PaperlayerView exact revision on arXiv

Source-linked AI summary

Federated Learning for Medical Applications: A Taxonomy, Current Trends, Challenges, and Future Research Directions

Ashish Rauniyar, Desta Haileselassie Hagos, Debesh Jha, Jan Erik Håkegård, Ulas Bagci, Danda B. Rawat, Vladimir Vlassov

arXiv:2208.03392v5cs.LGcs.CRcs.CV

TL;DR

Medical AI must improve diagnosis while meeting stringent security, privacy, and QoS requirements when sensitive data cannot be freely shared. This survey synthesizes FL for medical applications, especially global cancer diagnosis, reviewing frameworks, emerging technologies, use cases, challenges, and future directions. It reports evidence that FL-based medical models can outperform conventional or individually trained approaches, while emphasizing unresolved deployment, privacy, regulatory, and statistical challenges.

  • Problem

    Medical AI adoption is constrained by security, privacy, QoS, and data-sharing challenges surrounding sensitive healthcare information.

  • Method

    The paper surveys medical FL frameworks, emerging-technology integrations, diagnostic use cases, open-source software, challenges, and research directions.

  • Results

    FL-based medical models showed reported gains including 95.41% accuracy in FedHome, 33% higher glioblastoma detection accuracy, and 6.3% average improvement over individually trained breast-density models.

  • Takeaways & Limitations

    FL can support computer-aided diagnosis and collaborative medical model development without requiring direct access to sensitive patient data.

Abstract

from arXiv · show

With the advent of the IoT, AI, ML, and DL algorithms, the landscape of data-driven medical applications has emerged as a promising avenue for designing robust and scalable diagnostic and prognostic models from medical data. This has gained a lot of attention from both academia and industry, leading to significant improvements in healthcare quality. However, the adoption of AI-driven medical applications still faces tough challenges, including meeting security, privacy, and quality of service (QoS) standards. Recent developments in \ac{FL} have made it possible to train complex machine-learned models in a distributed manner and have become an active research domain, particularly processing the medical data at the edge of the network in a decentralized way to preserve privacy and address security concerns. To this end, in this paper, we explore the present and future of FL technology in medical applications where data sharing is a significant challenge. We delve into the current research trends and their outcomes, unravelling the complexities of designing reliable and scalable \ac{FL} models. Our paper outlines the fundamental statistical issues in FL, tackles device-related problems, addresses security challenges, and navigates the complexity of privacy concerns, all while highlighting its transformative potential in the medical field. Our study primarily focuses on medical applications of \ac{FL}, particularly in the context of global cancer diagnosis. We highlight the potential of FL to enable computer-aided diagnosis tools that address this challenge with greater effectiveness than traditional data-driven methods. We hope that this comprehensive review will serve as a checkpoint for the field, summarizing the current state-of-the-art and identifying open problems and future research directions.

I. INTRODUCTION

AI-driven medical applications improve healthcare but face security, privacy, and QoS challenges because sensitive data are difficult to share and centralize. This survey examines FL for medical applications, emphasizing emerging-technology integration, global cancer diagnosis, frameworks, challenges, and future directions.

  • Motivation: Centralized medical-data processing raises privacy and practical concerns, while edge-based distribution responds to real-time, latency, and privacy requirements.
  • Motivation: FL enables collaborative training across healthcare datasets without sharing sensitive patient data, addressing privacy concerns through decentralized processing.
  • Scope and Contributions: The survey integrates FL with emerging technologies and examines its medical applications, including healthcare frameworks, practical use cases, and global cancer diagnosis.
  • Scope and Contributions: Its cancer focus investigates whether FL can support computer-aided diagnostic tools with superior performance and generalizability compared with other data-driven healthcare models.
  • Scope and Contributions: The paper compares open-source FL software frameworks and identifies medical-domain challenges, open problems, and future research directions.

B. Paper Organizations

The paper progresses from FL concepts and medical frameworks to emerging-technology integrations, practical diagnostic applications, software frameworks, challenges, and future research directions.

  • Organization: Section II introduces FL concepts and medical-domain frameworks.
  • Organization: Later sections cover emerging technologies, practical medical applications, cancer-focused diagnostic tools, open-source software, challenges, and future directions.

II. FL CONCEPTS AND FRAMEWORKS IN THE MEDICAL DOMAIN

Medical FL keeps healthcare data at local organizations or devices while exchanging model updates with a coordinating server. This supports privacy-preserving collaboration but must accommodate heterogeneous data, devices, and regulatory requirements.

  • Advantages: Healthcare FL offers data sovereignty, reduced transfer, scalability, energy efficiency, and data localization, while security and privacy remain central concerns.
  • Privacy and Security: FL aggregates updates or gradients rather than raw healthcare data, while differential privacy, anonymization, and secure communication provide additional protections.
  • Classical Architecture: In classical FL, selected clients train on localized data, send model parameters to a central server, and receive the aggregated global model.
  • Algorithms: FedAvg iteratively aggregates client parameters until convergence or a specified iteration count, but lacks theoretical performance guarantees for non-iid data.
  • Medical Framework: Medical FL clients may include mobile, hospital, and home healthcare systems with differing data distributions and device constraints.
  • Aggregation: Aggregation combines decentralized updates into a global model intended to improve accuracy, generalizability, collaboration, and local customization.

A. Horizontal FL Framework

Horizontal FL applies when participating medical institutions share the same feature space but hold different patient samples. It enables joint model training without exchanging individual patient records.

  • Horizontal FL Framework: Horizontal FL connects institutions with identical features, such as age, gender, cholesterol, and blood pressure, across distinct patient populations.
  • Horizontal FL Framework: A shared model learns from distributed patient samples, helping address limited local sample sizes without sharing individual patient records.

B. Vertical FL Framework

Vertical FL enables collaboration across differing feature spaces, while blockchain-based FL supports decentralized, privacy-preserving healthcare learning and data management. The surveyed examples address personalization, security, interoperability, consent, and clinical applications.

  • Vertical FL: Vertical FL trains a global model from varied feature spaces across diverse datasets, such as hospital treatment histories and insurer claims.
  • Federated transfer learning: Federated transfer learning extends vertical FL with a pre-trained model to create customized health models when samples and feature spaces differ.
  • Federated transfer learning: FedHealth uses federated transfer learning to personalize organizational models, transferring human activity recognition knowledge to Parkinson’s disease classification.
  • Blockchain-based FL: Blockchain complements FL by providing decentralized coordination, secure health-data storage, auditable model updates, and privacy-preserving deployment.
  • Applications: Blockchain-FL architectures support secure collaboration across healthcare institutions, IoMT agents, clinical trials, precision medicine, and disease classification.
  • Healthcare integration: The surveyed blockchain-FL directions include interoperability, consent management, data monetization, research advancement, and regulatory compliance.

B. FL Enabled IoT-based Healthcare Monitoring

IoT and edge computing provide distributed healthcare data and computation, while FL enables collaborative analysis without centralizing sensitive patient information. Reported systems target real-time monitoring, personalization, efficiency, security, and diagnosis.

  • IoT-enabled healthcare: IoT and IoMT generate healthcare data streams from distributed devices, while FL evaluates data locally at edge devices to support privacy.
  • IoT-FL benefits: FL combined with IoT supports real-time processing, remote monitoring, early diagnosis, cost reduction, scalability, and patient engagement without sharing raw data.
  • Edge computing: Edge computing moves storage and computation closer to healthcare data sources, improving latency, resilience, availability, and bandwidth use.
  • Healthcare systems: FedHome integrates cloud and edge resources for personalized in-home monitoring and achieves 95.41% accuracy, exceeding conventional neural networks by over 7.49%.It uses class-balanced user data and a lightweight GCAE to reduce communication costs.
  • Healthcare systems: A fog-computing prototype demonstrates fourfold faster computation for mobile patients, while BodyEdge combines mobile clients, edge gateways, and multi-radio connectivity.
  • Edge-assisted FL: Edge-assisted FL avoids transporting distributed data to a central cloud and offers a privacy-respecting, cloud-resource-efficient solution.
  • Security and privacy: A lightweight edge privacy protocol protects gradient privacy while resisting equipment dropouts and collusion attacks through shared secrets, weight masks, signatures, and hashes.
  • Edge-FL capabilities: Edge-FL healthcare architectures address device diversity, local security, privacy-preserving training, and edge-based diagnosis and decision support.

IV. PRACTICAL APPLICATION OF FL IN MEDICAL DOMAIN

FL combines machine learning with the sensitivity of medical data, allowing institutions to collaborate on healthcare insights without compromising patient-record privacy and security.

  • FL enables medical institutions to collaborate and share insights and expertise without compromising the privacy and security of patient records.

A. FL-enabled COVID-19 Detection

The paper surveys federated learning applications for medical diagnosis, emphasizing collaborative analysis across institutions without sharing sensitive patient data. Reported use cases span COVID-19, cancer detection and segmentation, while generalization and dataset diversity remain important considerations.

  • COVID-19 detection: FL enables medical institutions to screen COVID-19 from chest X-ray images without sharing patient data.The review reports numerous COVID-19 X-ray training and deployment studies, alongside EXAM’s use of data from 20 global institutes to estimate oxygen requirements.
  • Cross-center learning: FL supports collaborative medical image analysis by aggregating models trained on distributed, heterogeneous datasets across multiple centers.Institutions can benefit from one another’s annotations without sharing them, although distributed training requires substantial computational power and memory.
  • Cancer detection: 33% higher accuracy than traditional methods was reported for FL-based glioblastoma detection using MRI data from 71 sites across 6 continents.The data remained within individual institutions, supporting development of a more robust and generalizable model for worldwide diagnosis.
  • Cancer detection: 6.3% average improvement over individually trained models and 45.8% relative generalizability improvement were reported for FL-based breast density classification across seven institutions.The review attributes these improvements to larger multi-institutional training datasets without data sharing.
  • Cancer segmentation: A federated brain-tumor segmentation model achieved a Dice coefficient of 0.8520, compared with 0.8620 for a model trained by sharing the dataset.The study used 2018 BraTS data because a centralized public dataset was unavailable.
  • Generalization: FL research addresses cross-hospital performance drops through methods such as Federated Domain Generalization and Episodic Learning in Continuous Frequency Space.These approaches target generalization to unseen hospitals and data outside the federation, including prostate MRI segmentation across six medical centers.

V. OPEN-SOURCE FL SOFTWARE FRAMEWORKS

This section surveys open-source FL software frameworks as tools for prototyping, experimentation, and validation in healthcare. It emphasizes transparency and contrasts open-source tools with proprietary frameworks whose documentation may hinder exploration.

  • Open-source FL frameworks support rapid prototyping, experimentation, and validation of healthcare algorithms.
  • Open-source software promotes transparency by allowing healthcare institutions to inspect code, verify security measures, and understand implemented algorithms.
  • Proprietary frameworks may hinder exploration of new FL methodologies because their documentation and implementation details are not comprehensive.

A. Federated AI Technology Enabler Framework (FATE)

The surveyed frameworks provide different capabilities for secure computation, decentralized machine learning, medical FL, biomedical deployment, and benchmarking. Their applicability is shaped by both functionality and practical limitations.

  • FATE: FATE provides secure computation protocols for collaborative data analysis under data protection regulations through a script-based interface.Its interface relies on command-line options and a specialized, underspecified domain-specific language.
  • TFF: TFF offers separate FL and Federated Core APIs for ML development, systems research, protocol simulation, and TensorFlow model compatibility.The framework lacks dedicated differential-privacy mechanisms.
  • OpenFL: OpenFL supports medical FL training while interoperating with TensorFlow, PyTorch, and other machine-learning and deep-learning pipelines.
  • Fed-BioMed: Fed-BioMed enables biomedical research through decentralized statistical analysis and machine-learning development using Python, PyTorch, and Scikit-learn.The initiative still requires development to become scalable and fully deployable in practical healthcare settings.
  • Table IV compares open-source FL software frameworks.

E. PySyft

The paper describes PySyft alongside other open-source FL frameworks and connects framework selection with healthcare privacy, communication, scalability, and reliability requirements. It also identifies OpenFL and Fed-BioMed as particularly suitable for medical applications.

  • PySyft: PySyft combines federated learning, differential privacy, and encrypted computations for research-oriented FL.It primarily supports deep-learning frameworks such as PyTorch and TensorFlow.
  • PySyft: PySyft cannot currently support projects requiring network communication without the additional PyGrid package.
  • Other frameworks: FedML supports on-device, distributed, and single-machine FL experiments, including Edge AI deployments on GPUs, smartphones, and IoT devices.
  • Framework suitability: The survey identifies OpenFL and Fed-BioMed as particularly suitable frameworks for medical applications because of their alignment with medical data and privacy needs.
  • Healthcare requirements: Healthcare FL faces communication constraints from bandwidth demands, geographically unstable connectivity, latency, and stringent transmission privacy requirements.These constraints are especially relevant for dispersed providers and time-sensitive applications.

B. Statistical Challenges

Healthcare FL must handle heterogeneous, biased, sparse, and non-I.I.D. data while accommodating unequal device capabilities and institutional data distributions. These statistical and computational conditions complicate optimization, generalization, efficiency, and scalability.

  • Data heterogeneity: Healthcare data heterogeneity arises across EHRs, imaging, wearables, genomics, formats, coding systems, and collection practices.
  • Bias and generalizability: Single-site training can introduce bias and limit generalizability because patient populations, disease prevalence, and treatment protocols vary across facilities.
  • Data sparsity: Data sparsity and limited labeling at smaller or remote clinics can hinder participation and reduce the usability of local medical data.
  • Non-I.I.D. distributions: Non-I.I.D. client distributions complicate federated optimization, increase straggler risk, and add modeling, analysis, and evaluation complexity.
  • Computational challenges: More than 80% of edge-device-generated data is predicted to be processed at the network edge by 2030.This trend increases interest in Edge-AI and makes device computation central to healthcare FL.
  • Device constraints: Healthcare FL requires lightweight, energy-efficient models because participating devices differ in computational capacity, storage, energy profiles, and reliability.
  • Cross-silo FL: Cross-silo data disparity may require more FL iterations to reach target global-model accuracy, increasing computation time and reducing responsiveness.
  • Scalability: FL training and deployment can be computationally expensive, making scalability difficult for resource-limited healthcare organizations.

D. Security and Privacy Challenges

Healthcare FL must address security, privacy, communication, statistical, and computational challenges before practical deployment. Key concerns include attacks on shared updates, heterogeneous medical data, participation incentives, and unresolved optimization issues.

  • Security and Privacy: FL healthcare systems can expose sensitive information through intercepted gradient updates and malicious client model updates.These threats include reverse engineering and backdoor attacks against the global model, despite raw data remaining local.
  • Future Directions: Future FL healthcare research must establish practical frameworks that address communication, statistical, computation, security, and privacy concerns.The paper characterizes healthcare FL as nascent and requiring substantial further research.
  • Optimization: Hyperparameter optimization remains difficult because clients may not contribute to federated training before receiving benefits from participation.This creates a coordination problem before the FL process begins.
  • Security and Privacy: Existing methods for distinguishing benign from malicious models remain limited to controlled settings and specific environments.The paper identifies practical medical deployment as requiring solutions beyond these constrained defenses.
  • System and Statistical Challenges: Medical FL faces communication bottlenecks, statistical heterogeneity, and computational constraints across numerous and diverse clients.Lightweight models, sparsification, subsampling, and quantization are proposed for communication efficiency, while heterogeneous distributions can degrade performance.
  • Participation and Data Quality: Participation incentives and client-quality assessment are needed to encourage honest contributions and improve model generalizability.The paper highlights rewarding high-quality data contributions and excluding clients with poor-quality data as complementary directions.

G. Generative Pre-trained Large Language Model (FL-GPT)

Generative models and language models are being explored for medical imaging, diagnosis, education, and clinical research within the broader FL healthcare landscape. The paper also identifies unresolved needs for unlabeled-data learning, benchmarking, regulatory evaluation, explainability, and real-world deployment.

  • Healthcare Applications: Pre-trained large language models support healthcare tasks including administrative document generation, disease diagnosis, professional education, and clinical research.Examples include BERT, GPT, Amazon Titan, AI21 Jurassic, Cohere, and ChatGPT.
  • Medical Image Generation: Generative image models can create realistic medical images, including hard-to-find tumors, for radiological model training.Synthetic examples include small liver tumors under 10 mm and rare tumors.
  • Medical Image Generation: FL optimization of generative imaging models can improve robustness and generalizability for diagnostic and prognosis assistance.The passage connects these applications to early disease diagnosis and treatment decisions.
  • Learning with Unlabeled Data: Most FL research uses supervised learning, while contrastive learning is proposed for unlabeled and non-IID medical data.Semi-supervised and unsupervised approaches may not achieve high performance on medical image datasets with deep-learning models.
  • Evaluation and Reproducibility: Open FL medical datasets and benchmarking frameworks are needed to support reproducibility, generalizability, and innovation.The paper calls for expanded implementations and publicly available datasets for empirical comparison.
  • Regulation and Explainability: Medical FL still requires regulatory risk guidelines and improved explainability before broader practical adoption.The paper calls for cooperation between computer science and law researchers to assess privacy compliance under regulations such as GDPR.
  • Conclusions: The survey identifies FL as a privacy-preserving framework for medical applications while acknowledging system, communication, statistical, security, and privacy challenges.Its discussion emphasizes applications related to global cancer burdens and computer-aided diagnosis tools.
  • Future Directions: Future work anticipates more secure protocols, energy-efficient communication paradigms, and translation of FL into real-world healthcare solutions.These directions are presented as responses to the requirements of FL networks and medical applications.
Loading 2208.03392v5…