{"id":"cdf93de8-ee0a-4abc-8223-039a6de75e49","arxiv_id":"2504.18020","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A federated adapter decomposed into global and client-specific units improves heterogeneous medical image segmentation over conventional federated baselines.","lead":"This paper presents a federated learning method for medical image segmentation that gives each hospital a personalized model instead of one shared compromise. It works by splitting small adapter modules into client-common and client-specific parts and updating those parts separately, without sharing images.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 4 computes 'normalized entropy' on raw, possibly negative contribution scores, so the δ threshold and the global/local split are not well-defined; the client-tailored decomposition's validity is unestablished.","rationale":"The central claim is that FCA obtains client-tailored segmentation by decomposing adapter units into global and local parts using contribution scores from a client discriminator. The most load-bearing condition is that the decomposition procedure is a valid signal of client-specificity. The reader's weakest assumption identified exactly this dependency. My stress-test sharpens it: even if the discriminator and contribution scores S were meaningful, Eq. 4 does not compute a normalized entropy because S is not normalized and may contain negative values. Consequently, the threshold δ in Eq. 5 and the resulting binary/smooth updates have no well-defined scale or interpretation. This is an internal inconsistency, not merely a matter of disagreement with existing methods. It is addressable: the authors could re-define D with a proper normalization and re-run the experiments. However, the absence of code, error bars, and threshold sensitivity means the empirical support is currently conditional. I therefore agree with the reader's CONDITIONAL verdict; my concern strengthens the conditions rather than changing the verdict class.","tokens_in":17072,"tokens_out":3253,"duration_ms":36861,"concrete_test":"Request or reconstruct one trained FCA checkpoint and dump S and D for several adapter layers across clients. Check the sign and scale of S entries. Then recompute D using a properly normalized probability distribution, e.g. p_{u,k,j} = softmax_j(S_{u,k,j}) or max(S,0) followed by row normalization, and regenerate the BFU masks (Eq. 5) and SFU weights (Eq. 7). If the masks or aggregation weights differ materially, or if the mDice numbers in Tables I and II shift by more than 0.5 on any client, then the current formulation of Eqs. 4-5 is not the mechanism behind the reported results.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-D1's Eqs. 4-5 are not a valid entropy test as written. The contribution score S from Eqs. 2-3 is built from element-wise products and an outer product of channel activations F_d, classifier weights W_d, and client probabilities P. W_d is learned and can be negative, and F_d is not guaranteed nonnegative, so S entries can be negative. Equation 4 then evaluates Σ_j S_{u,k,j} log2 S_{u,k,j} and divides by log2 K, calling the result a normalized entropy. This is undefined for negative values and is not a valid entropy for unnormalized scores: entropy requires a probability distribution with nonnegative entries summing to 1. Depending on feature scale, D can be negative, exceed 1, or change arbitrarily, so the threshold δ = 0.25 in Eq. 5 has no defined meaning. Every downstream step—global/local masks in BFU (Eq. 6) and the aggregation weights in SFU (Eq. 7)—inherits this issue, so the reported gains cannot be attributed to a principled decomposition of common versus client-specific units. The paper also provides no normalization convention, no sensitivity analysis for δ, and no synthetic validation that S separates known client-specific from client-invariant units. The GRL description is additionally in tension with standard gradient reversal semantics: a true gradient reversal layer encourages domain-invariant features, which would suppress the client-specific signal that S is meant to measure.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes the Federated Client-tailored Adapter (FCA), a framework for federated medical image segmentation based on the SAM-Med2D foundation model. The central idea is to insert lightweight adapters into the frozen encoder and decompose adapter units into client-invariant 'global' and client-specific 'local' components using a contribution score derived from a client discriminator. Two updating strategies are introduced: a binary strategy (BFU) that globally averages global units and leaves local units untouched, and a smooth strategy (SFU) that uses the contribution scores as probabilistic aggregation weights. Experiments on three distributed datasets (CXRS-HG, HLS, AMD-SD-HG) report consistent improvements over several federated baselines and personalized methods, and the framework is shown to generalize across three medical foundation models.","tokens_in":17394,"tokens_out":9726,"duration_ms":92132,"significance":"If the reported results hold, the paper makes a useful empirical contribution to personalized federated medical segmentation, demonstrating that parameter-efficient adapters plus client-specific aggregation can improve mDice over strong MFM-based baselines, with consistent gains across datasets and foundation models. The paper also includes a computational-efficiency analysis. However, the theoretical basis of the core decomposition mechanism is not sound as written, and several experimental details are missing, so the significance of the contribution is tempered by reproducibility and validity concerns.","major_comments":[{"comment":"Equation (4) defines D_{u,k} = (Σ_j S_{u,k,j} log2 S_{u,k,j}) / log2 K and calls it a normalized entropy over the client-wise score distribution. This is not a valid entropy because S_{u,k,j} is not a probability distribution: it is not normalized to sum to 1, and its entries are not guaranteed to be nonnegative, since they are products of the learned classifier weights W_d and the pooled features F_d, both of which can take negative values. Consequently, the logarithm is undefined for negative entries, and D can be negative or exceed 1. The threshold δ = 0.25 in Eq. (5) therefore has no well-defined meaning, and the subsequent global/local mask in Eq. (6) and the aggregation weights in Eq. (7) inherit this ill-posedness. The paper provides no normalization convention, no sensitivity analysis for δ, and no synthetic experiment showing that S separates known client-specific from client-invariant units. Please either redefine D as a proper normalized entropy over a nonnegative vector that sums to one, or provide a rigorous justification for the current formula.","section":"§III-D1, Eq. (4)"},{"comment":"The paper does not specify how the per-client contribution scores are aggregated to form the three-index array S_{u,k,j} used in Eqs. (4) and (7). In Section III-C, S(i) is computed for each sample and averaged locally, and the text states that 'the final contribution score S will be delivered to the central server and then broadcast to every client.' However, Eq. (7) uses S_{u,k,j} with both a source-client index k and a target-client index j, and the manuscript does not explain whether S_{u,k,j} is the score uploaded by client k, how server-side aggregation of these scores is performed, or how the algorithm handles normalization. Algorithm 1 likewise omits the server-side integration step for S. This makes the BFU and SFU updating rules non-reproducible as written.","section":"§III-D, Algorithm 1"},{"comment":"The description of the Gradient Reversal Layer is internally inconsistent. The text says the GRL is inserted 'to truncate the gradients from the above discrimination loss,' which is a stop-gradient operation, but the name 'Gradient Reversal Layer' in the literature denotes a layer that reverses gradients during backpropagation. If a true reversal is used, the discriminator would encourage domain-invariant features, which would suppress the client-specific signal that the contribution score S is intended to measure. If a stop-gradient is used, the layer should be described as such, and the implications for the validity of S as a client-specificity measure should be discussed. Please clarify the exact forward/backward behavior of this layer and justify that the resulting S is a meaningful estimator of unit-level client-specificity.","section":"§III-C, GRL description"},{"comment":"The HLS dataset is described as comprising four independent subsets (references [53]–[56]), but Table I reports results for exactly three clients on this dataset. The mapping from the four subsets to the three clients is not described, which prevents reproducibility of the HLS experiments. Please specify which subset(s) are assigned to each client, or how the fourth subset is used (e.g., held out, merged, or split).","section":"§IV.A.2 and Table I"}],"minor_comments":[{"comment":"The text 'our FCA-SBU achieves more precision segmentation' contains a typo: it should read 'our FCA-SFU.'","section":"§IV.C, text before Fig. 5"},{"comment":"The phrase 'sets a new start-of-art result' should read 'sets a new state-of-the-art result.'","section":"§IV.D"},{"comment":"The symbol W_d is used both for the discriminator network (in Eq. (1)) and for the classifier weight matrix (in Eqs. (2)–(3)); please use distinct notation to avoid ambiguity.","section":"§III-C, Eq. (1)"},{"comment":"The sentence 'If the weighted contribution scores are similar across all clients, it suggests that the parameter group in this unit is client-specific and thus can be considered a global parameter' appears to state the opposite of the intended meaning; 'client-specific' should likely read 'client-invariant.'","section":"§III-D1"},{"comment":"The list of compared methods includes 'IOP-FL [17]' twice; please remove the duplicate.","section":"§IV.C"},{"comment":"All main tables report single-run results without standard deviations; please report mean and standard deviation over multiple seeds, especially given the small margins on HLS, to support the superiority claims.","section":"Tables I–IV"}],"recommendation":"major_revision","confidential_remarks":"The empirical direction is promising and the scope fits the journal, but the method section contains a mathematically ill-defined core (Eq. (4)) and an underspecified aggregation of contribution scores, both of which must be fixed before the claims can be fully evaluated. I would advise requiring the authors to resolve these issues and to provide error bars before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the FCA paper is worth a serious look—the SFU aggregation is a genuine, if incremental, mechanism and the reported gains are consistent across all comparisons on three datasets. But the load-bearing decomposition score (Eq. 4) is not a valid entropy as written, and the paper needs error bars and a sensitivity analysis before the SOTA claim is credible.\n\nWhat is new: unit-level decomposition of a federated adapter into global and local parts, with a client-discriminator contribution score and a smooth probabilistic weighting in SFU. The paper positions this against coarse-grained decoupling (CD2-pFed, MAP, IOP-FL) and the comparison is honest. The consistent 1.2–3.5 mDice improvements over same-backbone baselines, plus the heterogeneity ablation in Table III, show the method works directionally.\n\nSoft spots, in proportion:\n\n(1) Eq. 4–5: the normalized-entropy test is undefined for negative S entries, and S from Eqs. 2–3 can be negative because W_d is learned and F_d is not guaranteed nonnegative. Calling the result an entropy and dividing by log2 K does not make it one. So δ=0.25 has no defined meaning, and the global/local split is not well-grounded. No sensitivity analysis for δ, no synthetic validation that S separates known client-specific from client-invariant units. This doesn't sink the empirical results, but it means the claimed decomposition mechanism is unestablished.\n\n(2) The GRL description is confused: standard GRL flips gradients and would actively suppress the client-specific signal, while the paper says it 'truncates' gradients so the auxiliary branch does not affect the main branch. If they mean stop-gradient, they should say so.\n\n(3) Evaluation hygiene: single runs, no error bars, no code, and HLS is described as four subsets but experiments report three clients. Addressable, but they undercut the SOTA claim.\n\nMinor: a sentence in Section III-D1 calls scores that are similar across clients both 'client-specific' and 'global'; the entropy math says they mean global, so it's likely a typo.\n\nThe stress-test note on Eq. 4 lands. I don't think it's a fatal flaw because the empirical wins are consistent, but it needs fixing. The central argument holds up directionally.\n\nWho this is for: researchers combining PEFT with personalized FL for medical segmentation. A serious referee should engage. I'd recommend conditional acceptance with required error bars, a corrected or clarified Eq. 4, a GRL/stop-gradient clarification, and dataset-client alignment.","headline":"Consistent gains from a plausible unit-level adapter decomposition, but the entropy-based split (Eq. 4) is mathematically undefined as written and the paper needs error bars before the SOTA claim is taken seriously.","tokens_in":17930,"tokens_out":3978,"would_cite":true,"duration_ms":38030,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Federated adapters split shared and private units to beat global models","keywords":["federated learning","medical image segmentation","parameter-efficient fine-tuning","client-tailored personalization","adapter decomposition","domain heterogeneity","X-ray chest images","smooth federated updating"],"falsifier":"Train FCA on synthetic clients that differ only in a clinically irrelevant factor (e.g., global brightness or mild blur) while keeping the anatomical labels and distributions identical; if the discriminator still pushes many units into the local set and the local/global split visibly changes the segmentation outputs, the score is capturing nuisance artifacts. A direct check is to compare the global-unit mask selected by Eq. 5 with the mask obtained after permuting client labels during discriminator training: if the selected units change substantially, the decomposition is not a stable evidence of client-specificity.","tokens_in":16890,"feed_emoji":"🏥","tokens_out":7184,"duration_ms":69105,"temperature":0.7,"pith_summary":"Conventional federated learning for medical image segmentation produces a single globally averaged model that trains unstably and converges slowly when hospitals hold very different data distributions. This paper proposes the Federated Client-tailored Adapter (FCA), which keeps a large medical foundation model frozen, inserts lightweight adapters, and splits each adapter into shared client-invariant units and private client-specific units. A client discriminator computes a contribution score for every unit, and the paper uses that score to decide which parameters are updated by global federated averaging and which are updated only on the client. Two strategies implement this split: a binary one and a smooth probabilistic one. The paper claims that FCA stabilizes training, gives each client a tailored model instead of a compromised one, and achieves state-of-the-art results on three large-scale distributed medical segmentation datasets.","feed_headline":"Federated adapters split shared and private units to beat global models","feed_subtitle":"Hospitals keep a tailored model while sharing only client-invariant adapter weights; smooth allocation works best.","key_machinery":"The load-bearing object is the Global-local Decomposer (GLD), a branch attached to every adapter layer (GLDConv for channels, GLD-FC for neurons) that turns a pretext client-discrimination task into a per-unit contribution score. For an input feature, global average pooling gives $F_d$, and the discriminator's classifier weights $W_d$ and predicted client probabilities are combined as $S = \\mathrm{broadcast}(F_d) \\odot W_d$ and $S = \\hat{S} \\otimes P(\\hat{y} \\mid F_{\\mathrm{in}})$ (Eqs. 2-3). The normalized entropy of $S$ over clients (Eq. 4) is then used to mask units as global or local in the binary strategy, or to weight client-to-client parameter updates in the smooth strategy (Eqs. 5-7). A gradient reversal layer keeps the discriminator loss from perturbing the segmentation features. This machinery is what lets the framework keep one global update channel for shared knowledge and one local update channel for personalization.","core_discovery":"The central claim is that client-wise heterogeneity in federated medical segmentation is best handled not by making one global model more robust, but by explicitly separating each client's parameters into a globally shared subset and a locally held subset. To find that split, the framework trains a client discriminator on average-pooled features and computes, for each adapter channel or neuron, a contribution score $S$ that says how strongly that unit votes for each client identity. Units whose scores are near-uniform across clients are treated as client-invariant and federated globally; units whose scores concentrate on one or few clients are treated as client-specific and updated locally. The binary strategy (BFU) hard-thresholds the normalized entropy of $S$, while the smooth strategy (SFU) uses $S$ directly as weights in the aggregation formula. On the CXRS-HG, HLS, and AMD-SD-HG datasets, the paper reports mean Dice scores of 64.15, 92.44, and 63.40 for FCA-SFU, exceeding the MFM-enhanced FedAvg baselines by 3.51, 1.65, and 2.98 points respectively and outperforming all compared personalized federated baselines.","pith_inferences":["The same contribution-score idea could transfer to other parameter-efficient structures (LoRA, prompts, or bias terms) and to federated classification or detection, since nothing in the mechanism is segmentation-specific.","The discriminator's score $S$ is trained to separate clients by any signal available, so it may treat easy-to-see acquisition differences (brightness, blur, resolution) as client-specific even when those differences carry no anatomical meaning; a test with controlled nuisance-only shifts would reveal how much of the decomposition is domain specificity versus clinically meaningful personalization.","The fixed entropy threshold $\\delta = 0.25$ in the binary strategy is an empirical choice; a per-layer or adaptive threshold, or a warmup schedule that lets the discriminator stabilize before fixing the split, could be tested against the reported results."],"forward_implications":["Each hospital deploying FCA ends up with its own model tuned to its own data distribution, so federation no longer forces a single compromise model on everyone.","The decomposition is agnostic to the specific medical foundation model: the paper reports consistent gains when the same FCA-SFU pipeline is built on SAM-Med2D, H-SAM, or Med-SA.","Because only lightweight adapter parameters are split and shared, the overhead over the plain MFM baseline is small (about 0.09M additional parameters out of 180.5M).","The smooth strategy's consistent advantage over the binary one suggests that soft, contribution-weighted aggregation is a better recipe for heterogeneous medical data than hard global/local assignment.","The instability problem reported for conventional FL under class imbalance and distribution diversity is reduced: on the CXRS-HG variant, adding full heterogeneity costs FCA-SFU almost nothing, while it hurts the FedAvg baseline substantially."],"supporting_citations":[{"why":"Supplies the vanilla FedAvg federated averaging protocol and the main baseline that the proposed client-tailored updating strategies modify.","marker":"[15]"},{"why":"Provides SAM-Med2D, the medical foundation model and adapter structure on which the FCA framework is built and evaluated.","marker":"[10]"},{"why":"IOP-FL is the strongest personalized federated segmentation baseline that FCA compares against in the main tables.","marker":"[17]"},{"why":"FedSeg is the class-heterogeneous federated segmentation baseline used to show the challenge of client-wise class imbalance.","marker":"[18]"},{"why":"FedProx is a proximal-term federated optimization baseline included in the comparisons.","marker":"[34]"},{"why":"HarmoFL is a heterogeneity-targeting federated method included as a baseline for medical image segmentation.","marker":"[46]"},{"why":"PerFedAvg is a personalized federated baseline compared in the main experiments.","marker":"[47]"},{"why":"FedCross is a cross-client personalized federated baseline used in both main datasets and the AMD-SD-HG comparison.","marker":"[48]"},{"why":"MAP is a model aggregation and personalization baseline that decouples global and local components, a direct comparator for FCA's decoupling idea.","marker":"[49]"},{"why":"H-SAM is one of the alternative medical foundation models used to test the generalization ability of FCA.","marker":"[58]"}],"fun_headline_variants":["Federated adapter splits shared and private units for stable segmentation","Client-tailored adapters outperform global models in federated segmentation","Splitting adapters into common and individual parts stabilizes federated learning","FCA: personalizing federated segmentation via adapter unit splitting"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole framework rests on trusting that the contribution score computed from the client discriminator is a genuine measure of whether each adapter unit is client-specific, rather than an artifact of easy-to-tell-apart image differences or of discriminator overfitting.","fun_headline_variants_meta":{"raw":{"variants":["Federated adapter splits shared and private units for stable segmentation","Client-tailored adapters outperform global models in federated segmentation","Splitting adapters into common and individual parts stabilizes federated learning","FCA: personalizing federated segmentation via adapter unit splitting"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000639,"raw_usage":{"total_tokens":2970,"prompt_tokens":1002,"completion_tokens":1968,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":1894}},"tokens_in":618,"tokens_out":1968,"duration_ms":15863,"temperature":1.0,"reasoning_tokens":1894,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:26:32.802976+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train FCA on synthetic clients that differ only in a clinically irrelevant factor (e.g., global brightness or mild blur) while keeping the anatomical labels and distributions identical; if the discriminator still pushes many units into the local set and the local/global split visibly changes the segmentation outputs, the score is capturing nuisance artifacts. A direct check is to compare the global-unit mask selected by Eq. 5 with the mask obtained after permuting client labels during discriminator training: if the selected units change substantially, the decomposition is not a stable evidence of client-specificity.","supporting_citations":[{"cited_title":"Iop-fl: inside-outside person- alization for federated medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"IOP-FL is the strongest personalized federated segmentation baseline that FCA compares against in the main tables."},{"cited_title":"Fedseg: Class-heterogeneous federated learning for semantic segmentation,","cited_arxiv_id":null,"evidence_quote":"FedSeg is the class-heterogeneous federated segmentation baseline used to show the challenge of client-wise class imbalance."},{"cited_title":"Harmofl: Harmonizing local and global drifts in federated learning on heterogeneous medical images,","cited_arxiv_id":null,"evidence_quote":"HarmoFL is a heterogeneity-targeting federated method included as a baseline for medical image segmentation."},{"cited_title":"Personalized and privacy-preserving federated heterogeneous medical image analysis with pppml-hmi,","cited_arxiv_id":null,"evidence_quote":"PerFedAvg is a personalized federated baseline compared in the main experiments."},{"cited_title":"Federated cross learning for medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"FedCross is a cross-client personalized federated baseline used in both main datasets and the AMD-SD-HG comparison."},{"cited_title":"Map: model aggregation and personalization in federated learning with incomplete classes,","cited_arxiv_id":null,"evidence_quote":"MAP is a model aggregation and personalization baseline that decouples global and local components, a direct comparator for FCA's decoupling idea."},{"cited_title":"Unleashing the potential of sam for medical adaptation via hierarchical decoding,","cited_arxiv_id":null,"evidence_quote":"H-SAM is one of the alternative medical foundation models used to test the generalization ability of FCA."}],"review_version":1}