acceptodds
Withdrawn from ICLR 2027

FedOPS: Order-sensitive Parameter Sanitization for cross‑client Federated Learning in Medical Image Segmentation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recently, cross-client federated learning has received growing attention in medical image segmentation, owing to its superior performance and improved security guarantees compared with conventional server-based federated learning paradigms. However, existing methods suffer from two key challenges: 1) Severe data heterogeneity (non-IID) across clients often causes conflicting gradient directions during parameter exchanges, resulting in training oscillations that degrade convergence rate and final segmentation accuracy. 2) Critical parameter contamination during parameter transmission may cause interference parameters to overwhelm critical ones, resulting in incorrect pixel predictions. To overcome these issues, we propose a novel Order-sensitive Parameter Sanitization (FedOPS) framework for decentralized federated foundation models, which ensures more secure and stable training. Specifically, to promote stable convergence and effective knowledge propagation between different clients, we develop a consistency-aware client scheduling (CCS) strategy that orders clients based on feature representation similarity. To prevent critical parameter contamination, we introduce a cross-client parameter sanitization (CPS) mechanism that leverages historical update directions to selectively discard interfering updates. Extensive experiments on multiple benchmarks demonstrate that FedOPS consistently outperforms existing methods, highlighting its effectiveness for federated medical image segmentation.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Cross-client Federated Learning, Medical Image Segmentation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_gDgaOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on decentralized cross-client federated learning for medical image segmentation. Instead of relying on a central server for model aggregation, FedOPS adopts a sequential cross-client training paradigm in which model parameters are directly transmitted among clients. The method mainly addresses two issues: the client training order may affect knowledge propagation and convergence stability, and unstable or interfering parameter updates may be propagated across clients. To address these issues, the authors introduce two components. First, Consistency-aware Client Scheduling (CCS) orders clients according to the similarity of their feature representations so that similar clients are trained consecutively. Second, Cross-client Parameter Sanitization (CPS) evaluates the historical update-direction consistency of each trainable parameter and generates a binary mask to preserve relatively stable parameters while filtering unstable ones before transmission to the next client. The framework is built on SAM, with the pretrained encoder frozen and only the adapters and decoder being trained and transmitted. Experiments are conducted on Prostate, Fundus, and Nuclei segmentation datasets, with comparisons against FedUNet, FedCross, FedSAM, FedMSA, and TM-FAL, together with ablation studies on CCS and CPS.

Critical strengths
  1. The paper addresses the practical problem of unstable sequential cross-client training under heterogeneous data distributions.
  2. FedOPS has a simple and modular design that separates client scheduling (CCS) from parameter filtering (CPS).
  3. The method is computationally lightweight by freezing the SAM encoder and training and transmitting only the adapters and decoder.
  4. The experimental evaluation is relatively comprehensive, covering three medical image segmentation benchmarks and several federated learning baselines.
  5. The paper provides useful ablation and analysis experiments on CCS, CPS, backbone choice, the threshold parameter, communication rounds, and model complexity.
Critical weaknesses
  1. The CCS is an incremental extension of FedCross that adds feature-similarity-based client reordering, while the reformation is limited.

  2. The construction of the client feature representation used by CCS is insufficiently described, including its extraction, aggregation, and update strategy.

  3. The paper contains terminology inconsistencies, for example, using different full names for CCS and CPS in the main text and the caption of Figure 2.

Questions for authors
  1. How is the client feature representation used by CCS extracted, aggregated across local samples, and updated during training?
  2. Is the client training order fixed before training or recomputed across communication rounds as client representations change?
  3. Could independent parameter-wise masking in CPS disrupt coordinated updates among correlated parameters and introduce additional update drift?
Reviewer_TTmHOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary
  1. This paper proposes a decentralized federated learning framework for medical image segmentation, incorporating two key mechanisms: Consistency-aware Client Scheduling (CCS) and Cross-client Parameter Sanitization (CPS).

  2. The method is evaluated on three medical image segmentation benchmarks and achieves competitive performance.

Critical strengths

The paper addresses an important problem in decentralized federated medical image segmentation, particularly the challenges of client heterogeneity and unstable knowledge transfer.

Critical weaknesses
  1. Some key assumptions lack sufficient justification. For example, the authors state that "Random or arbitrary client sequences MAY introduce conflicting updates" when motivating CCS in page 4, line 197 and that "Critical parameters TEND to exhibit more stable update directions across training rounds, whereas interference parameters often undergo more diverse and inconsistent directional changes" when motivating CPS in page 5, line 232. Are there empirical observations or prior studies supporting these claims? These assumptions are fundamental to the proposed method.

  2. Confusing notation. DD denotes cosine distance in Eq. (2), but parameter differences in Eqs. (7)–(8). Also, Eqs. (10)–(11) use the same symbol CC for different quantities. A different symbol, such as C′C', should be used for the refined consistency score.

  3. Eqs. (7)–(8) define parameter differences as wt−wt−1w^t-w^{t-1}, while Eq. (11) uses their signs to refine the consistency score. Why does the direction of weight changes reliably indicate parameter consistency? The underlying rationale is unclear.

  4. Directly multiplying WW by the mask sets parameters to zero when (M=0M=0). However, CPS appears intended to discard undesirable updates rather than zero out existing weights. Shouldn't these parameters remain unchanged instead?

  5. Some typos. On page 4, the subscripts of Vt\mathcal{V}t and ct+1c{t+1} appear to be incorrectly formatted. The authors should carefully check the mathematical notation throughout the paper.

  6. Are the datasets heterogeneous across clients? CPS filters directionally inconsistent updates, but such updates may contain valuable domain-specific knowledge. Could this mechanism suppress the learning of heterogeneous information rather than facilitate it?

  7. Dice and IoU both measure region overlap and provide somewhat redundant information. Additional boundary-based metrics, such as HD95 and ASD, should be included to evaluate boundary accuracy and segmentation quality more comprehensively.

Questions for authors

This paper addresses an important problem in decentralized federated medical image segmentation and proposes two simple mechanisms, CCS and CPS. The experimental results across three benchmarks demonstrate promising performance. However, the paper suffers from unclear presentation and insufficient empirical justification for its key motivations. More importantly, the core methodology contains a potentially serious flaw in the masking formulation, raising concerns about its technical correctness. Given these issues, I recommend rejection.

Reviewer_nG45OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies cross-client federated medical image segmentation with SAM. It introduces consistency-aware client scheduling (CCS), which orders clients by feature similarity, and cross-client parameter sanitization (CPS), which masks parameters using their historical directions of change. Experiments on prostate, fundus, and nuclei segmentation report improvements in average performance. The paper also includes component ablations and comparisons of client orders and masking strategies.

Critical strengths
  1. The paper addresses a concrete issue in sequential cross-client training. Since different clients update the same model in sequence, the order of those updates is worth studying. CCS is accompanied by the client-order comparison in Figure 4, which gives experimental support for this part of the design.
  2. The evaluation covers prostate, fundus, and nuclei segmentation, spanning different medical image types. Average performance improves on all three tasks, and the four configurations in Table 4 examine the separate contributions of CCS and CPS. These results provide initial evidence that the combined approach is useful.
Critical weaknesses
  1. The definition of the CPS consistency score needs to be clearer. Equation (10) accumulates the fraction of nonnegative parameter changes, while Equation (11) takes its complement for a negative current change. It is not clear which state is carried into the next round. Under the historical-fraction interpretation, a parameter whose update sign alternates would still have a score near 0.5 and pass the default threshold of 0.1. Equation (13) also zeros parameter values, which differs from masking only the current update. Parameter pruning can be a valid design, but the paper needs to specify the complete recurrence and explain why this statistic identifies useful parameters. Retention rates or a contribution analysis would help establish what CPS actually removes.
  2. The stability and attribution of the performance gains remain uncertain. The methods in the main tables use different backbones and fine-tuning scopes, so their full performance differences cannot be attributed to CCS and CPS. Table 4 is a useful controlled ablation, but the client order and model configuration of the version without either component need to be specified. The paper should also report variability across independent runs. Finally, the leave-one-client-out protocol does not clearly explain which trained client model is used to predict on the held-out institution. These details matter when the main improvements are roughly one to two Dice percentage points.
Questions for authors
  1. In Equations (10)-(12), is the historical nonnegative-change fraction stored separately from the direction-adjusted score? Please give the initialization and exact recurrence. At the default threshold, which patterns of parameter change are actually removed?
  2. For Table 4, does the configuration without CCS and CPS use the same SAM backbone and fine-tuning scope as the full method, and what client order does it use? Please also explain how the final model for each held-out institution is chosen and provide independent-run variability for the principal comparison.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.