CoChat
Sign Up
Cited by 518Privacy-Preserving Technologies in DataDomain Adaptation and Few-Shot LearningArtificial Intelligence in Healthcare and Education

Federated learning for medical image analysis: A survey

Open access
Guan, Hao, Yap, Pew-Thian, Bozoki, Andrea, Liu, Mingxia·2024·Pattern Recognition
DOI 10.1016/j.patcog.2024.110424PMC PMC10976951
Open this paper in CoChat — free accountOpen PDF

Abstract

Machine learning in medical imaging often faces a fundamental dilemma, namely, the small sample size problem. Many recent studies suggest using multi-domain data pooled from different acquisition sites/centers to improve statistical power. However, medical images from different sites cannot be easily shared to build large datasets for model training due to privacy protection reasons. As a promising solution, federated learning, which enables collaborative training of machine learning models based on data from different sites without cross-site data sharing, has attracted considerable attention recently. In this paper, we conduct a comprehensive survey of the recent development of federated learning methods in medical image analysis. We have systematically gathered research papers on federated learning and its applications in medical image analysis published between 2017 and 2023. Our search and compilation were conducted using databases from IEEE Xplore, ACM Digital Library, Science Direct, Springer Link, Web of Science, Google Scholar, and PubMed. In this survey, we first introduce the background of federated learning for dealing with privacy protection and collaborative learning issues. We then present a comprehensive review of recent advances in federated learning methods for medical image analysis. Specifically, existing methods are categorized based on three critical aspects of a federated learning system, including client end, server end, and communication techniques. In each category, we summarize the existing federated learning methods according to specific research problems in medical image analysis and also provide insights into the motivations of different approaches. In addition, we provide a review of existing benchmark medical imaging datasets and software platforms for current federated learning research. We also conduct an experimental study to empirically evaluate typical federated learning methods for medical image analysis. This survey can help to better understand the current research status, challenges, and potential research opportunities in this promising research field.

AI snapshot

Written from the paper’s full text.

TL;DR

This survey reviews federated learning (FL) methods for medical image analysis published from 1 January 2017 to 31 October 2023, organizing them by client-end, server-end, and client-server communication perspectives. It also reviews FL software platforms and benchmark medical imaging datasets, and empirically compares representative FL methods on ADNI for Alzheimer’s disease classification.

Key findings

  • The survey categorizes existing FL methods for medical image analysis into three system-level groups: client-end methods, server-end methods, and client-server communication methods, with client-end work addressing domain shift, limited data/labels, and heterogeneous computation/data resources.
  • It reviews FL software platforms including PySyft, OpenFL, PriMIA, and Fed-BioMed, and benchmark datasets including ADNI, ABIDE, BraTS, RSNA brain CT, UK Biobank, IXI, CheXpert, ChestX-ray, COVID-19 CXR, COVIDx, ACDC, M&M, HAM10000, ISIC, Kaggle Diabetic Retinopathy, PROMISE12, TCGA, fastMRI, and MedMNIST.
  • In the ADNI experiment, ADNI-1 contained 428 subjects (199 AD, 229 normal controls) and ADNI-2 contained 360 subjects (159 AD, 201 normal controls); analysis used 90 AAL brain ROIs and mean gray matter volumes, with 80% training and 20% testing repeated five times.
  • The paper reports that the “Mix” strategy performed best, the “Cross” strategy worst, FL methods achieved satisfactory performance, and among FL methods FedAvg and FedProx were more advantageous than FedSGD.

Methods

  • Survey design: literature search across IEEE Xplore, ACM Digital Library, Science Direct, Springer Link, Web of Science, Google Scholar, and PubMed for papers published from 1 January 2017 to 31 October 2023; results were refined by removing duplicates and papers not closely related to medical imaging, then analyzed and organized into the survey.
  • Experimental study: AD vs. normal-control classification on ADNI using logistic regression, evaluated with accuracy, sensitivity, specificity, and AUC; compared Cross, Single, Mix, FedAvg, FedSGD, and FedProx.
  • Experimental details: 80/20 train/test split repeated five times; FedAvg and FedSGD used 10 local iterations; FedProx used a proximal term coefficient μ = 0.1 and local training once.

Limitations

  • The paper states major challenges including data heterogeneity/client shift, privacy leakage and poisoning attacks, technological limitations such as computational resources, communication costs, network resource allocation, and synchronization, and long-term viability issues such as scalability, sustainability, evolving regulations, and clients joining or leaving.
  • The paper states future needs including federated domain adaptation, multi-modality fusion, generalizability to unseen clients, weakly supervised learning, security defenses, blockchain/decentralization, medical video analysis, large-scale benchmarks, interpretability, and real-world implementation.
  • Evident from the text: the experimental evaluation is limited to one dataset (ADNI), one task (AD vs. NC), one simple model (logistic regression), and two clients; the survey search and refinement process is described but lacks formal systematic-review protocol details such as PRISMA-style counts or inclusion/exclusion criteria.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1016/j.patcog.2024.110424

Cites

152 of 152 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 54 cited sources checked · 0 supported · 0 flagged

  • Findingnone given

    Among the FL methods, aggregation of model weights (FedAvg, FedProx) is more advantageous than a fusion of the gradients of each client model (FedSGD).

    “Among the FL methods, we find that aggregation of model weights (i.e., FedAvg, FedProx) can be more advantageous than a fusion of the gradients of each client model (i.e., FedSGD).”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The ADNI-1 and ADNI-2 subsets, after excluding duplicate subjects, provide an appropriate simulation of independent federated clients for evaluating FL methods.
  • Findingreasoning

    Federated learning methods achieve satisfactory performance.

    “Federated learning methods achieve satisfactory performance.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The ADNI-1 and ADNI-2 subsets, after excluding duplicate subjects, provide an appropriate simulation of independent federated clients for evaluating FL methods.
  • Premisecitation

    Medical images from different sites cannot be easily shared to build large datasets for model training due to privacy protection reasons.

    “However, medical images from different sites cannot be easily shared to build large datasets for model training due to privacy protection reasons.”
    • Also stated in: § 1. Introduction · boosted
      “However, sharing medical imaging data between different sites/centers is intractable due to strict privacy protection policies such as Health Insurance Portability and Accountability Act (HIPAA) [10] and General Data Protection Regulation (GDPR) [11].”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [10] US Department of Health and Human Services, 2020

      Not checked in the free run.

    • cites [11] General Data Protection Regulation, 2019

      Not checked in the free run.

  • Premisecitation

    Some survey papers cover very broad areas in medicine and healthcare applications without detailed introduction on federated learning in medical image analysis.

    “And some survey papers cover very broad areas in medicine and healthcare applications [23,24], without detailed introduction on federated learning in medical image analysis.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [23] Nguyen, 2022, Federated learning for smart healthcare: A survey, ACM Comput. Surv. source

      Not checked in the free run.

    • cites [24] Pfitzner, 2021, Federated learning in a medical context: A systematic literature review, ACM Trans. Internet Technol. (TOIT) source

      Not checked in the free run.

  • Premisecitation

    Medical image analysis faces a fundamental small-sample-size problem.

    “Machine learning in medical imaging often faces a fundamental dilemma, namely, the small sample size problem.”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [8] Raudys, 1991, Small sample size effects in statistical pattern recognition: Recommendations for practitioners, IEEE Trans. Pattern Anal. Mach. Intell. source

      Not checked in the free run.

    • cites [9] Vabalas, 2019, Machine learning algorithm validation with a limited sample size, PLOS ONE source

      Not checked in the free run.

  • May combine two claims. 28
  • Classified as probably not a substantial claim. 39
  • The quote only partly states this claim. 2

  • May combine two claims.

    Premisecitation

    There have been several survey papers on federated learning, but further technical details about facilitating federated learning in medicine and healthcare are not yet covered.

    “There have been several survey papers on federated learning [15][16][17][18][19][20], but further technical details about facilitating federated learning in medicine and healthcare are not yet covered.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [15][16][17][18][19][20]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    Domain adaptation is a sound machine learning technique that has been widely used in medical image analysis.

    “Domain adaptation is a sound machine learning technique that has been widely used in medical image analysis [3].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [3] Guan, 2022, Domain adaptation for medical image analysis: A survey, IEEE Trans. Biomed. Eng. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given

    Domain adaptation aims to reduce domain shift among different medical image datasets and enhance the generalizability of a machine learning model.

    “It aims to reduce domain shift (in terms of certain distances) among different medical image datasets and enhance the generalizability of a machine learning model.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Premisereasoning

    Federated learning depends on model aggregation/fusion techniques to jointly train a global model which is then sent/broadcast to each site for fine-tuning and deployment.

    “Instead, as shown in Fig. 1, federated learning depends on model aggregation/fusion techniques to jointly train a global model which is then sent/broadcast to each site for fine-tuning and deployment.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Premisenone given

    Weight aggregation in federated learning plays a crucial role by combining model updates from multiple decentralized clients to form a single, improved global model.

    “Weight aggregation in federated learning plays a crucial role by combining the model updates from multiple decentralized clients to form a single, improved global model.”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Premisereasoning

    After image harmonization, all medical images are expected to have similar styles, thus reducing domain shift.

    “After harmonization, all the medical images are expected to have similar styles, thus reducing domain shift.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    The generalizability of the global model learned by federated learning should outperform each local model learned through local training.

    “The ultimate output of FL is the learned model ℳ* which is broadcast to each client for deployment, and the generalizability of ℳ* by FL should outperform each local model ℳ i (typically with the same model architecture as ℳ*) learned through local training.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given

    Most existing machine/deep learning methods can be plugged and integrated into a federated learning framework.

    “Most existing machine/deep learning methods can be plugged and integrated into an FL framework.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisereasoning

    Limited labeled medical images in a client can cause a client model to be badly trained, which can negatively influence the entire federation.

    “A client model may be badly trained with limited labeled data, which can cause negative influences on the entire federation.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: In real-world practice, there are often limited medical images in one client/site, and labeled medical images are even fewer due to the high cost of image annotation/labeling.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given

    Federated learning provides a generic framework for distributed learning with privacy preservation.

    “Federated learning (FL) provides a generic framework for distributed learning with privacy preservation.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisereasoning

    Domain shift among clients in federated learning may cause difficult convergence of the global model and performance degradation of some clients.

    “In an FL system, domain shifts may cause difficult convergence of the global model and performance degradation of some clients.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisereasoning

    Different medical sites often have significantly different computation resources and amounts of data, which may lead to slow convergence of federated learning model optimization.

    “Different medical sites, however, often have significantly different computation resources and amounts of data (images), which may lead to slow convergence of FL model optimization.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: Different medical sites often have significantly different computation resources and amounts of data.

  • Classified as probably not a substantial claim.

    Premisecitation

    A recent survey also reviews the application of federated learning on medical image analysis.

    “A recent survey also reviews the application of federated learning on medical image analysis [27].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [27] Sohan, 2023, A systematic review on federated learning in medical image analysis, IEEE Access source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisecitation

    Federated learning enables collaborative training of machine learning models based on data from different sites without cross-site data sharing.

    “As a promising solution, federated learning, which enables collaborative training of machine learning models based on data from different sites without cross-site data sharing, has attracted considerable attention recently.”
    • Also stated in: § 1. Introduction · plain
      “federated learning [12][13][14] has become a spotlight research topic in recent years, which aims to train machine learning models in a collaborative manner without exchanging/sharing data among different sites.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [12][13][14]

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Collecting, sharing, and processing of personal data are strictly constrained, and violating these laws and regulations may face high-cost penalties.

    “Collecting, sharing, and processing of personal data are strictly constrained, and violating these laws and regulations may face high-cost penalties [30].”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [30] Satariano, 2019, Google is fined $57 million under Europe’s data privacy law, N.Y. Times

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    PriMIA is an open-source framework for privacy-preserving decentralized deep learning with medical images.

    “The Privacy-preserving Medical Image Analysis (PriMIA) [93] is an open-source framework for privacy-preserving decentralized deep learning with medical images.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [93] Kaissis, 2021, End-to-end privacy preserving deep learning on multi-institutional medical imaging, Nat. Mach. Intell. source

      Not checked in the free run.

  • The quote only partly states this claim.

    Premisecitation

    Most existing FL research in medical image analysis uses databases from different medical sites/centers for simulation and experiment.

    “The first way is to directly use databases from different medical sites/centers [48,100].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [48] Li, 2020, Multi-site fMRI analysis using privacy-preserving federated learning and domain adaptation: ABIDE results, Med. Image Anal. source

      Not checked in the free run.

    • cites [100] Dayan, 2021, Federated learning for predicting clinical outcomes in patients with COVID-19, Nature Med. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given

    Fed-BioMed is an open-source federated learning software for real-world medical applications.

    “Fed-BioMed 5 is an open-source federated learning software for real-world medical applications.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    ADNI is the largest and most influential benchmark for Alzheimer's Disease research, including ADNI-1, ADNI-2, ADNI-GO and ADNI-3.

    “ADNI-The Alzheimer's Disease Neuroimaging Initiative (ADNI) [101,102] is the largest and most influential benchmark for the research of Alzheimer's Disease (AD), including ADNI-1, ADNI-2, ADNI-GO and ADNI-3.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [101] Mueller, 2005, The Alzheimer’s disease neuroimaging initiative, Neuroimaging Clin. source

      Not checked in the free run.

    • cites [102] Jack, 2008, The Alzheimer’s disease neuroimaging initiative (ADNI): MRI methods, J. Magn. Reson. Imaging source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    ABIDE contains both structural and functional brain images independently collected from more than 24 imaging laboratories/sites around the world.

    “ABIDE contains both structural and functional brain images independently collected from more than 24 imaging laboratories/sites around the world.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [103] Di Martino, 2014, The autism brain imaging data exchange: towards a large-scale evaluation of the intrinsic brain architecture in autism, Mol. Psychiatry source

      Not checked in the free run.

  • May combine two claims.

    Premisecitation

    Silva et al. used Fed-BioMed to conduct multi-center analysis for structural brain imaging data (MRI) across different datasets and verified its effectiveness.

    “Silva et al. [98] use Fed-BioMed to conduct multi-center analysis for structural brain imaging data (MRI) across different datasets and verify its effectiveness.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [98] Silva, 2020, Fed-biomed: A general open-source frontend framework for federated learning in healthcare

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisenone given

    OpenFL is an open-source FL framework initially developed for use in medical imaging.

    “The Open Federated Learning (OpenFL) 3 is an open-source FL framework initially developed for use in medical imaging.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    BraTS is a benchmark dataset for brain tumor segmentation.

    “Brain Tumor Image Segmentation Benchmark (BraTS) [104] is a benchmark dataset for brain tumor segmentation.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [104] Menze, 2014, The multimodal brain tumor image segmentation benchmark (BRATS), IEEE Trans. Med. Imaging source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    ADNI provides structural brain MRI, functional MRI, and PET from 1900+ subjects and 59 centers.

    “Structural brain MRI, functional MRI, and positron emission tomography (PET) from 1900+ subjects and 59 centers are provided for analysis and research.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [101] Mueller, 2005, The Alzheimer’s disease neuroimaging initiative, Neuroimaging Clin. source

      Not checked in the free run.

    • cites [102] Jack, 2008, The Alzheimer’s disease neuroimaging initiative (ADNI): MRI methods, J. Magn. Reson. Imaging source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    PySyft is an open-source FL library enabling secure and private machine learning by wrapping popular deep learning frameworks.

    “PySyft [95] 1 is an open-source FL library enabling secure and private machine learning by wrapping popular deep learning frameworks.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [95] Ziller, 2021, Pysyft: A library for easy federated learning

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    RSNA is a large-scale multi-institutional CT dataset for intracranial hemorrhage detection.

    “RSNA brain CT-Radiological Society of North America (RSNA) [105] is a large-scale multi-institutional CT dataset for intracranial hemorrhage detection.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [105] Flanders, 2020, Construction of a machine learning dataset through collaboration: the RSNA 2019 brain CT hemorrhage challenge, Radiol.: Artif. Intell.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    fastMRI is a large-scale dataset for medical image reconstruction using machine learning, containing more than 1500 knee MRIs and DICOM images from 10,000 clinical knee MRIs.

    “The fastMRI [117,118] 15 is a large-scale dataset for medical image reconstruction using machine learning approaches.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [117] Knoll, 2020, fastMRI: A publicly available raw k-space and DICOM dataset of knee images for accelerated MR image reconstruction using machine learning, Radiol.: Artif. Intell.

      Not checked in the free run.

    • cites [118] Muckley, 2021, Results of the 2020 fastMRI challenge for machine learning MR image reconstruction, IEEE Trans. Med. Imaging source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given

    Kaggle Diabetic Retinopathy dataset is a large-scale dataset of color digital retinal fundus images for diabetic retinopathy detection, including 17,563 pairs of images.

    “The Kaggle Diabetic Retinopathy (Retina) 13 is a large-scale dataset of color digital retinal fundus images for diabetic retinopathy detection.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    The ADNI dataset used in the experiment comprises ADNI-1 with 1.5T T1-weighted MRIs from 428 subjects (199 AD, 229 NC) and ADNI-2 with 3.0T T1-weighted MRIs from 360 subjects (159 AD, 201 NC).

    “This dataset comprises two subsets: ADNI-1, featuring 1.5T T1-weighted MRIs from 428 subjects (199 with Alzheimer's Disease (AD) and 229 normal controls (NCs)), and ADNI-2, with 3.0T T1weighted MRIs from 360 subjects (159 AD patients and 201 NCs).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [101] Mueller, 2005, The Alzheimer’s disease neuroimaging initiative, Neuroimaging Clin. source

      Not checked in the free run.

    • cites [102] Jack, 2008, The Alzheimer’s disease neuroimaging initiative (ADNI): MRI methods, J. Magn. Reson. Imaging source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    ChestX-ray contains 112,120 X-ray images of 30,805 patients with 14 disease labels.

    “ChestX-ray (also known as ChestX-ray14) 8 is a large and publicly-available medical image dataset that contains 112,120 X-ray images (in frontalview) of 30,805 patients with 14 disease labels.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [108] X. Wang, Y. Peng, L. Lu, Z. Lu, M. Bagheri, R.M. Summers, Chestx-ray8: Hospital-scale chest X-ray database and benchmarks on weakly-supervised classification and localization of common thorax diseases, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp. 2097–2106. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    M&Ms Challenge is a publicly available cardiac MRI dataset containing 375 participants from 6 hospitals in Spain, Canada, and Germany, acquired by 4 different scanners.

    “Multi-Center, Multi-Vendor, and Multi-Disease Cardiac Segmentation (M&Ms) Challenge [112] 10 is a publicly available cardiac MRI dataset.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [112] Campello, 2021, Multi-centre, multi-vendor and multi-disease cardiac segmentation: The M&Ms challenge, IEEE Trans. Med. Imaging source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    COVID-19 Chest X-ray is a publicly-available database containing 3616 COVID-19 positive cases, 10,192 normal controls, 6012 lung opacity, and 1345 viral pneumonia cases.

    “COVID-19 Chest X-ray (also known as COVID-19 CXR) [109] 9 is a publicly-available database of chest X-ray images, containing 3616 COVID-19 positive cases, 10,192 normal controls, 6012 lung opacity (non-COVID infection), and 1345 viral pneumonia cases.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [109] Chowdhury, 2020, Can AI help in screening viral and COVID-19 pneumonia?, IEEE Access source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    COVIDx is a large-scale fully accessible database comprising 13,975 chest X-ray images of 13,870 patients.

    “COVIDx dataset [110] is a large-scale and fully accessible database comprising 13,975 chest X-ray images of 13,870 patients.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [110] Wang, 2020, Covid-net: A tailored deep convolutional neural network design for detection of COVID-19 cases from chest X-ray images, Sci. Rep.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    TCGA is a large-scale landmark cancer genomics database providing whole-slide images for normal controls and cancers for histology and microscopy research.

    “The Cancer Genome Atlas (TCGA) [116] 14 is a large-scale landmark cancer genomics database.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [116] Cancer Genome Atlas Research Network, 2013, The cancer genome atlas pan-cancer analysis project, Nature Genet. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    CheXpert is a large-scale dataset including 224,316 chest radiographs of 65,240 patients, acquired from Stanford University Medical Center.

    “[107] is a large-scale dataset including 224,316 chest radiographs of 65,240 patients.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [107] J. Irvin, et al., Chexpert: A large chest radiograph dataset with uncertainty labels and expert comparison, in: Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 33, 2019, pp. 590–597. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    BraTS contains brain MRIs acquired by various scanners from around 19 independent institutions.

    “It contains brain MRIs acquired by various scanners from around 19 independent institutions.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [104] Menze, 2014, The multimodal brain tumor image segmentation benchmark (BRATS), IEEE Trans. Med. Imaging source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    RSNA contains 874,035 images compiled from three institutions: Stanford University, Thomas Jefferson University Hospital, and Universidade Federal de São Paulo.

    “RSNA contains 874,035 images which are compiled and archived from three different institutions, i.e., Stanford University (Palo Alto, USA), Thomas Jefferson University Hospital (Philadelphia, USA), and Universidade Federal de São Paulo (São Paulo, Brazil).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [105] Flanders, 2020, Construction of a machine learning dataset through collaboration: the RSNA 2019 brain CT hemorrhage challenge, Radiol.: Artif. Intell.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    ACDC is a large publicly available and fully annotated dataset for cardiac MRI assessment, consisting of 150 patients divided into 5 categories.

    “Automatic Cardiac Diagnosis Challenge (ACDC) [111] is a large publicly available and fully annotated dataset for cardiac MRI assessment.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [111] Bernard, 2018, Deep learning techniques for automatic MRI cardiac multi-structures segmentation and diagnosis: is the problem solved?, IEEE Trans. Med. Imaging source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    UK Biobank is a large-scale biomedical database containing genetic and health data from half a million participants, with four imaging centers and brain scans and cardiac MRI.

    “UK Biobank [106] is a large-scale, influential biomedical database and research resource containing genetic and health data from half a million participants.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [106] Miller, 2016, Multimodal population brain imaging in the UK Biobank prospective epidemiological study, Nature Neurosci. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    ISIC challenge dataset is a large-scale database containing a series of challenges for skin lesion image analysis, and has become a standard benchmark dataset for dermatoscopic image analysis.

    “The International Skin Imaging Collaboration (ISIC) challenge dataset [114] 12 is a large-scale database, containing a series of challenges for skin lesion image analysis.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [114] Cassidy, 2022, Analysis of the ISIC image datasets: Usage, benchmarks and recommendations, Med. Image Anal. source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    IXI Dataset consists of around 600 MR images from healthy subjects, acquired from three different hospitals in London.

    “IXI Datasetfoot_7 consists of around 600 MR images from healthy subjects.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    The 'mix' strategy's best performance is because it combines all training data and the learning model can access the largest amount of data information.

    “This is because it combines all the training data of the clients and the learning model can get access to the largest amount of data information than the other methods.”
    Quote in the paperSubstantial claim
    • Rests on an assumption: The ADNI-1 and ADNI-2 subsets, after excluding duplicate subjects, provide an appropriate simulation of independent federated clients for evaluating FL methods.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Premisenone given

    Certain FL algorithms demand substantial computational resources, posing challenges for the underlying hardware infrastructure.

    “For instance, certain FL algorithms demand substantial computational resources, posing challenges for the underlying hardware infrastructure.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Premisenone given

    Most existing FL studies only focus on images of a single modality.

    “Most existing FL studies only focus on images of a single modality.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    Duplicate subjects across ADNI-1 and ADNI-2 were excluded for independence.

    “Duplicate subjects across ADNI-1 and ADNI-2 were excluded for independence.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Premisenone given

    FL methods can encounter difficulties during real-world implementation, such as compatibility with existing hospital systems, integration challenges, and user adoption hurdles.

    “FL methods, while promising, can encounter difficulties during real-world implementation, such as compatibility with existing hospital systems, integration challenges, and user adoption hurdles.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Premisenone given

    In real-world scenarios, unforeseen challenges emerge, such as clients leaving or joining the training process, as well as unexpected technical and connectivity issues.

    “In real-world scenarios, unforeseen challenges emerge, such as clients leaving or joining the training process, as well as unexpected technical and connectivity issues.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisecitation

    FL still faces privacy and security risks, including privacy leakage and poisoning attacks.

    “But further research reveals that FL still faces privacy and security risks, including privacy leakage [87][88][89] and poisoning attacks [169,170].”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [87][88][89]

      Not checked in the free run.

    • cites [169] Lyu, 2022, Privacy and robustness in federated learning: Attacks and defenses, IEEE Trans. Neural Netw. Learn. Syst.

      Not checked in the free run.

    • cites [170] Xia, 2023, Poisoning attacks in federated learning: A survey, IEEE Access source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given

    The majority of research on FL in medical imaging has primarily focused on algorithm development and simulation.

    “The majority of research on FL in medical imaging has primarily focused on algorithm development and simulation.”
    Quote in the paperNot a substantial claim

  • The quote only partly states this claim.

    Premisenone given

    Most existing FL studies focus on model training and test within a fixed federation system, where a global model is trained on and applied to the same client datasets.

    “Most existing FL studies focus on model training and test within a fixed federation system.”
    Quote in the paperSubstantial claim
50 of 50

  • The paper relies on this without support.

    In real-world practice, there are often limited medical images in one client/site, and labeled medical images are even fewer due to the high cost of image annotation/labeling.

    “In real-world practice, there are often limited medical images in one client/site, and labeled medical images are even fewer due to the high cost of image annotation/labeling.”

  • The paper relies on this without support.

    Different medical sites often have significantly different computation resources and amounts of data.

    “Different medical sites, however, often have significantly different computation resources and amounts of data (images), which may lead to slow convergence of FL model optimization.”

  • The paper relies on this without support.

    The first paper on federated learning was released in 2016.

    “The first paper on federated learning was released in the year of 2016, thus the searching process for this survey ranges from 1st January 2017 to 31st October 2023 (see Fig. 2).”

  • The paper relies on this without support.

    The ADNI-1 and ADNI-2 subsets, after excluding duplicate subjects, provide an appropriate simulation of independent federated clients for evaluating FL methods.

    “This dataset comprises two subsets: ADNI-1, featuring 1.5T T1-weighted MRIs from 428 subjects (199 with Alzheimer's Disease (AD) and 229 normal controls (NCs)), and ADNI-2, with 3.0T T1weighted MRIs from 360 subjects (159 AD patients and 201 NCs).”
    “Duplicate subjects across ADNI-1 and ADNI-2 were excluded for independence.”
4 of 4

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Federated learning enables multi-institutional research, supporting statistic-based discoveries across geographical barriers and promoting a collaborative research environment that may accelerate advancements in personalized medicine, disease prediction, and treatment efficacy.

    “FL enables multi-institutional research, supporting statistic-based discoveries across geographical barriers and promoting a collaborative research environment that may accelerate advancements in personalized medicine, disease prediction, and treatment efficacy [27].”

    From Federated Learning in Smart Healthcare: A Comprehensive Review on Privacy, Security, and Predictive Analytics with IoT Integration (2024) · cites this paper as [27]

    Split into 2 atoms — 1 confirmed, 1 partially supported, 0 disputed, 0 misattributed, 0 unverifiable. Expand the claim to see per-atom verdicts.

    Experimental — accuracy not yet measured

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Privacy-law citations point to federated-learning papers

    Minor

    The passage cites references 10 and 11 for HIPAA and GDPR, respectively. Those bibliography entries are papers on federated learning, not the named privacy laws.

  • FedAvg attribution points to a different reference

    Minor

    The text attributes FedAvg to McMahan et al. [12], but reference 12 is by Kairouz et al. The McMahan et al. paper appears as reference 10.

2 of 2

Cites

Flagged first, then by citations

Research works printed but not linked: 8

  1. K. Bonawitz, H. Eichner, W. Grieskamp, D. Huba, A. Ingerman, V. Ivanov, C. Kiddon, J. Konečnỳ, S. Mazzocchi, B. McMahan, et al., Towards federated learning at scale: System design, in: Proceedings of Machine Learning and Systems, Vol. 1, 2019, pp. 374–388.
  2. Satariano, 2019, Google is fined $57 million under Europe’s data privacy law, N.Y. Times, 21
  3. Chaitanya, 2020, Contrastive learning of global and local features for medical image segmentation with limited annotations, Adv. Neural Inf. Process. Syst., 33, 12546
  4. Smith, 2017, Federated multi-task learning, Adv. Neural Inf. Process. Syst., 30
  5. Geiping, 2020, Inverting gradients-how easy is it to break privacy in federated learning?, Adv. Neural Inf. Process. Syst., 33, 16937
  6. M. Malekzadeh, B. Hasircioglu, N. Mital, K. Katarya, M.E. Ozfatura, D. Gündüz, Dopamine: Differentially private federated learning on medical data, in: The Second AAAI Workshop on Privacy-Preserving Artificial Intelligence, PPAI-21, 2021, pp. 1–9.
  7. X. Li, K. Huang, W. Yang, S. Wang, Z. Zhang, On the Convergence of FedAvg on Non-IID Data, in: Proceedings of International Conference on Learning Representations, 2020, pp. 1–12.
  8. T. Li, A.K. Sahu, M. Zaheer, M. Sanjabi, A. Talwalkar, V. Smith, Federated optimization in heterogeneous networks, in: Proceedings of Machine Learning and Systems, Vol. 2, 2020, pp. 429–450.

Cited by

Sorted by citations

Related

Authors
Guan, H., Yap, P.-T., Bozoki, A., & Liu, M.
Affiliations
University of North Carolina at Chapel Hill (US)
Corresponding author
Liu, Mingxia
ORCID
Venue
Pattern Recognition
Publisher
Elsevier BV
Published
2024-07
Volume
151
Pages
110424
ISSN
0031-3203 (ISSN-L)
DOI
10.1016/j.patcog.2024.110424
PMC
PMC10976951
OpenAlex
W4392699793
Cited by
518
References
152
Citations by year
2026: 236 · 2025: 222 · 2024: 59 · 2023: 1
FWCI
76.14
Type
journal article
Language
en
Open access
Licence
Locations
  • Pattern Recognition (publishedVersion)
  • PubMed (publishedVersion)
  • PubMed Central (acceptedVersion)
  • arXiv (Cornell University) (submittedVersion, cc-by-nc-nd)
  • arXiv (Cornell University)
Funders
  • Foundation for the National Institutes of Health
  • Alzheimer's Disease Neuroimaging Initiative
  • National Institutes of Health
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Privacy-Preserving Technologies in Data
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
Keywords
federated learning, medical image analysis, privacy-preserving machine learning, small sample size problem, collaborative training, communication techniques

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-09.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-09.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review