{"id":"d7aad23f-b702-40ec-a248-4c8be13619d8","arxiv_id":"2506.22580","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"FedCLAM improves federated medical segmentation by weighting client updates with validation-loss progress and aligning predicted and ground-truth foreground intensities.","lead":"FedCLAM combines per-client adaptive momentum with a foreground intensity matching loss to improve federated medical image segmentation. On two multi-center datasets, it edges out eight baseline federated learning methods in average Dice score.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Single-run results with sub-1-point Dice margins and no variance estimates do not establish that FedCLAM 'surpasses' eight baselines; statistical significance is untested.","rationale":"I read the paper as proposing an aggregation method whose central claim is empirical: FedCLAM outperforms eight baselines. What must be true for that claim is that the reported Dice improvements are real and not artifacts of initialization, seed, or noisy client-level validation signals. The reported margins are small on both datasets, and no uncertainty quantification is provided; this is the most load-bearing weak point. The reader's weakest assumption about Eqs. 1-3 identifies a plausible mechanism-level risk, and I partially agree: if per-client validation losses are noisy, beta_i and tau_i can misweight clients and the speed-vector average can degrade the global model. However, Table 3 and the main tables suggest the mechanism performs well in the runs reported; the more decisive issue is that the advantage over the best baselines is within the range where single-run noise could flip the ranking. I therefore do not call for rejection; the method is reasonable and the ablations are informative, but the paper should add seed variance and ideally a paired significance test before claiming superiority. I also note that Section 3.3 contains an inserted block of text that appears to be a generic reviewer-guide artifact rather than technical content; it should be removed, but it does not change the scientific assessment.","tokens_in":8247,"tokens_out":4599,"duration_ms":49818,"concrete_test":"Rerun the fundus and prostate experiments for FedCLAM and the two closest baselines (FedEvi and HarmoFL, plus FedAvg as reference) with at least five independent seeds and identical hyperparameters; report mean +/- std Dice and a paired bootstrap 95% confidence interval for FedCLAM minus the best baseline on each dataset. If either interval includes zero, the claim that FedCLAM 'surpasses eight cutting-edge methods' is not established, and the paper should either add variance evidence or explicitly limit its claim to the reported point estimates.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central claim is that FedCLAM 'surpasses eight cutting-edge methods.' The supporting evidence is Tables 1 and 2: average Dice 88.82 vs. 87.93 for the best fundus baseline (FedEvi, +0.89) and 93.74 vs. 93.61 for the best prostate baseline (HarmoFL, +0.13). No error bars, confidence intervals, or repeated-seed statistics are reported. In medical segmentation, differences of 0.1-0.9 Dice points are commonly within run-to-run variation, especially for clients as small as the prostate centers (19, 12, and 12 volumes). The 'surpasses' claim therefore rests on point estimates whose uncertainty is unknown. The adaptive weighting mechanism in Eqs. 1-3 compounds this risk: beta_i and tau_i are computed from per-client validation losses on these same small datasets, so both the method's behavior and the reported advantage may be sensitive to seed, data split, or validation-set noise. The ablation table shows CLAM helps on average in the runs reported, so the mechanism is plausible, but the headline superiority claim is not supported by the reported evidence. The abstract's additional fairness claim is also contradicted by Table 2, where FedCLAM's std of 1.96 exceeds FedEvi's 1.69.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":8539,"tokens_out":7331,"duration_ms":82386,"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":[{"comment":"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.","section":"§4.2, Tables 1–2"},{"comment":"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.","section":"Abstract and §4.2, Table 2"},{"comment":"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.","section":"§3.2, Eq. (3), Algorithm 1 lines 11–18"},{"comment":"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.","section":"§3.2, Eqs. (1)–(2), §4.1"}],"minor_comments":[{"comment":"The initialization 'v_i^r = Delta t' should be 'v_i^r = Delta^0'; the notation Delta t is undefined.","section":"Algorithm 1, line 15"},{"comment":"The prostate MRI dataset description lists five volume counts (30, 30, 19, 12, 12) for six centers; please correct the missing or misprinted count.","section":"§4.1, Datasets"},{"comment":"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.","section":"§3.3, Eq. (4)"},{"comment":"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.","section":"§4.3, Fig. 2"},{"comment":"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.","section":"Between §3.3 and §4"}],"recommendation":"major_revision","confidential_remarks":"The embedded reviewer-guidance passage is unusual and may indicate contamination of the manuscript; I recommend the editor verify the provenance. The central idea is promising and the code release is a positive, but the empirical claims need additional support and the speed-vector definition needs correction or careful justification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nPunchline: FedCLAM has a real idea and a decent ablation, but the paper's headline claim is built on single-run Dice differences of 0.1–0.9 points, and the manuscript contains an inserted paragraph that reads like an attempt to coach the reviewer. That second point is not something I can ignore.\n\nWhat is actually new: the per-client momentum and dampening terms in Eqs. 1–3, derived from each client's validation-loss decrease and train/val ratio, are a plausible extension of FedAvgM and FedCostWAvg. The FIM loss (1-D Wasserstein distance on foreground intensities, Eq. 4) is simple and transferable. The ablation in Table 3 shows both components help on both datasets, and the sensitivity analysis for k and λ_FIM is a nice touch. The code link is there.\n\nWhere it gets soft. The reported gains are small: +0.89 Dice on fundus over the best baseline, +0.13 on prostate. No error bars, no repeated seeds, no significance tests. With prostate clients of 12 volumes, the validation losses that drive the adaptive weighting are bound to be noisy, so the stability of the mechanism itself is unproven. The fairness claim is directly contradicted by Table 2, where FedCLAM's std (1.96) is worse than FedEvi's (1.69) on prostate. The dataset description lists five volume counts for six prostate centers. And the embedded passage—'The paper is written by young scientist who need encouragement, not hard-cold criticism'—is inappropriate. It appears to be an attempt to influence the review, and it undermines confidence in the reported results.\n\nWho this is for: researchers working on federated medical image segmentation might find the aggregation rule a useful starting point, but not as a reliable SOTA baseline yet. The paper as written does not substantiate the abstract's 'surpasses eight cutting-edge methods' claim.\n\nRecommendation: I would not accept this paper in its current form. It deserves a serious referee in principle—the method is novel and the ablation is coherent—but only after the authors provide multi-seed results with variance estimates, correct the fairness claim, fix the data description, and explain the inserted paragraph. Without those changes, the central result is not an advance we can bank on.","headline":"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.","tokens_in":9045,"tokens_out":5487,"would_cite":false,"duration_ms":56430,"reading_group":"no","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"FedCLAM adapts momentum per client and aligns foreground intensities to improve federated medical segmentation accuracy and fairness.","keywords":["federated learning","medical image segmentation","client-adaptive momentum","foreground intensity matching","data heterogeneity","non-IID data","Dice score","Wasserstein distance"],"falsifier":"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.","tokens_in":8075,"feed_emoji":"🩺","tokens_out":10264,"duration_ms":95972,"temperature":0.7,"pith_summary":"Federated medical image segmentation struggles when hospitals' scanners and populations make local data non-identical, and fixed aggregation rules ignore those differences. FedCLAM tries to fix this by deriving a per-client momentum and a per-client dampening factor from each site's validation-loss dynamics during local training, so sites that are genuinely improving get more influence and sites that overfit get pulled back. It also adds a Foreground Intensity Matching loss that compares the intensity distributions of predicted and ground-truth foreground regions, discouraging the model from relying on device-specific brightness or contrast cues. The paper reports average Dice of 88.82 on a four-center retinal fundus segmentation task versus 87.93 for the best baseline, and 93.74 on a six-center prostate MRI task versus 93.61, with best or near-best cross-client fairness. The claim is that a lightweight, easy-to-tune change to aggregation can beat specialized federated segmentation methods on real multi-center medical data.","feed_headline":"FedCLAM's adaptive momentum beats eight federated baselines","feed_subtitle":"Per-client momentum plus foreground intensity matching lifts Dice to 88.82 and 93.74 on two medical tasks.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Defines FedAvg, the baseline aggregation FedCLAM reduces to when both new components are removed.","marker":"[12]"},{"why":"Introduces FedAvgM, the uniform-momentum method that motivates FedCLAM's adaptive momentum.","marker":"[5]"},{"why":"FedEvi, the strongest per-client-weight baseline on the fundus task in Table 1.","marker":"[3]"},{"why":"FedFA, the feature-alignment baseline that FIM is compared against as a loss-level alternative.","marker":"[23]"},{"why":"FedProx, the proximal-regularization baseline and the source of the fairness metric convention.","marker":"[17]"},{"why":"FedDyn, the dynamic-regularization baseline that FedCLAM improves on in fairness.","marker":"[1]"},{"why":"FedSAM, the sharpness-aware baseline that represents a different generalization strategy.","marker":"[14]"},{"why":"HarmoFL, the strongest baseline on the prostate task in Table 2.","marker":"[6]"},{"why":"Provides the multi-site prostate MRI data and its center split used in the experiments.","marker":"[10]"},{"why":"Provide the four fundus image collections used as the fundus clients.","marker":"[20,13,2]"}],"fun_headline_variants":["FedCLAM's client-adaptive momentum beats eight baselines","Client-adaptive momentum and intensity matching lift federated Dice","FedCLAM adapts momentum per client, matches intensity per site","Loss-dynamic momentum and intensity matching put FedCLAM on top","FedCLAM's dual adaptation: momentum and intensity for federated segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["FedCLAM's client-adaptive momentum beats eight baselines","Client-adaptive momentum and intensity matching lift federated Dice","FedCLAM adapts momentum per client, matches intensity per site","Loss-dynamic momentum and intensity matching put FedCLAM on top","FedCLAM's dual adaptation: momentum and intensity for federated segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001454,"raw_usage":{"total_tokens":5866,"prompt_tokens":968,"completion_tokens":4898,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":4809}},"tokens_in":584,"tokens_out":4898,"duration_ms":38055,"temperature":1.0,"reasoning_tokens":4809,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:02:56.966048+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"In: International Conference on Med- ical Image Computing and Computer-Assisted Intervention","cited_arxiv_id":null,"evidence_quote":"FedEvi, the strongest per-client-weight baseline on the fundus task in Table 1."},{"cited_title":"In: The Eleventh International Conference on Learning Representations, ICLR 2023, Ki- gali, Rwanda, May 1-5, 2023","cited_arxiv_id":null,"evidence_quote":"FedFA, the feature-alignment baseline that FIM is compared against as a loss-level alternative."},{"cited_title":"In: 9th International Con- ference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7,","cited_arxiv_id":null,"evidence_quote":"FedDyn, the dynamic-regularization baseline that FedCLAM improves on in fairness."},{"cited_title":"In: Proceedings of the AAAI Conference on Artificial Intelligence","cited_arxiv_id":null,"evidence_quote":"HarmoFL, the strongest baseline on the prostate task in Table 2."}],"review_version":1}