Nonetheless, variability in infrastructure
the external elastic membrane (EEM) area, lumen area, plaque across healthcare institutions can complicate this integration
burden area, plaque burden index, EEM volume, lumen vol- process, potentially leading to inefficiencies [6][7].
ume, and plaque burden volume. The plaque burden index was
particularly significant, as it quantifies the proportion of plaque C. Statistical and System Heterogeneity
burden relative to the EEM area [8][7]. A Bland-Altman Heterogeneity in data is a prominent challenge in FL, partic-
plot was utilized to assess the agreement between manual ularly in healthcare settings where data sources may be non-
measurements and automatic estimations generated by the independent and identically distributed (non-IID). Statistical
global model, thereby enhancing the reliability of diagnostic heterogeneity arises from differences in demographic factors,
classifications [8]. data acquisition protocols, and instrumentation biases among
various institutions. For example, variations in the quality of
C. System Performance and Challenges CT scans can lead to inconsistencies in the correlation between
Federated learning demonstrated a potential for effective imaging data and corresponding electronic health records
diagnosis and classification of plaques through intravascular (EHR), degrading FL performance and requiring additional
ultrasound images, facilitating inter-hospital collaborations [7]. communication rounds for convergence [16].
However, challenges such as statistical and model hetero- Moreover, system heterogeneity, which refers to differences
geneity were prevalent, arising from non-independent and in computational capabilities and resource availability across
identically distributed (non-IID) healthcare data across various clients, complicates FL further. Hospitals with advanced in-
institutions. This heterogeneity was shown to significantly frastructure may achieve better performance than those with
impair FL performance, leading to accuracy drops of up limited resources, affecting the overall efficiency of model
to 50%, necessitating additional communication rounds for aggregation and training [16][7].
model convergence [16].
D. Deficiencies in FL Frameworks
D. Resource Consumption and Efficiency Synchronous FL, where all clients must complete their
Resource consumption was another critical area identified, training and send updates before model aggregation, can be
with only 12% of reviewed studies measuring computational inefficient in real-world scenarios. Delays in data acquisition
costs in FL evaluation. Essential metrics included training or the unavailability of devices can lead to idle times and
time, encryption time, and overall CPU and memory consump- resource under-utilization. Smaller healthcare entities may not
tion [6]. Additionally, privacy and security evaluations were be active during every communication round, resulting in
found to be fundamental yet inadequately addressed, with only delays that degrade the system’s performance. Thus, finding
16% of studies assessing vulnerabilities to attacks like model effective aggregation strategies that can accommodate these
inversion [6]. limitations remains a critical challenge [6].
JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 4

E. Addressing Ethical and Legal Issues [7] L. U. Khan, W. Saad, Z. Han, E. Hossain, and C. S. Hong, “Federated
learning for internet of things: Recent advances, taxonomy, and open
As FL promotes data sharing while maintaining privacy, challenges,” IEEE Communications Surveys & Tutorials, vol. 23, no. 3,
addressing ethical and legal concerns remains a top priority. pp. 1759–1799, 2021.
Researchers must navigate the complexities of health data [8] Y. Han, J. Wei, and A. Yilmaz, “Uas navigation in the real world using
visual observation,” in 2022 IEEE Sensors. IEEE, 2022, pp. 1–4.
governance to develop AI systems that are not only high- [9] T. Yang, X. Yu, M. J. McKeown, Z. J. Wang et al., “When federated
performing but also respect patient rights and confidentiality. learning meets medical image analysis: A systematic review with chal-
Future studies should focus on creating frameworks that bal- lenges and solutions,” APSIPA Transactions on Signal and Information
Processing, vol. 13, no. 1, 2024.
ance innovation with ethical standards, thereby ensuring that [10] M. M. Mello, J. ADLER-MILSTEIN, K. L. Ding, and L. Savage, “Legal
advancements in AI benefit patient care without compromising barriers to the growth of health information exchange—boulders or
privacy [17][9]. pebbles?” The Milbank Quarterly, vol. 96, no. 1, pp. 110–143, 2018.
[11] C.-H. Hsiao, K. Chen, T.-Y. Peng, and W.-C. Huang, “Federated learning
for coronary artery plaque detection in atherosclerosis using ivus imag-
F. Enhancing Collaboration and Generalizability ing: A multi-hospital collaboration,” arXiv preprint arXiv:2412.15307,
2024.
The establishment of large-scale collaborative initiatives, [12] S. Niknam, H. S. Dhillon, and J. H. Reed, “Federated learning for
such as international consortia and multi-center studies, will wireless communications: Motivation, opportunities, and challenges,”
IEEE Communications Magazine, vol. 58, no. 6, pp. 46–51, 2020.
be vital in advancing FL in healthcare. These collaborations [13] M. Joshi, A. Pal, and M. Sankarasubbu, “Federated learning for health-
can help enhance model generalizability across diverse pop- care domain-pipeline, applications and challenges,” ACM Transactions
ulations and clinical environments, addressing the limitations on Computing for Healthcare, vol. 3, no. 4, pp. 1–36, 2022.
[14] J. Xu, B. S. Glicksberg, C. Su, P. Walker, J. Bian, and F. Wang,
often encountered when models are developed within a single “Federated learning for healthcare informatics,” Journal of healthcare
institution [6]. By pooling resources and expertise, researchers informatics research, vol. 5, pp. 1–19, 2021.
can significantly improve the robustness and applicability of [15] T. J. Loftus, M. M. Ruppert, B. Shickel, T. Ozrazgat-Baslanti, J. A.
Balch, P. A. Efron, G. R. Upchurch Jr, P. Rashidi, C. Tignanelli, J. Bian
AI algorithms in medical diagnostics. et al., “Federated learning for preserving data privacy in collaborative
healthcare research,” Digital Health, vol. 8, p. 20552076221134455,
2022.
G. Future Research Opportunities [16] X. Zhang, X. Zhao, Y. Wu, H. Zheng, and Y. Li, “Federated learning for
Future research directions should explore not only the medical image classification: Advances, challenges and opportunities,”
SSRN, 2023. [Online]. Available: https://ssrn.com/abstract=4369198
technical aspects of federated learning but also its practical [17] S. Huang, S. Diao, Y. Wan, and C. Song, “Research on multi-agency
implications in clinical settings. This includes investigating collaboration medical images analysis and classification system based on
the integration of simpler, more interpretable models that can federated learning,” in Proceedings of the 2024 International Conference
on Biomedicine and Intelligent Technology, 2024, pp. 40–44.
operate effectively in real-world healthcare applications while [18] T. Li, A. K. Sahu, A. Talwalkar, and V. Smith, “Federated learning:
providing transparency in decision-making processes [18][6]. Challenges, methods, and future directions,” IEEE signal processing
Additionally, a rigorous analysis of convergence dynamics in magazine, vol. 37, no. 3, pp. 50–60, 2020.
[19] F. R. da Silva, R. Camacho, and J. M. R. Tavares, “Federated learning
federated optimization is needed to understand the perfor- in medical image analysis: A systematic survey,” Electronics, vol. 13,
mance of FL algorithms across various healthcare scenarios, no. 1, p. 47, 2023.
facilitating the development of more efficient training strate-
gies [19].
Through these initiatives, federated learning has the poten-
tial to transform medical image diagnosis, leading to enhanced
patient care and more efficient clinical workflows.

R EFERENCES
[1] E. Albalawi, M. TR, A. Thakur, V. V. Kumar, M. Gupta, S. B. Khan,
and A. Almusharraf, “Integrated approach of federated learning with
transfer learning for classification and diagnosis of brain tumor,” BMC
Medical Imaging, vol. 24, no. 1, p. 110, 2024.
[2] D. Solans, M. Heikkila, A. Vitaletti, N. Kourtellis, A. Anagnostopoulos,
I. Chatzigiannakis et al., “Non-iid data in federated learning: A system-
atic review with taxonomy, metrics, methods, frameworks and future
directions,” arXiv preprint arXiv:2411.12377, 2024.
[3] Y. Wu, C. Desrosiers, and A. Chaddad, “Facmic: Federated adaptative
clip model for medical image classification,” in International Conference
on Medical Image Computing and Computer-Assisted Intervention.
Springer, 2024, pp. 531–541.
[4] N. Mouhni, A. Elkalay, M. Chakraoui, A. Abdali, A. Ammoumou, and
I. Amalou, “Federated learning for medical imaging: An updated state
of the art,” Ing. Syst. D’Inf, vol. 27, pp. 143–150, 2022.
[5] I. Alekseenko, A. Karargyris, and N. Padoy, “Distance-aware non-iid
federated learning for generalization and personalization in medical
imaging segmentation,” in MIDL 2024, Medical Imaging with Deep
Learning, Paris, France, 03-05 juillet 2024, 2024.
[6] N. Koutsoubis, A. Waqas, Y. Yilmaz, R. P. Ramachandran, M. Scha-
bath, and G. Rasool, “Future-proofing medical imaging with privacy-
preserving federated learning and uncertainty quantification: A review,”
arXiv preprint arXiv:2409.16340, 2024.