REVIEW 4 major objections 5 minor 24 references
FedCLAM: Client Adaptive Momentum with Foreground Intensity Matching for Federated Medical Image Segmentation
T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read FedCLAM adapts momentum per client and aligns foreground intensities to improve federated medical segmentation accuracy and fairness.
desk verdict The core idea is a genuine tweak to momentum-based federated aggregation, but the reported superiority is not statistically supported and a stray paragraph in the text tries to steer the review. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the per-client speed vector $v_i^r = \beta_i v_i^{r-1} + (1-\tau_i)\Delta^r$, where $\Delta^r$ is the mean pseudo-gradient of the round and the adaptive coefficients are $\beta_i = \sigma(k(L^{\rm init}_{val,i}-L_{val,i})/L_{val,i})$ and $\tau_i = 1-(L_{train,i}/L_{val,i})^\alpha$. These coefficients turn a standard momentum update into one that trusts clients with improving validation loss and slows clients whose training loss has dropped far below their validation loss. The second component is the Foreground Intensity Matching loss, a 2-Wasserstein distance between the sorted intensities of the predicted and ground-truth foreground pixels, added to the segmentation loss as $\lambda_{\rm FIM} L_{\rm FIM}$. This term does the work of aligning intensity statistics across sites without sharing data or adding specialised network layers.
What would settle it
Re-run the prostate experiment with each client's validation pool cut to five volumes, or replace the validation-loss decrease with a shuffled client-specific signal of the same distribution; if FedCLAM's average Dice no longer beats HarmoFL, the adaptive-weight mechanism is not carrying the claimed gain.
Extended reading notes
Core claim
The paper's central claim is that a federated aggregator can read each client's validation-loss dynamics during a local round and use them to decide how heavily to trust that client's update, and that adding an intensity-distribution matching term to the segmentation loss reduces the harm of scanner-driven heterogeneity. Concretely, clients with steeper relative validation-loss decreases receive a larger momentum term $\beta_i=\sigma(k(L^{\rm init}_{val,i}-L_{val,i})/L_{val,i})$, clients whose training loss has fallen far below their validation loss are dampened by $\tau_i=1-(L_{train,i}/L_{val,i})^\alpha$, and the global model moves along the average of the resulting per-client speed vectors. The Foreground Intensity Matching loss adds a 2-Wasserstein penalty between the sorted intensities of predicted and ground-truth foreground pixels, pushing the network to ignore site-specific brightness and contrast cues. Under this mechanism, the paper reports average Dice of 88.82 on the four-center fundus task against 87.93 for the best baseline, and 93.74 on the six-center prostate MRI task against 93.61, with the best or near-best cross-client standard deviation.
Load-bearing premise
The load-bearing assumption is that a client's per-round validation loss is a dependable, comparable measure of genuine improvement; on the smallest sites, with roughly a dozen validation volumes, that number is mostly noise.
Editorial extensions
If this is right
- If the reported gains hold, weighting clients by validation progress rather than by fixed momentum or training loss alone is a viable improvement to federated aggregation.
- The FIM loss is a separable component: the ablation shows it improves FedAvg on its own, so other aggregation schemes could adopt it without adopting CLAM.
- FedCLAM still outperforms FedAvg when FIM is removed, indicating the adaptive momentum carries most of the benefit and intensity alignment adds on top.
- The default hyperparameters ($k=1$, $\lambda_{\rm FIM}=10^{-2}$) give consistent results across both tasks, which lowers the tuning burden for new medical FL deployments.
- Lower cross-client standard deviation in the fundus results suggests the weighting scheme helps under-performing institutions more than uniform aggregation does.
Reading between the lines
- Beyond the reported experiments, the validation-loss signal may need smoothing or multi-round averaging on sites as small as the 12-volume prostate clients, because per-round validation loss at that size is noisy enough to misweight clients.
- Extending FIM to multi-class segmentation would require aligning foreground intensities per anatomical class; the current single-foreground formulation could be confounded when a structure has heterogeneous internal intensities.
- The adaptive-momentum rule is not image-specific in principle, but carrying it to non-imaging federated tasks would first require checking that client loss scales are comparable.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. FedCLAM proposes two additions to federated medical image segmentation: (1) a client-adaptive momentum aggregation rule in which per-client momentum beta_i and dampening tau_i are derived from the relative decrease in validation loss and the train/validation loss ratio, and (2) a Foreground Intensity Matching (FIM) loss that penalizes the 2-Wasserstein distance between predicted and ground-truth foreground intensity distributions. The method is evaluated on a 4-client fundus segmentation task and a 6-client prostate MRI task, with U-Net as backbone, against eight FL baselines. The reported results show higher average Dice than all baselines on both tasks (88.82 vs 87.93 best baseline on fundus; 93.74 vs 93.61 on prostate) and a component ablation (Table 3) indicating that both CLAM and FIM contribute. The code is publicly released.
Significance. The proposed direction is relevant: aggregation rules that adapt to per-client training dynamics and intensity-alignment losses that do not require data sharing are both useful building blocks for cross-silo medical FL. The paper is clearly written, the method is lightweight, and the ablation supports that each component helps on the reported runs. The release of code and the use of two real multicentre datasets are strengths. However, the headline 'surpasses eight cutting-edge methods' claim is not yet supported by the evidence as presented: the reported margins are small, no variance or significance information is given, and the fairness claim is in part contradicted by the reported standard deviations. The method's central mechanism also needs clarification because the speed-vector update uses the global pseudo-gradient rather than per-client updates.
major comments (4)
- [§4.2, Tables 1–2] The abstract's central claim that FedCLAM 'surpasses eight cutting-edge methods' rests on average Dice differences of +0.89 (fundus) and +0.13 (prostate) over the best baseline. No repeated-seed statistics, confidence intervals, or significance tests are reported anywhere. With prostate clients as small as 12 volumes, such sub-1-point differences are within plausible run-to-run variation. Please add multiple seeds with variance estimates and a significance test (or an explicit statement that the differences are not statistically significant), and temper the abstract and conclusion accordingly. The ablation in Table 3 supports the method's plausibility, but not the headline superiority claim.
- [Abstract and §4.2, Table 2] The claim of 'better fairness' is contradicted by Table 2: on the prostate task FedCLAM's cross-client std of 1.96 is higher than FedEvi's 1.69, FedFA's 1.65, and FedSAM's 1.74. The text in §4.2 correctly says 'very competitive fairness' for prostate, but the abstract and introduction state 'better fairness across clients than state-of-the-art FL methods.' Please align the claims with the reported numbers, or provide evidence for fairness gains using a different fairness metric.
- [§3.2, Eq. (3), Algorithm 1 lines 11–18] In the speed-vector update, Delta^r is defined as the average of local updates (the global pseudo-gradient), and the same Delta^r is fed to every client's speed vector v_i^r. Thus a client's own local update Delta_i^r appears only inside the global average; client-specificity enters solely through beta_i and tau_i acting on shared past global pseudo-gradients. This does not implement the stated goal that 'clients with greater validation improvements guide the global model more prominently.' If the intended update is v_i^r = beta_i * v_i^{r-1} + (1 - tau_i) * Delta_i^r, please correct Eq. (3) and Algorithm 1; if the global Delta is intended, please explain concretely how per-client momentum states provide client-adaptive aggregation beyond a time-varying global momentum coefficient.
- [§3.2, Eqs. (1)–(2), §4.1] The adaptive signals beta_i and tau_i are computed from validation (and training) losses on each client. Prostate clients have as few as 12 volumes, so per-round validation losses are likely noisy; the paper provides no analysis of how sensitive FedCLAM's aggregation is to validation-set size, validation split, or cross-client loss calibration. Please add a sensitivity analysis (e.g., varying the validation fraction on the prostate data, or showing per-client loss trajectories) to support that the mechanism is reliable in the small-client regime.
minor comments (5)
- [Algorithm 1, line 15] The initialization 'v_i^r = Delta t' should be 'v_i^r = Delta^0'; the notation Delta t is undefined.
- [§4.1, Datasets] The prostate MRI dataset description lists five volume counts (30, 30, 19, 12, 12) for six centers; please correct the missing or misprinted count.
- [§3.3, Eq. (4)] The definition of n and the sorting procedure should be stated precisely; it is unclear how the soft weighting by predicted probability yields equal-length vectors when the foreground sizes of prediction and ground truth differ.
- [§4.3, Fig. 2] The sensitivity analysis is reported for the fundus dataset only; please report or motivate the absence of a comparable analysis for prostate and for the alpha hyperparameter.
- [Between §3.3 and §4] An unrelated reviewer-guidance passage appears in the submitted PDF between Section 3.3 and Section 4; it is not part of the technical content and should be removed.
Circularity Check
No circularity found: FedCLAM's components are evaluated empirically against external baselines and held-out test metrics, with no fitted parameter renamed as a prediction.
full rationale
FedCLAM's derivation chain is not circular. The client-adaptive momentum in Eq. 1 is computed from the relative decrease in each client's validation loss before and after local training, and the dampening factor in Eq. 2 is computed from the ratio of training loss to validation loss; these quantities are measured locally during training and only modulate the speed-vector average in Eq. 3. The reported outcome is test-set Dice in Tables 1 and 2, which is independently assessed on held-out test data. Nothing in Eqs. 1-3 uses the reported test Dice, the average Dice, or the final global model's test performance as an input, so the method's advantage cannot reduce by construction to the quantities it computes. The FIM loss in Eqs. 4-5 is an additional training objective aligning predicted and ground-truth foreground intensity distributions, while the segmentation loss still contains Dice or cross-entropy terms, and its contribution is separately evaluated through the ablation in Table 3; the FIM term does not by definition guarantee the observed Dice improvement. Hyperparameters k and lambda_FIM are varied on the fundus dataset in Fig. 2 to recommend defaults, which is standard hyperparameter selection rather than fitting a parameter to the reported result and then calling it a prediction. The reference list contains no self-citations by the present authors that carry the load of the central claim; the method is justified by its own local-loss signals and by external baseline comparisons. Concerns about small client sizes, absent error bars, and sub-point Dice margins are statistical-evidence issues, not circularity. The central claim is therefore empirically evaluated rather than equivalent to its inputs by definition.
Assumptions & free parameters
free parameters (3)
- k (sigmoid steepness) =
1.0
- alpha (dampening exponent) =
1.0
- lambda_FIM (FIM loss weight) =
1e-2
assumptions (4)
- domain assumption Validation loss (L_init_val, L_val) is a reliable signal of client contribution quality
- domain assumption Matching predicted and ground-truth foreground intensity distributions reduces site-specific bias
- domain assumption Clients have labeled local validation sets available each round
- domain assumption U-Net with Instance Norm is an adequate segmentation backbone for both tasks
Cite this review
Pith. "Pith review of FedCLAM: Client Adaptive Momentum with Foreground Intensity Matching for Federated Medical Image Segmentation." pith.science (2026). https://pith.science/paper/K7LK6USK
@misc{pith2026250622580,
author = {Pith},
title = {Pith review of: FedCLAM: Client Adaptive Momentum with Foreground Intensity Matching for Federated Medical Image Segmentation},
year = {2026},
howpublished = {\url{https://pith.science/paper/K7LK6USK}},
note = {Machine review of arXiv:2506.22580}
}
read the original abstract
Federated learning is a decentralized training approach that keeps data under stakeholder control while achieving superior performance over isolated training. While inter-institutional feature discrepancies pose a challenge in all federated settings, medical imaging is particularly affected due to diverse imaging devices and population variances, which can diminish the global model's effectiveness. Existing aggregation methods generally fail to adapt across varied circumstances. To address this, we propose FedCLAM, which integrates \textit{client-adaptive momentum} terms derived from each client's loss reduction during local training, as well as a \textit{personalized dampening factor} to curb overfitting. We further introduce a novel \textit{intensity alignment} loss that matches predicted and ground-truth foreground distributions to handle heterogeneous image intensity profiles across institutions and devices. Extensive evaluations on two datasets show that FedCLAM surpasses eight cutting-edge methods in medical segmentation tasks, underscoring its efficacy. The code is available at https://github.com/siomvas/FedCLAM.
Figures
Reference graph
Works this paper leans on
-
[1]
Acar,D.A.E.,Zhao,Y.,Navarro,R.M.,Mattina,M.,Whatmough,P.N.,Saligrama, V.: Federated learning based on dynamic regularization. In: 9th International Con- ference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7,
work page 2021
-
[2]
Image Analysis and Stereology39(3), 161–167 (2020)
Batista, F.J.F., Diaz-Aleman, T., Sigut, J., Alayon, S., Arnay, R., Angel-Pereira, D.: Rim-one dl: A unified retinal image database for assessing glaucoma using deep learning. Image Analysis and Stereology39(3), 161–167 (2020)
work page 2020
-
[3]
In: International Conference on Med- ical Image Computing and Computer-Assisted Intervention
Chen, J., Ma, B., Cui, H., Xia, Y.: Fedevi: Improving federated medical image seg- mentation via evidential weight aggregation. In: International Conference on Med- ical Image Computing and Computer-Assisted Intervention. pp. 361–372. Springer (2024)
work page 2024
-
[4]
Cheng, Z., Huang, X., Wu, P., Yuan, K.: Momentum benefits non-iid federated learning simply and provably. In: The Twelfth International Conference on Learn- ing Representations (2024), https://openreview.net/forum?id=TdhkAcXkRi
work page 2024
-
[5]
CoRR abs/1909.06335 (2019), http: //arxiv.org/abs/1909.06335
Hsu, T.H., Qi, H., Brown, M.: Measuring the effects of non-identical data distri- bution for federated visual classification. CoRR abs/1909.06335 (2019), http: //arxiv.org/abs/1909.06335
arXiv 2019
-
[6]
In: Proceedings of the AAAI Conference on Artificial Intelligence
Jiang, M., Wang, Z., Dou, Q.: Harmofl: Harmonizing local and global drifts in federated learning on heterogeneous medical images. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 36, pp. 1087–1095 (2022)
work page 2022
-
[7]
arXiv:1912.04977 [cs, stat] (Dec 2019), http:// arxiv.org/abs/1912.04977, arXiv: 1912.04977
Kairouz, P., McMahan, H.B., Avent, B., Bellet, A., Bennis, M., Bhagoji, A.N., Bonawitz, K., Charles, Z., Cormode, G., Cummings, R., et al.: Advances and Open Problems in Federated Learning. arXiv:1912.04977 [cs, stat] (Dec 2019), http:// arxiv.org/abs/1912.04977, arXiv: 1912.04977
arXiv 1912
-
[8]
Li, X., Huang, K., Yang, W., Wang, S., Zhang, Z.: On the convergence of fedavg on non-iid data (2020), https://arxiv.org/abs/1907.02189
arXiv 2020
Show all 24 references
-
[9]
CoRR abs/2103.06030 (2021), https://arxiv.org/abs/2103.06030
Liu, Q., Chen, C., Qin, J., Dou, Q., Heng, P.: Feddg: Federated domain generaliza- tion on medical image segmentation via episodic learning in continuous frequency space. CoRR abs/2103.06030 (2021), https://arxiv.org/abs/2103.06030
2021 arXiv
-
[10]
IEEE Transactions on Medical Imaging 39(9), 2713–2724 (Sep 2020)
Liu, Q., Dou, Q., Yu, L., Heng, P.A.: Ms-net: Multi-site network for improving prostate segmentation with heterogeneous mri data. IEEE Transactions on Medical Imaging 39(9), 2713–2724 (Sep 2020). https://doi.org/10.1109/tmi.2020.2974574, http://dx.doi.org/10.1109/TMI.2020.2974...
2020
-
[11]
CoRR abs/2111.08649 (2021), https://arxiv.org/abs/2111.08649
Mächler, L., Ezhov, I., Kofler, F., Shit, S., Paetzold, J.C., Loehr, T., Wiestler, B., Menze, B.H.: Fedcostwavg: A new averaging for better federated learning. CoRR abs/2111.08649 (2021), https://arxiv.org/abs/2111.08649
2021 arXiv
-
[12]
In: Artificial intelligence and statistics
McMahan, B., Moore, E., Ramage, D., Hampson, S., y Arcas, B.A.: Communication-efficient learning of deep networks from decentralized data. In: Artificial intelligence and statistics. pp. 1273–1282. PMLR (2017)
2017
-
[13]
Medical image analysis59, 101570 (2020)
Orlando, J.I., Fu, H., Breda, J.B., Van Keer, K., Bathula, D.R., Diaz-Pinto, A., Fang, R., Heng, P.A., Kim, J., Lee, J., et al.: Refuge challenge: A unified frame- work for evaluating automated methods for glaucoma assessment from fundus pho- tographs. Medical image analysis59...
2020
-
[14]
Qu, Z., Li, X., Duan, R., Liu, Y., Tang, B., Lu, Z.: Generalized federated learning via sharpness aware minimization (2022), https://arxiv.org/abs/2206.02618
2022 arXiv
-
[15]
CoRR abs/2003.00295 (2020), https://arxiv.org/abs/2003.00295
Reddi, S.J., Charles, Z., Zaheer, M., Garrett, Z., Rush, K., Konečný, J., Kumar, S., McMahan, H.B.: Adaptive federated optimization. CoRR abs/2003.00295 (2020), https://arxiv.org/abs/2003.00295
2020 arXiv
-
[16]
In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, Oc- tober 5-9, 2015, proceedings, part III 18
Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, Oc- tober 5-9, 2015, proceedings, part III 18....
2015
-
[17]
CoRR abs/1812.06127 (2018), http://arxiv.org/abs/1812.06127
Sahu, A.K., Li, T., Sanjabi, M., Zaheer, M., Talwalkar, A., Smith, V.: On the convergence of federated optimization in heterogeneous networks. CoRR abs/1812.06127 (2018), http://arxiv.org/abs/1812.06127
2018 arXiv
-
[18]
Scientific Reports 10(1), 12598 (Jul 2020)
Sheller, M.J., Edwards, B., Reina, G.A., Martin, J., Pati, S., Kotrotsou, A., Milchenko, M., Xu, W., Marcus, D., Colen, R.R., Bakas, S.: Federated learn- ing in medicine: facilitating multi-institutional collaborations without sharing pa- tient data. Scientific Reports 10(1), ...
2020
-
[19]
Sheller, M.J., Reina, G.A., Edwards, B., Martin, J., Bakas, S.: Multi-institutional deep learning modeling without sharing patient data: A feasibility study on brain tumor segmentation. In: Brainlesion: Glioma, Multiple Sclerosis, Stroke and Trau- matic Brain Injuries: 4th Int...
2019
-
[20]
JSM Biomed Imaging Data Pap2 (01 2015)
Sivaswamy, J., Krishnadas, S., Chakravarty, A., Joshi, G., Ujjwal: A comprehensive retinal image dataset for the assessment of glaucoma from the optic nerve head analysis. JSM Biomed Imaging Data Pap2 (01 2015)
2015
-
[21]
CoRRabs/1607.08022 (2016), http://arxiv.org/ abs/1607.08022
Ulyanov, D., Vedaldi, A., Lempitsky, V.S.: Instance normalization: The missing ingredient for fast stylization. CoRRabs/1607.08022 (2016), http://arxiv.org/ abs/1607.08022
2016 arXiv
-
[22]
CoRRabs/2008.07665 (2020), https: //arxiv.org/abs/2008.07665
Yeganeh, Y., Farshad, A., Navab, N., Albarqouni, S.: Inverse distance aggregation for federated learning with non-iid data. CoRRabs/2008.07665 (2020), https: //arxiv.org/abs/2008.07665
2020 arXiv
-
[23]
In: The Eleventh International Conference on Learning Representations, ICLR 2023, Ki- gali, Rwanda, May 1-5, 2023
Zhou, T., Konukoglu, E.: Fedfa: Federated feature augmentation. In: The Eleventh International Conference on Learning Representations, ICLR 2023, Ki- gali, Rwanda, May 1-5, 2023. OpenReview.net (2023), https://openreview.net/ forum?id=U9yFP90jU0
2023
-
[2021]
OpenReview.net (2021), https://openreview.net/forum?id=B7v4QMR6Z9w
2021
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.