PaperlayerView exact revision on arXiv

Source-linked AI summary

FedDG: Federated Domain Generalization on Medical Image Segmentation via Episodic Learning in Continuous Frequency Space

Quande Liu, Cheng Chen, Jing Qin, Qi Dou, Pheng-Ann Heng

arXiv:2103.06030v1cs.CV

TL;DR

The paper addresses federated domain generalization, where models trained on decentralized source domains must generalize to completely unseen hospitals despite data-distribution shifts. It introduces ELCFS, combining privacy-preserving continuous frequency-space interpolation with boundary-oriented episodic learning. Experiments on two medical image segmentation tasks demonstrate superior performance over state-of-the-art methods and support the approach through ablation studies.

  • Problem

    Federated learning improves collaboration while keeping data decentralized, but existing medical segmentation models neglect generalization to completely unseen domains outside the federation.

  • Method

    ELCFS transfers multi-source distribution information through continuous frequency-space interpolation and uses boundary-oriented episodic learning for local training.

  • Results

    Superior performance over state-of-the-art methods is demonstrated on retinal fundus and prostate MRI segmentation tasks, with supporting ablation experiments.

  • Takeaways & Limitations

    The approach enables federated models to access multi-source distributions without privacy leakage and provides a learning scheme potentially extendable to other segmentation problems.

  • Takeaways & Limitations

    Existing domain-generalization methods typically require centralized multi-domain data, which conflicts with privacy-preserving federated learning on decentralized datasets.

Abstract

from arXiv · show

Federated learning allows distributed medical institutions to collaboratively learn a shared prediction model with privacy protection. While at clinical deployment, the models trained in federated learning can still suffer from performance drop when applied to completely unseen hospitals outside the federation. In this paper, we point out and solve a novel problem setting of federated domain generalization (FedDG), which aims to learn a federated model from multiple distributed source domains such that it can directly generalize to unseen target domains. We present a novel approach, named as Episodic Learning in Continuous Frequency Space (ELCFS), for this problem by enabling each client to exploit multi-source data distributions under the challenging constraint of data decentralization. Our approach transmits the distribution information across clients in a privacy-protecting way through an effective continuous frequency space interpolation mechanism. With the transferred multi-source distributions, we further carefully design a boundary-oriented episodic learning paradigm to expose the local learning to domain distribution shifts and particularly meet the challenges of model generalization in medical image segmentation scenario. The effectiveness of our method is demonstrated with superior performance over state-of-the-arts and in-depth ablation experiments on two medical image segmentation tasks. The code is available at "https://github.com/liuquande/FedDG-ELCFS".

1. Introduction

The paper identifies federated domain generalization as learning from decentralized source domains while generalizing directly to completely unseen domains. It proposes ELCFS, which transfers distribution information through continuous frequency-space interpolation and uses boundary-oriented episodic learning for local training.

  • Federated domain generalization targets direct generalization from multiple decentralized source domains to completely unseen target domains.
  • Existing federated medical segmentation studies focus on internal-client performance and neglect generalization to unseen domains outside the federation.Unseen hospitals may use different imaging scanners and protocols, creating distribution shifts.
  • Federated domain generalization is challenging because clients cannot directly access multi-source distributions during decentralized learning.Conventional domain-generalization methods often require samples from multiple source distributions during training.
  • ELCFS transfers distribution information across clients through continuous frequency-space interpolation while retaining core semantic phase information locally for privacy protection.The method exchanges amplitude spectra associated with low-level style and keeps phase spectra associated with high-level semantics at each client.
  • Boundary-oriented episodic learning exposes local optimization to domain shifts and enhances generalizability around ambiguous segmentation boundaries.
  • Experiments on retinal fundus and prostate MRI segmentation show superior performance over state-of-the-art methods and support the approach through analytical experiments.The evaluation covers four retinal fundus datasets and six prostate MRI datasets.

2. Related Work

Related work covers federated learning for privacy-preserving multi-site collaboration and domain-generalization methods for unseen targets. The paper emphasizes that existing domain-generalization approaches commonly require centralized multi-domain data, conflicting with federated decentralization.

  • Federated Learning in Medical Imaging: Federated learning trains a global model by aggregating local-client parameters while keeping medical data at each institution.Federated averaging is identified as a communication-efficient approach for distributed deep-network training.
  • Federated Learning in Medical Imaging: Prior federated medical-imaging studies improve performance on internal clients but do not address generalization to unseen domains outside the federation.
  • Domain Generalization: Domain generalization learns from multiple source domains so a model can directly generalize to unseen target domains.Existing approaches include domain-invariant representation learning and meta-learning with episodic meta-train and meta-test splits.
  • Domain Generalization: Typical domain-generalization methods require centralized access to multi-domain data, violating privacy in federated settings with decentralized datasets.
  • Domain Generalization: Other domain-generalization approaches avoid data centralization through architecture changes, self-supervision, training heuristics, or data augmentation.

3. Method

ELCFS addresses federated domain generalization by sharing cross-client distribution information through continuous frequency-space interpolation, then training locally with episodic domain shifts and boundary-oriented regularization. The method preserves semantic phase information while generating transformed appearances for generalizable segmentation learning.

  • 3.1. Federated Domain Generalization: FedDG learns from distributed source domains so a model can generalize directly to a completely unseen testing domain.The federated setting limits each client to its individual distribution, constraining access to multi-source information needed for generalization.
  • 3.2. Continuous Frequency Space Interpolation: ELCFS builds a shared amplitude-spectrum distribution bank while retaining each image’s phase spectrum locally to preserve semantic content and privacy.Amplitude captures low-level distribution or style information, whereas phase captures high-level semantics; raw images cannot be reconstructed from amplitude alone.
  • 3.2. Continuous Frequency Space Interpolation: Continuous interpolation replaces selected low-frequency amplitude components with external-client components, producing transformed images that bridge distinct domain distributions.The binary mask controls the exchanged low-frequency region, and the interpolation ratio λ is dynamically sampled from [0.0, 1.0].
  • 3.3. Boundary-oriented Episodic Learning: Each local client uses original images as meta-train data and frequency-transformed counterparts as meta-test data to simulate domain shift during episodic learning.The local scheme first updates parameters with segmentation Dice loss, then evaluates the updated parameters on held-out transformed data.
  • 3.3. Boundary-oriented Episodic Learning: The boundary-oriented objective targets ambiguous anatomy boundaries by promoting domain-invariant boundary delineation during meta-optimization.Its boundary loss uses cosine-similarity-based InfoNCE over region-level feature pairs, while the overall objective combines segmentation Dice loss with the boundary objective.

4. Experiments

Experiments evaluate ELCFS across federated domain-generalization settings for fundus and prostate MRI segmentation, comparing it with federated-compatible DG methods and analyzing its components and design choices. ELCFS improves unseen-domain performance, with continuous frequency interpolation and boundary-oriented episodic learning contributing complementary benefits.

  • Experimental setting: Experiments cover optic disc and cup segmentation on four fundus datasets and prostate segmentation on six MRI datasets.Evaluation follows leave-one-domain-out generalization, training on distributed source domains and testing on an unseen target domain.
  • Comparison with DG methods: Qualitative cases show ELCFS accurately segments structures and delineates boundaries on unknown-distribution images where other methods sometimes fail.The comparison includes unseen-domain cases for both fundus and prostate MRI tasks.
  • Ablation analysis: Removing either CFSI or BEL decreases generalization performance across unseen settings, indicating complementary roles in ELCFS.CFSI generates distributions for BEL, while BEL helps exploit those generated distributions.
  • Ablation analysis: Continuous interpolation smoothly bridges distinct generated-data distributions, while fixed interpolation produces several distinct distributions in t-SNE visualization.The authors associate the denser continuous distribution space with more domain-invariant local learning.
  • Ablation analysis: Continuous sampling of λ outperforms fixed positive values, with sampling over [0.0, 1.0] producing the best results compared with λ = 0.Any fixed λ > 0 improves performance over transferring no distribution information.
  • Ablation analysis: Generalization performance increases as more clients participate, and ELCFS consistently outperforms FedAvg across client-count settings.The analysis uses prostate MRI segmentation on two unseen sites while increasing participating clients from 1 to K −1.

5. Conclusion

The paper formulates federated domain generalization and introduces a method combining continuous frequency-space interpolation with boundary-oriented episodic learning. It demonstrates superior efficacy on two medical image segmentation tasks and identifies broader potential for privacy-preserving multi-source access in federated learning.

  • The paper proposes federated domain generalization and addresses it with continuous frequency-space interpolation plus boundary-oriented episodic learning.
  • The method demonstrates superior efficacy on two important medical image segmentation tasks.
  • Privacy-preserving access to multi-source distributions may also address other federated-learning problems, including data heterogeneity.
  • The boundary-oriented learning scheme is generally extendable to other segmentation problems.

7. Supplementary Material

The supplementary material documents the experimental data sources, scanning protocols, statistical analysis, and frequency-space image transformations. It shows that increasing interpolation changes image style toward another client while preserving semantic content.

  • Experimental data sources: The retinal fundus experiments use four clinical centers drawn from three public datasets, while the prostate MRI task uses six data sources.
  • Scanning protocols: Tables 3 and 4 provide scanning-protocol details for the fundus and prostate MRI data sources.
  • Statistical analysis: The study compares methods with paired t-tests using Dice scores and a 0.05 significance level across unseen-site generalization settings.
  • Statistical analysis: Table 5 reports p-values for statistical comparisons between the proposed approach and competing methods on overall Dice score.
  • Frequency-space interpolation: Increasing interpolation ratio λ gradually transfers a local image’s style toward another client’s target style without changing semantic content.
Loading 2103.06030v1…