{"id":"4b985224-3d25-4788-8b16-35843dc83b37","arxiv_id":"2502.00619","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Attribute-conditioned mixture-of-experts gating (dMoE) improves subgroup equity in medical image segmentation, framed as mode-switching optimal control.","lead":"A medical image segmentation method, dMoE, feeds patient attributes like race, age, or tumor stage into a mixture-of-experts router to improve fairness across subgroups. The authors frame the design as an optimal control problem and report state-of-the-art fairness-scaled Dice scores on three segmentation datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported CIs do not exclude plain MoE on most ES metrics, so the attribute-conditioned router—the paper's core novelty—is not statistically shown to improve fairness segmentation.","rationale":"The reader's weakest_assumption correctly identifies a real gap: Eq. 16 requires f to be linear in its second argument, but the implemented experts are two-layer ReLU MLPs, so the control-theoretic equality does not hold for the architecture. I agree this reduces the 'mode-switching optimal control' statement to an analogy rather than a derivation. I do not make it the headline concern because the experiments in Sec. 4 do not rely on that derivation, and the authors place an explicit 'if' before Eq. 16. The load-bearing question for the empirical claim is whether dMoE improves over MoE by more than sampling noise. The reported marginal CIs overlap for every ES metric, and no paired difference inference is supplied, so the attribute-conditioned router—the only difference from MoE—is not statistically established as the source of the gains. This is not an accusation of misconduct; it is a request for the inference implied by 'significantly outperforming.' The paired cluster bootstrap described above would settle it. Since the manuscript needs that additional analysis regardless of the theory issue, I keep the reader's CONDITIONAL verdict unchanged.","tokens_in":18457,"tokens_out":9840,"duration_ms":87529,"concrete_test":"Run a paired cluster bootstrap over patients: for 1,000 resamples, resample patients with replacement, recompute ES-Dice and subgroup Dice for dMoE and MoE on the identical resamples, and form percentile 95% CIs for the paired difference (dMoE - MoE) on each dataset and attribute. Apply Holm-Bonferroni correction across the three datasets and all reported subgroups, and add CIs for T1 and T4 in Table 3. If the corrected interval excludes 0 for most comparisons, the attribute-conditioned gating is validated; if it includes 0, the distinctive dMoE contribution is not statistically demonstrated and the paper should present the method as a MoE variant with point-estimate gains only.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The comparison isolating the 'distribution-aware' contribution is dMoE vs. the default MoE baseline (Sec. 4.3). In Table 1 rim, ES-Dice is 0.743 (0.723-0.763) for dMoE vs. 0.733 (0.713-0.752) for MoE; in Table 2, ES-Dice is 0.801 (0.745-0.847) vs. 0.796 (0.741-0.844) and ES-IoU 0.725 vs. 0.721; in Table 3, ES-Dice is 0.499 (0.469-0.531) vs. 0.452 (0.415-0.492). All these reported 95% CIs overlap, and the manuscript supplies no paired difference interval or p-value for the dMoE-vs-MoE comparison. The strongest subgroup claims (T1 Dice 0.718, n=11; T4 Dice 0.778, n=21) are given without subgroup CIs, while Sec. 4.4.2 calls the T4 result 'significantly outperforming.' Because dMoE differs from MoE only by attribute-conditioned gating, overlapping intervals on this comparison leave the distinctive mechanism unverified; the SOTA claim rests on point estimates whose noise is not quantified at the comparison level.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Distribution-aware Mixture of Experts (dMoE), a sparse-gating layer whose router is conditioned on a demographic or clinical attribute flag, and inserts it into the encoders of segmentation backbones (TransUNet, 3D ResUNet). The authors reinterpret standard MoE as feedback control and dMoE as attribute-wise mode-switching control (Section 3.3), then evaluate on Harvard-FairSeg (race), HAM10000 (age), and an in-house prostate CT dataset (T-stage, Gleason Grade Group, PSA), reporting equity-scaled Dice/IoU (ES-Dice/ES-IoU) together with subgroup Dice/IoU and comparisons against ADV, DRO, FEBS, FairDiff, and a default MoE baseline. The headline claims are SOTA ES-Dice values of 0.743 on the FairSeg rim task, 0.801 on HAM10000, and 0.499 on the prostate CT dataset, with particularly large subgroup gains reported for T1 (n=11) and T4 (n=21) prostate cases.","tokens_in":18793,"tokens_out":17815,"duration_ms":163504,"significance":"The empirical program is substantive: evaluation spans 2D and 3D segmentation, demographic and clinical attributes, and out-of-center test cohorts (Yongin and Gangnam Severance), and the main comparisons are benchmarked, so the fairness results are not artifacts of the control-theoretic framing. The released code, the ablations over module placement and parameter sharing, and the extension to clinical attributes (T-stage, GG, PSA) beyond demographics are genuine strengths. If the empirical gains hold up under proper paired inference, dMoE is a practical, architecture-agnostic mechanism with modest overhead. However, three load-bearing issues currently separate the claims from the evidence: the theoretical derivation in Section 3.3 relies on conditions that the implemented architecture violates; the decisive dMoE-versus-MoE comparison is not statistically supported as reported; and the appendix ablation tables are internally inconsistent with the main tables. The control-theoretic narrative is an analogy rather than a circularity, but the paper presents it as a derivation.","major_comments":[{"comment":"The derivation in Eqs. (13)-(16) identifies dMoE with mode-switching optimal control only if the transformation f is linear in its second argument and if the gating is a dense kernel-weighted sum over all experts; neither condition holds for the implemented module. Section 4.2 specifies each expert as a two-layer MLP with ReLU and dropout, which is nonlinear in its parameters, and Eqs. (3)-(5) with Top-k=2 implement sparse top-2 routing, whereas Eq. (16) uses the dense kernel weights G_i = K(h_t, h^i_t) summed over all i. The equality in Eq. (16) therefore does not hold for the actual dMoE layer, so the sentence \"This formulation naturally leads us to mode-switching control\" overstates what has been shown; Eqs. (10)-(11) are a restatement of training as a terminal-cost problem, with no optimal-control result (e.g., necessary conditions) derived or used. I recommend reframing Section 3.3 explicitly as an analogy or heuristic motivation, or providing a version of Eq. (16) that holds under the actual top-k routing and nonlinear experts, and tempering the abstract's claim of a \"comprehensive analysis of its underlying mechanisms.\" This issue does not by itself invalidate the empirical comparisons, which are benchmarked independently of the framing.","section":"Section 3.3, Eqs. (13)-(16); Section 4.2"},{"comment":"The comparison that isolates the paper's core novelty, dMoE versus the default MoE ablation, is not statistically supported as reported. The bootstrapped 95% CIs overlap in every dataset at the All level: rim ES-Dice 0.743 (0.723-0.763) versus 0.733 (0.713-0.752); cup ES-Dice 0.832 (0.810-0.853) versus 0.830 (0.809-0.847); HAM ES-Dice 0.801 (0.745-0.847) versus 0.796 (0.741-0.844); prostate ES-Dice 0.499 (0.469-0.531) versus 0.452 (0.415-0.492). No paired difference interval or test is reported anywhere, even though a paired bootstrap over per-subject Dice differences would be the natural procedure given the resampling setup described in Section 4.3. The subgroup headline numbers (T1 Dice 0.718 with n=11, T4 Dice 0.778 with n=21, Age<20 Dice 0.880 with n=24) are given without subgroup CIs, and Section 4.4.2 calls the T4 result \"significantly outperforming\" without any test. The manuscript should either report paired inference for dMoE versus MoE (for example, a bootstrap CI on the per-case difference), add subgroup CIs and justify the use of \"significantly,\" or temper the SOTA claim to point estimates whose uncertainty is acknowledged.","section":"Section 4.3, Section 4.4.2, Tables 1-3"},{"comment":"The ablation tables are internally inconsistent with the main results and with one another. For HAM10000, the configuration labeled \"Ours\" in Table 11 (encoder-only, layer-wise, mode-switching) reports ES-Dice 0.841, whereas the same configuration in Table 2 reports ES-Dice 0.801, even though the All Dice values agree at 0.884. For the prostate dataset, Table 12 contains three \"Ours\" rows describing the same configuration with different values: ES-Dice 0.546, 0.546, and 0.499 with All Dice 0.711, 0.710, and 0.650 across the Location, Parameters, and Optimal-Control subsections, while Table 3 reports 0.499 and 0.650 for that configuration. In addition, Table 12 shows non-feedback control with ES-Dice 0.509, which is higher than the mode-switching value of 0.499 on the same task; this contradicts the claim in Appendix A.5 that mode-switching control \"consistently improved both ES-Dice and Dice scores\" (Dice improves, 0.650 versus 0.615, but ES-Dice does not). The authors should reconcile these numbers, state which run is canonical, and qualify the consistency claim, since the ablation currently does not support the stated conclusion.","section":"Appendix A.5, Tables 11-12 versus Tables 2-3"}],"minor_comments":[{"comment":"There are several wording and typesetting errors: \"flattend\" should be \"flattened\" in the paragraph before Eq. (2); the description after Eq. (5) says KeepTop-k \"retrains only the top-k\" but should say \"retains\"; Section 4.4.1 contains \"dMoe\" instead of \"dMoE\"; Section 4.4.3 contains \"These finding show\"; the Figure 3 caption says \"Violin plots depicts\" instead of \"depict.\"","section":"Section 3.2 and Eq. (5)"},{"comment":"The segmentation loss in Eq. (6) is typeset incorrectly: the expectation operator has a malformed subscript, the loss function symbol is garbled, and since Eq. (7) defines the prediction as a function of both the image and the attribute flag, Eq. (6) should make the dependence on the attribute explicit rather than writing the loss only over X.","section":"Eq. (6)"},{"comment":"The abstract contains two consecutive sentences announcing code availability (\"The source code will be made available. The source code is available at https://github.com/tvseg/dMoE.\"); one of them should be deleted.","section":"Abstract"},{"comment":"The checkpoint-selection protocol is unclear: \"The best performance is selected from checkpoints saved at 100-epoch intervals\" is ambiguous for HAM10000, which is trained for only 100 epochs, and no validation split is mentioned for either 2D task. If \"best performance\" was evaluated on the test benchmark, the reported numbers would be optimistically biased; the selection criterion and the data used for selection should be stated explicitly (the 3D task describes early stopping on a validation set, so the 2D protocol should likewise be specified).","section":"Section 4.2"},{"comment":"The HAM10000 subgroup counts do not sum to the stated total: 121+469+328+120+24 = 1062, while the tables state n=1061. In addition, the age column labels \"≥80,\" \"≥60,\" \"≥40,\" \"≥20,\" and \"<20\" are cumulative thresholds, but the reported counts (e.g., 469 for \"≥60\" versus 328 for \"≥40\") show the groups are actually disjoint bins; the labels should be ranges such as \"80+,\" \"60-79,\" \"40-59,\" \"20-39,\" and \"<20\" to avoid confusion.","section":"Table 6 and Table 2"},{"comment":"The FairDiff ES-metrics are \"recalculated using Eq. (19) based on metrics reported in the original paper.\" Please describe how the per-subgroup Dice/IoU values required by Eq. (20) were obtained for FairDiff, and state whether the identical recalculation procedure was applied to the other externally reported baselines, so that the comparison is apples-to-apples.","section":"Table 1 footnotes"},{"comment":"The reference Langley (2000) appears in the bibliography, but I could not locate an in-text citation for it; it should be cited where relevant or removed.","section":"References"},{"comment":"Equations (12) and (17) are written in differential form (\"d h_t = ... dt\") with no stochastic term, while the rest of the presentation is in terms of discrete layer updates and ODEs; the differential notation should be made consistent with the deterministic setting, or the stochastic convention should be introduced explicitly.","section":"Eqs. (12) and (17)"}],"recommendation":"major_revision","confidential_remarks":"The same-configuration discrepancy between the main tables and the ablation tables (Tables 2 versus 11 and Tables 3 versus 12) is the most serious data-integrity issue in this submission. Before any revision is accepted, I recommend requiring the authors to verify with the released code which numbers are reproducible; if the appendix values (e.g., HAM ES-Dice 0.841, prostate ES-Dice 0.546) are the correct ones, the abstract and main-text SOTA numbers would need to be revised. The checkpoint-selection protocol for the 2D tasks should also be checked against the code, since no validation set is described. The paper's empirical core is plausible and the multi-center 3D evaluation is a genuine strength, but the paired statistical analysis requested in the major comments is necessary before the dMoE-versus-MoE contribution can be regarded as established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a reasonable engineering contribution: attribute-conditioned gating for mixture-of-experts, applied to medical segmentation fairness. It runs on three datasets with different attributes, integrates with both transformer and CNN backbones, and includes ablations. The dMoE module is simple and plausible, and the authors report bootstrapped CIs for the headline metrics and share code. I'll give credit for breadth and for not overclaiming in the text, even if the abstract says \"state-of-the-art.\"\n\nThe soft spot is the statistical evidence for the core claim. The comparison that isolates the attribute-conditioned router is dMoE vs. plain MoE, and the reported 95% CIs overlap on essentially every ES metric across all three tables. The T1 and T4 subgroup numbers, where the gains look largest, have no CIs at all despite n=11 and n=21. So the paper does not actually show that the distribution-aware gating improves on a well-chosen baseline. That is not a minor issue; it is the central novel mechanism.\n\nThe control-theoretic framing is also weaker than advertised. Equation 16 requires f to be linear in its second argument to turn the kernel-weighted sum into a mixture of experts. But the experts are two-layer MLPs with ReLU, so that equality fails for the deployed architecture. The \"mode-switching optimal control\" interpretation is an analogy, not a derivation. The authors do say \"if f is linear,\" but they never connect that condition to their implementation. Reframing it as intuition or inspiration would be honest; presenting it as a derivation overstates the theory.\n\nNone of this sinks the method. The empirical direction is consistent, the ablations support mode-switching over plain feedback, and the code release is useful. But the paper needs paired difference tests or bootstrap-level comparison, subgroup CIs for the small groups, and either a fixed theory or a more modest claim about what the control analogy provides. As it stands, the headline result is a set of point estimates with overlapping intervals.\n\nWho should read this: researchers working on fairness in medical imaging, especially those who want a simple, plug-in router baseline. It is not a paradigm shift, and I would not cite it as evidence that attribute-conditioned gating works; I would cite it as a comparative approach. Still, a serious referee could extract a solid conditional paper from this material if the authors do the statistical and theoretical cleanup. I'd send it to review, not desk-reject, but the reviewer should push hard on what the distinctive contribution actually shows.","headline":"A reasonable plug-in fairness mechanism whose empirical edge over its own MoE baseline is not statistically established, and whose control-theoretic derivation rests on an assumption the implemented experts violate.","tokens_in":19297,"tokens_out":2046,"would_cite":false,"duration_ms":22859,"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":"Conditioning sparse expert routing on demographic or clinical attributes improves segmentation fairness in medical imaging.","keywords":["medical image segmentation","fairness learning","mixture of experts","sparse gating","optimal control theory","equity-scaled Dice","demographic attributes","clinical attributes"],"falsifier":"Evaluate both sides of Eq. (16) on real hidden states from a trained dMoE layer; any difference beyond floating-point noise shows that the derivation's key equality fails for the implemented experts. Alternatively, replace the ReLU expert MLPs with linear experts and retrain: if the equity-scaled gains disappear, the nonlinear experts rather than the mode-switching control design are carrying the improvement.","tokens_in":18292,"feed_emoji":"⚖️","tokens_out":10329,"duration_ms":96262,"temperature":0.7,"pith_summary":"The paper's central claim is that segmentation models trained on imbalanced clinical data can be made fairer by conditioning a sparse mixture-of-experts gating network on an attribute flag, such as race, age, or tumor stage. The proposed dMoE module keeps a set of shared expert networks and learns a separate top-k router for each attribute value, so that underrepresented subgroups effectively switch among specialized subnetworks while still sharing what they can. On the FairSeg, HAM10000, and an in-house prostate radiotherapy CT dataset, dMoE reports the best equity-scaled Dice and IoU among the compared fairness methods, with the clearest gains in the least-represented subgroups. The paper also interprets this design as mode-switching control within an optimal-control view of neural network training, arguing that external distributional information acts as a switch among control modes. If the empirical claims hold, the method offers a practical way to turn available demographic and clinical labels into measurable fairness improvements in medical image segmentation.","feed_headline":"Attribute-aware gating boosts fairness in medical image segmentation","feed_subtitle":"Keying sparse expert routers to race, age, or tumor stage improves equity-scaled Dice across three segmentation datasets.","key_machinery":"The load-bearing object is the dMoE layer itself, defined by Eqs. (2)-(5): a sparse gating network $\\mathbf{G}_{attr}$ that takes the flattened hidden state and an attribute flag, injects Gaussian noise, applies KeepTop-$k$, and forms a weighted mixture of shared expert MLPs whose output is added as a residual to the hidden state. The control-theoretic bridge is the identity in Eq. (16), which rewrites a kernel-weighted sum of expert parameters as a mixture of expert outputs when the layer transformation is linear in its second argument; this identity connects the kernel approximation of the control policy to the MoE gating weights, and the attribute-specific matrices in Eq. (18) implement mode-switching. The paper's derivation is conditional on that linearity, while the implemented experts are two-layer ReLU MLPs, so the identity holds conceptually rather than numerically for the trained architecture.","core_discovery":"dMoE is a plug-in layer for segmentation networks: after a layer block, the hidden embeddings are flattened, an attribute-specific router computes noisy top-k weights over eight shared expert MLPs, and the weighted expert outputs are added back as a residual. Each attribute value (race group, age band, T-stage) trains its own router, while the experts are shared across the whole population, which the authors argue lets minority subgroups benefit from the model capacity without losing the common representation. Across three datasets and two backbone families, the paper reports consistent improvements in ES-Dice and ES-IoU relative to plain MoE, FEBS, DRO, adversarial training, and generative augmentation, including a rise in ES-Dice from 0.733 with MoE to 0.743 with dMoE on FairSeg rim segmentation, from 0.796 to 0.801 on HAM10000, and from 0.452 to 0.499 on the prostate CT dataset. The authors read this result through control theory: the layer-wise hidden-state update is a discrete dynamical system, standard MoE is a feedback controller, and dMoE is a mode-switching controller whose mode is selected by the attribute flag.","pith_inferences":["Beyond the paper: because Eq. (16) requires linearity that the ReLU experts violate, the reported gains, not the control-theoretic derivation, are what establish dMoE's value; testing the same routing with linear experts would separate the two.","Beyond the paper: the model uses the attribute flag at inference time, so a natural stress test is to withhold or corrupt the attribute at deployment and measure whether fairness degrades, which the paper does not report.","Beyond the paper: the same gating pattern could transfer to other dense prediction tasks and to multi-attribute settings, for example age and stage jointly, where a hierarchical or composed router might handle combined imbalances."],"forward_implications":["If dMoE's results are correct, attribute-conditioned routing can be added to existing transformer and CNN segmentation backbones without changing the loss function, converting demographic or clinical labels already present in many clinical datasets into fairness gains.","The reported experiments imply that fairness in medical segmentation can be improved by architectural conditioning on known attributes, not only by loss reweighting, adversarial debiasing, or generative augmentation.","The success on T-stage, Gleason grade, and PSA-level attributes suggests the same mechanism can address clinical factors that are not directly visible in the image, such as disease severity and treatment intent.","The ablation studies attribute the improvement specifically to the mode-switching (attribute-aware) control component rather than to the added MoE capacity alone."],"supporting_citations":[{"why":"Defines the sparse gated mixture-of-experts layer that dMoE extends with attribute-conditioned routers.","marker":"Shazeer et al., 2017"},{"why":"Provides the FairSeg dataset, the FEBS fairness baseline, and the equity-scaled Dice/IoU metrics used for evaluation.","marker":"Tian et al., 2024"},{"why":"Supplies the FairDiff generative augmentation baseline compared on FairSeg.","marker":"Li et al., 2024"},{"why":"Supplies the distributionally robust optimization baseline for fairness comparison.","marker":"Sagawa et al., 2019"},{"why":"Supplies the adversarial fairness training baseline for comparison.","marker":"Madras et al., 2018"},{"why":"TransUNet is the transformer backbone into which dMoE is inserted for the 2D experiments.","marker":"Chen et al., 2021"},{"why":"The 3D U-Net is the CNN backbone used for the 3D radiotherapy target segmentation experiments.","marker":"Çiçek et al., 2016"},{"why":"Provides the feedback-control formulation that the paper uses to reinterpret MoE as feedback control and dMoE as mode-switching control.","marker":"Åström & Murray, 2021"}],"fun_headline_variants":["Attribute-aware routers boost fairness in medical segmentation","Control-theoretic MoE levels up fairness in segmentation","dMoE: Fairer medical segmentation via attribute-specific experts","Sparse experts keyed to demographics improve segmentation equity","Plug-in fairness: distribution-aware MoE for medical imaging"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument assumes that averaging expert parameters and then applying the layer gives the same result as applying the layer to each expert and averaging the outputs, but the implemented experts are nonlinear ReLU networks, so this equality does not hold for the actual model.","fun_headline_variants_meta":{"raw":{"variants":["Attribute-aware routers boost fairness in medical segmentation","Control-theoretic MoE levels up fairness in segmentation","dMoE: Fairer medical segmentation via attribute-specific experts","Sparse experts keyed to demographics improve segmentation equity","Plug-in fairness: distribution-aware MoE for medical imaging"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00075,"raw_usage":{"total_tokens":3351,"prompt_tokens":968,"completion_tokens":2383,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":2306}},"tokens_in":584,"tokens_out":2383,"duration_ms":18747,"temperature":1.0,"reasoning_tokens":2306,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T18:16:07.180252+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate both sides of Eq. (16) on real hidden states from a trained dMoE layer; any difference beyond floating-point noise shows that the derivation's key equality fails for the implemented experts. Alternatively, replace the ReLU expert MLPs with linear experts and retrain: if the equity-scaled gains disappear, the nonlinear experts rather than the mode-switching control design are carrying the improvement.","supporting_citations":[{"cited_title":"Fairseg: A large-scale medical image segmentation dataset for fairness learning using segment anything model with fair error-bound scaling","cited_arxiv_id":null,"evidence_quote":"Provides the FairSeg dataset, the FEBS fairness baseline, and the equity-scaled Dice/IoU metrics used for evaluation."}],"review_version":1}