Abstract

Federated Learning (FL) is revolutionizing medical imaging by enabling collaborative model training across institutions without sharing sensitive patient data, ensuring compliance with stringent privacy regulations such as HIPAA and GDPR. This approach is particularly valuable in low- and middle-income countries (LMICs), where access to annotated training data and trained medical professionals is limited. However, FL adoption in LMICs faces significant barriers, including limited high-performance computing resources and unreliable internet connectivity. To address these challenges, we introduce FedNCA, a novel federated learning system tailored for medical image segmentation tasks. FedNCA leverages the lightweight MedNCA architecture, enabling training on low-cost edge devices, such as widely available smartphones, while minimizing communication costs. Additionally, the architecture’s robustness against data reconstruction attacks enhances privacy and security. By overcoming infrastructural and security challenges, FedNCA paves the way for inclusive, efficient, and privacy-preserving medical imaging solutions, fostering equitable healthcare advancements in resource-constrained regions.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/2510_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/MECLabTUDA/FedNCA

Link to the Dataset(s)

Fetal Abdominal Structures Segmentation dataset: https://data.mendeley.com/datasets/4gcpm9dsc3/1 MIMIC-III dataset: https://physionet.org/content/mimiciii/1.4/

BibTex

@InProceedings{LemNic_Equitable_MICCAI2025,
        author = { Lemke, Nick and Konstantin, Mirko and Krumb, Henry John and Kalkhof, John and Stieber, Jonathan and Mukhopadhyay, Anirban},
        title = { { Equitable Federated Learning with NCA } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15973},
        month = {September},
        page = {172 -- 182}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    This paper describes the potential of using the previously developed Med-NCA architecture for federated learning (FL). The authors call this FedNCA. Within the paper they describe various benefits of FedNCA, which seem to derive primarily from the smaller model size relative to UNet-type deep learning models.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.

    The paper provides a relatively high-level description of how the FedNCA system could operate and its potential benefits regarding its lightweight and efficient design.

  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.

    It is not clear that the FedNCA model differs from the previously published Med-NCA model. The methods provided are general and lack mathematical detail. While the smaller number of parameters in FedNCA might facilitate operations performed with homomorphic encryption, no calculations on encrypted data appear to have been performed. Instead, a comparison of encryption/decryption times is provided (Fig 5). Along with the relatively smaller transmission cost in Fig 3, these appear to be products of model size. The benefits of FedNCA appear to derive entirely from the relatively smaller model size. Various aspects of the FedNCA are tested, but they are tested separately, not as part of an integrated model. It is further not clear which dataset the results pertain to. The necessity for AI in geographical areas lacking medical professionals, particularly the need for collaborative training of models, is uncertain.

  • Please rate the clarity and organization of this paper

    Poor

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (1) Strong Reject — must be rejected due to major flaws

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper provides little innovation beyond a previously published model. While the tests may demonstrate the model’s potential for a federated learning application, it is limited.

  • Reviewer confidence

    Very confident (4)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper
    • Authors created a secure way to share medical AI models between hospitals that uses very little internet data.

    • They tested how well their models identify important areas in medical images and measured how much data they need to send when sharing updates.

    • They checked if their system works on regular smartphones in real-world conditions.
    • They studied how their approach makes patient data protection much faster and more efficient.
  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • The paper is clearly written and presents content that is valuable to the medical research community. Therefore, it is also relevant and of interest to the MICCAI community.
    • The proposed model helps developing countries: FedNCA makes AI-powered medical imaging accessible in places with limited resources and poor internet.
    • At just ~284KB, the model runs smoothly on basic smartphones and tablets, unlike bulkier alternatives.
    • The model requires up to 2000 times less data transfer than competing models, making it practical even with unreliable connections.
    • The proposed architecture supports encryption that’s 1800 times faster than alternatives due to its streamlined design.
    • The proposed framework was successfully tested on ultrasound and X-ray images using real-world devices, matching performance of more complex models.
    • The Code is made open-source for anyone to use and improve.
    • References are relevant to the studied application.
    • The figures are easy to understand and nicely presented.
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • The model was only tested in comparison to a few other models.
    • The framework has been only evaluated on relatively few images, which may not represent all real-world situations.
    • The model hasn’t yet been tested in actual clinical settings with different types of devices and user expertise.
    • The proposed model doesn’t fully explain what you might lose in image quality when using this simpler approach.
    • While encryption is mentioned, specific security protections aren’t thoroughly explained.
    • The paper doesn’t provide practical instructions for setting up and maintaining the system in resource-limited healthcare settings.
  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html

    N/A

  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (4) Weak Accept — could be accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    This paper introduces FedNCA, a tool that helps bring AI medical imaging to places with few resources. The tiny size (284KB) and low data needs make it great for areas with poor internet and basic phones. The good points are clear, it works on cheap devices, uses very little internet data, protects patient information, and is free for anyone to use. These strengths address real problems in bringing healthcare technology to developing countries. However, there are important issues. The testing used too few images. The paper doesn’t explain how regular healthcare workers would set this up and use it. We also need more details about how well it protects patient data and what quality trade-offs exist. I suggest accepting the paper but asking the authors to improve these weak areas. They should test on more diverse images, provide setup instructions for non-technical users, and better explain security features. With these improvements, this work could help bring AI benefits to underserved healthcare settings worldwide.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #3

  • Please describe the contribution of the paper

    The paper introduces FedNCA, a federated learning (FL) framework tailored for medical image segmentation in low-resource environments - specifically targeting low- and middle-income countries (LMICs). The core idea is to leverage Neural Cellular Automata (NCA), implemented as a lightweight Med-NCA backbone from prior work, to drastically reduce the parameter count (by orders of magnitude compared to traditional architectures like U-Net). This reduction facilitates not only efficient training on low-cost devices (e.g., smartphones) but also minimizes communication overhead. In addition, the paper addresses privacy concerns through the integration of a homomorphic encryption scheme (CKKS), claiming significant efficiency gains in encryption processes over heavier networks. Experimental evaluations on ultrasound and chest X-ray segmentation tasks are provided, with comparisons to baseline U-Net, TransUNet, and their compressed variants.

  • Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
    • The paper tackles an important problem of enabling resource efficient federated learning
    • The paper is well-written with motivations clearly defined and the reasons behind their decisions effectively developed
    • The use of NCAs is not only practical but also helps promulgate knowledge of an underutilized class of models in the medical community
    • The paper provides empirical validation and practical deployment considerations (via low cost devices), albeit without any error-bars
  • Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
    • In general the experiments could use some strengthening. Currently no standard deviations/errors are available for any of the reported metrics, which makes it difficult to ascertain the consistency of this approach
    • While the implementation is stated as publicly available, key experimental details remain somewhat high-level (e.g total number of training iterations, encryption parameters etc.)
    • Methodological novelty is limited. The work does not introduce a novel method per se but, (quite effectively) combines existing methods in a targeted fashion.
  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2025/en/REVIEWER-GUIDELINES.html
    • Algorithm 1 should also include indicate the use of a top-k selection procedure.
    • Results in Figure 3 need better elucidation. How is Fed UNet different from the other UNets? Are the top-k versions also 4-bit?
    • In the experiments, are the weights of the NCA full float-32? This should be clarified.
    • The training / test split for Ultrasound is a bit odd - significantly more samples used in testing. Is this common practice in federated learning?

    • Minor
    • In last sentence of conclusion “By combining efficient segmentation with federated training, Med-NCA ensures AI models … “ - perhaps you meant FedNCA?

    • In general, more hyperparameters need be explicitly stated for the NCA as well as baseline UNets i.e hidden dimensions, optimizer, training iterations etc. Additionally, as the authors do not report results across multiple runs, it is difficult to know the stability/variance of this approach.
  • Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.

    (5) Accept — should be accepted, independent of rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    Accept. While the papers falls a bit short on implementation details, it can be updated via a minor revision or the code release. As it stands currently, the paper is well written, elucidating both the landscape it addresses and the approach it took. The methodology is practical and shows strong empirical evidence.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A




Author Feedback

We thank the reviewers for recognizing FedNCA’s low transfer demands, lightweight computing requirements, and quick encryption capabilities. Our paper is submitted for the health equity track, where the technical contributions are less algorithmic and more practical in nature. We believe some misunderstanding of R2 stems from judging the paper as a purely methodological advancement.

  • R1 on too few comparison models: We compare to three common FL baseline algorithms and two SOTA models (CNN-, and transformer-based). An evaluation including more baselines is out of scope for this work.
  • R1 on heterogeneous data (including limited datasets, different types of devices, different user expertise): We agree that an evaluation in a lab lacks many details for clinical translation, such as different data distributions and expertise. This paper lays the foundation for inclusive real-world federated learning with homogeneous data distributions. We are working on a follow-up project investigating FedNCA with more data, including heterogeneous data distributions.
  • R1 on specific security protections: We agree that encryption in federated learning is an important concept and hence have emphasized the threat of privacy leakage on the second page. However, a detailed review of security issues would go beyond the scope of this manuscript, but will be subject of a follow-up work.
  • R1 on missing instructions for setup and maintenance: We agree that orchestrating a federated learning algorithm among different clinics can be challenging. Our source code, which we will publish on GitHub, contains detailed instructions on setting up and starting the FedNCA training. Yet, a full clinical translation is beyond the scope of this paper.
  • R1 on missing qualitative results: Our quantitative results suggest that no detail in segmentation is lost. For the camera-ready version, we include a few qualitative results in the supplementary material.

  • R2 on difference to Med-NCA: We modified the backbone Med-NCA model to be more sparse by reducing the number of convolutional layers and the latent dimension. Additionally, we changed the loss function and removed the random patch selection used in the original architecture.
  • R2 on lack of mathematical details: We acknowledge the importance of this topic but, due to space limitations, refer readers to Cheon et al. [25] for the details of the CKKS algorithm.
  • R2 on missing end-to-end experiments: Our experiments demonstrate the feasibility of a federated learning setup on smartphones. Due to the lack of sophisticated libraries compatible with Android, an entire training run on smartphones is out of scope for this paper. We are working on a follow-up project that will include such experiments.
  • R2 on missing motivation of FL for LMICs: AI can significantly reduce the workload of medical professionals, enabling the delivery of accessible and cost-effective healthcare, particularly in regions facing shortages of healthcare workers [3]. However, developing traditional AI solutions relies heavily on resources that are often limited in LMICs[4, 5]. Federated Learning presents a promising approach to address this challenge in areas where data availability is scarce [1*].

  • R3 on additional clarifications: We will report the variance in the supplementary of the camera-ready version. Due to space constraints, we report any experimental details in the public source code. Fed UNet is the same as the other UNets without sparsification or quantization. The top-k versions and training of FedNCA are always float-32. We wanted to create a challenging scenario with a few images for each client, leaving more samples for testing. We will make those things clearer in the camera-ready version.

[1*] Fabila, Jorge, et al. “Democratizing AI in Africa: Federated Learning for Low-Resource Edge Devices.” Meets Africa Workshop. Springer Nature Switzerland, 2024




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    The authors should clarify the points raised by the reviewers in their rebuttal.

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    N/A

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Accept

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



back to top