{"id":"ab295a9c-932e-4672-9a66-778a7e57b083","arxiv_id":"2511.18894","paper_version":6,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"MetaDCSeg improves medical segmentation under noisy labels by combining meta-learned pixel weights with a dynamic boundary-distance attention module.","lead":"A new framework called MetaDCSeg uses pixel-by-pixel meta-learning plus boundary-distance weighting to make medical image segmentation more robust to noisy annotations. Tested on four medical datasets, it reports large accuracy gains over prior methods, but the validation relies on synthetic noise and the paper contains internal inconsistencies.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Core meta-weight definition is internally inconsistent: §3.2.3 requires α+β=1 per pixel, but the A.3 update rectifies and globally normalizes weights, and Eq. 9 sums pre-normalization values; the central module has no well-defined implementation.","rationale":"The reader's weakest_assumption focused on synthetic noise and the clean meta-validation set, which is a valid external-validity concern. But the more load-bearing issue is internal: the algorithm that produces the reported numbers is not uniquely specified. This is not a stylistic mismatch; it changes the actual loss and the DCD weighting. Under the complementary-constraint reading, γ_i^hw = α_i^hw + β_i^hw = 1, so the DCD center weights in Eq. (10) reduce to unweighted averages and the module cannot have the effect claimed in the ablation. Under the Appendix A.3 reading, the per-pixel constraint is absent and the uncertainty map γ is a by-product of a whole-image normalization rather than a per-pixel reliability measure. The proof in A.4.1 silently switches between these readings (it first uses normalized α̂,β̂, then re-introduces per-pixel complementary intuitions). Because the paper does not indicate which version was executed, the experimental tables cannot be independently checked. I agree with the reader's REJECT verdict: the results may be correct in spirit, but as written the central claim is not reproducible. The anonymous code link is a positive sign, but it has not been inspected against the equations; the proposed test would settle whether the code matches either variant and whether the headline gains survive a consistent implementation.","tokens_in":942,"tokens_out":1210,"duration_ms":77238,"concrete_test":"Run the released repository (anonymous.4open.science/r/Meseg-6F8A) as-is and instrument the weight branch to log (α,β) after the rectification/normalization step (Eqs. 23–24) and the γ used in Eq. 9. Check whether α+β=1 holds for every pixel and whether γ is constant. Then re-run the MSD Heart 40% noise experiment once with per-pixel normalization enforcing α+β=1 and once with the repository's global normalization, comparing mIoU/DSC/HD to Table 1. If neither variant reproduces Table 1, the reported gains are not tied to a well-defined algorithm.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism is the pixel-wise meta-weight pair (α,β). Section 3.2.3 states that each pair satisfies the complementary constraint α_i^hw + β_i^hw = 1 with values in [0,1], and Eq. (3) uses these weights in the bootstrapping loss. However, Appendix A.3's online approximation (Eqs. 21–24) updates α and β independently by gradient descent, rectifies them with max(·,0), and then normalizes by Z_i = Σ_{h,w}[α̃+β̃] over the entire image, not per pixel. After this update, α̂_i^hw + β̂_i^hw is generally not 1 and varies across pixels. The convergence proof in A.4.1 uses these globally normalized weights, while the DCD module (Eq. 9) defines γ_i^hw = α̃_i^hw + β̃_i^hw using the pre-normalization, rectified weights. Thus there are at least three mutually incompatible definitions of the quantities that drive both the loss and the boundary weighting. If one instead enforced α+β=1 per pixel as the text claims, then γ would be identically 1 and the center weighting in Eq. (10) would be constant, eliminating the DCD contribution. The paper does not say which variant produced Tables 1–4; without this, the headline claim of consistent gains is not reproducible from the description. This is load-bearing because the reweighting scheme is the paper's core contribution and is used in every experiment.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MetaDCSeg, a medical image segmentation framework designed to be robust to noisy annotations and boundary ambiguity. The method combines a pixel-wise meta-learning scheme that learns per-pixel weights for real and pseudo labels with a Dynamic Center Distance (DCD) module that reweights boundary pixels using distances to foreground, background, and boundary feature centers. The authors report experiments on MSD Heart, MSD Brain, PROMISE12, and Kvasir-SEG under synthetic noise levels of 0%, 20%, 40%, and 60%, and claim consistent improvements over several state-of-the-art baselines, with ablations and complexity analysis. The paper also includes a theoretical convergence analysis in the appendix.","tokens_in":18295,"tokens_out":7406,"duration_ms":79755,"significance":"If the method were precisely specified and the experiments reproducible, the idea of learning spatially varying trust in noisy labels is a reasonable and potentially useful contribution to medical segmentation. The manuscript has several strengths: it reports four datasets, ablates the proposed modules, examines meta-dataset size and computational cost, and provides an anonymous code-link. However, the central reweighting mechanism is described inconsistently across the main text and appendix, the convergence proof is incomplete, and one baseline (MADGNet on PROMISE12) produces implausible results on clean data. These issues are load-bearing for the paper's central claim of consistent, significant state-of-the-art gains, so the paper is not acceptable in its present form.","major_comments":[{"comment":"The central reweighting mechanism is internally inconsistent. §3.2.3 requires α_hw+β_hw=1 with α,β∈[0,1] per pixel, and Eq. (3) is built on that complementary pair. Appendix A.3 updates the two weights independently by gradient descent, rectifies them with max(·,0), and normalizes by Z_i=Σ_{h,w}[α̃+β̃] over the entire image (Eq. 24). After this update, α̂+β̂ is not 1 per pixel and varies spatially. Equation (9) defines γ from the pre-normalization rectified values, while the convergence proof in A.4.1 uses the normalized values. These are three mutually incompatible definitions. Conversely, if α+β=1 were enforced, γ would be identically 1, making the center weighting in Eq. (10) constant and eliminating the DCD contribution. The paper does not state which variant produced Tables 1–4. Because this reweighting is the paper's core contribution, the headline empirical claims are not reproduc","section":"§3.2.3, Eq. (3); §3.3.3, Eq. (9); Appendix A.3, Eqs. (21)–(24)"},{"comment":"The convergence proof is incomplete and partly circular. Equation (32) replaces a cross term by a sum of squares using 'the fact that meta-weights are updated to align with validation gradients'; no such fact is proven. The one-step update in Eqs. (28)–(30) only shows that α moves in a direction of positive gradient alignment; it does not imply the equality used in Eq. (32), and the β update is not analyzed. Moreover, the claimed condition λ<√(2/(ησ²ML)) does not follow from Eq. (35): λ cancels when the inequality is divided by λ², and the resulting inequality involves sums of inner products with validation gradients that are not bounded by σ and |V| alone. The theorem therefore does not establish monotonic decrease of the validation loss. It should either be proved rigorously or removed/qualified.","section":"Appendix A.4.1, Theorem 1"},{"comment":"The MADGNet row reports mIoU=8.92 and DSC=16.38 at 0% noise, and mIoU=5.46 at 40% noise. MADGNet is a published state-of-the-art medical segmentation method; near-zero scores on clean expert labels for prostate MRI are implausible and indicate a broken baseline implementation. Its 'collapse' values are therefore not a valid comparison, and the text statements about MADGNet on PROMISE12 are unsupported. The baseline must be rerun and the table corrected. Since the paper claims consistent gains across datasets, a broken baseline in one dataset also casts doubt on the reliability of the other baseline implementations.","section":"Table 4, PROMISE12"},{"comment":"All robustness experiments are conducted only under synthetic label corruption (rotation, morphological erosion/dilation, ellipse replacement). The paper defines the target noise as real annotation boundary inconsistency, but no evidence is provided that this protocol captures how human annotators actually err. Without validation on real noisy labels or at least a comparison against annotator disagreement, the claimed transfer of the reported gains to real-world medical annotations is not established. This is not an internal inconsistency, but it is load-bearing for the paper's applied claims.","section":"§4.1 and Appendix A.2"}],"minor_comments":[{"comment":"The implementation details say 100 epochs total with SGD, while Table 8 lists per-dataset epochs of 120–180 and different learning rates. Please harmonize these statements.","section":"§4.1 vs. Appendix A.1"},{"comment":"The text mentions a 'GMM threshold β=3.0', but no GMM procedure is defined in the method. Also, α is used both for momentum (α=0.9) and for the meta-learned real-label weight, which is confusing; please use distinct symbols.","section":"§4.1"},{"comment":"The caption begins with 'Figure 2. Figure 2. Overview...' — remove the duplication.","section":"Figure 2"},{"comment":"Theorem 2 bounds DCD without clipping, but A.5.1 introduces DCD_max clipping; the stability bound should be updated to reflect the clipping mechanism, otherwise the two analyses are not consistent.","section":"A.5.1 vs. A.4.2"},{"comment":"The cost efficiency definition is ad hoc and the denominator can be non-positive or ill-defined for small overheads. Please state the assumptions or replace it with standard FLOPs/memory/time reporting.","section":"Eq. (14)"}],"recommendation":"reject","confidential_remarks":"The anonymous repository may, in principle, resolve which variant of the algorithm produced the tables. If the authors can show that a single, well-defined implementation reproduces all reported numbers and that the PROMISE12 MADGNet baseline is corrected, a resubmission could be considered. As written, the contradictory specification of the core module and the invalid baseline are too central to overlook."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Let me give you the short version: this is a plausible combination of L2B-style meta-reweighting with a new boundary-distance module, but as written the core meta-weight definition is inconsistent between the main text and the appendix, and the convergence proof assumes what it needs to show. I would not trust the reported gains until those are fixed.\n\nWhat's actually new is the Dynamic Center Distance module. It computes feature distances from each boundary pixel to foreground, background, and boundary centers, then uses those to reweight boundary loss terms. That's a sensible heuristic and it is cleanly separable from the meta-learning part. The paper also does real work: four datasets, three noise levels, ablations, a meta-dataset size study, and a code link. The ablation ordering — meta-learning matters most, DCD adds a modest consistent amount, dice loss contributes least — is believable.\n\nThe soft spots, in proportion.\n\nFirst and most important: the central weight definition is not pinned down. Section 3.2.3 says alpha + beta = 1 per pixel. Appendix A.3 updates both independently, rectifies them, and then normalizes by a single scalar Z_i computed over the entire image, so the sum is generally not 1 per pixel. Equation (9) defines gamma from the pre-normalization tilde weights. At least three different objects are in play, and the paper never says which one produced Tables 1-4. Because this reweighting is the core contribution, the method is not reproducible from the description. The stress-test note is right about this, though it overstates one point: even if gamma were identically 1, the DCD distances would still vary and the module would not be a no-op. The contradiction is still load-bearing.\n\nSecond, the convergence proof in A.4.1 is circular. Equation (32) uses the fact that the meta-weights are updated to align with validation gradients, which is exactly what the proof is supposed to establish. The rest is a bound with a condition on lambda, not a monotonic decrease argument.\n\nThird, the MADGNet baseline on PROMISE12 clean gives an mIoU of 8.92. A near-zero score on clean data indicates a broken reimplementation, and it undermines confidence in the other baseline numbers. The text also says the 13.96% gain at 60% noise on PROMISE12 is over HAMIL; the table shows it is actually over L2B. Both are fixable but need to be.\n\nFourth, the evaluation uses only synthetic noise. The protocol is explicit and the noise levels are calibrated, but there is no validation on real noisy annotations. The clean meta-validation set is also a practical limitation.\n\nWho gets value: people working on noisy-label medical segmentation who want to see a concrete combined meta-learning + boundary-weighting approach. But the paper is not ready as a refereed contribution. The core definition and proof need to be fixed, and the baseline table redone. I would send it to peer review rather than desk-reject — the idea is sane and the experimental scope is serious — but I would not accept it in this form.","headline":"Plausible L2B-style meta-reweighting plus a boundary-distance module, but the paper's core weight definition contradicts itself and the proof assumes its conclusion; the experiments are broad but not clean enough to support the headline claims.","tokens_in":18788,"tokens_out":5053,"would_cite":false,"duration_ms":48087,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that medical segmentation noise is spatially concentrated at anatomical boundaries, and that per-pixel meta-learned trust weights combined with a boundary-aware feature-distance loss let segmentation networks stay accurate","keywords":["noisy label learning","medical image segmentation","pixel-wise meta-learning","boundary uncertainty","dynamic center distance","label noise robustness","bi-level optimization","segmentation with noisy annotations"],"falsifier":"Run MetaDCSeg on natural noisy labels, e.g., masks derived from multiple expert annotators with known inter-observer disagreement, without synthetic corruption; if mIoU gains over baselines largely disappear compared to the synthetic-noise experiments, the noise model is the load-bearing element. Conversely, remove the clean meta-set (e.g., replace it with noisy labels) and check whether performance collapses toward baseline; a large drop would show the clean-set requirement is not just an implementation convenience.","tokens_in":17770,"feed_emoji":"🩺","tokens_out":5971,"duration_ms":57567,"temperature":0.7,"pith_summary":"MetaDCSeg claims that medical segmentation noise is not uniform: most annotation errors live at anatomical boundaries, so global noise-robust tricks fail. The paper introduces a two-part framework: a bi-level meta-learning loop that learns a trust weight for every pixel (how much to believe the noisy ground truth vs. the model's own prediction), and a Dynamic Center Distance module that measures each boundary pixel's feature-space distance to foreground, background, and boundary centers and spends extra loss on the ambiguous ones. Across four medical benchmarks with 20–60% synthetic boundary noise, the method consistently outperforms existing baselines, with the largest gains at high noise (e.g., 13.96 mIoU points on PROMISE12). The framework also improves clean-data results, and a 2% clean meta-validation set is enough to drive most of the benefit.","feed_headline":"Meta-learning per pixel tames boundary noise in medical segmentation","feed_subtitle":"A small clean set teaches the model which pixels to trust, lifting accuracy by ~14 points under 60% noise.","key_machinery":"Two coupled mechanisms carry the argument. (1) Pixel-wise bootstrapping loss with meta-learned complementary weights: at each pixel, the training loss is α·CE(p, y_real) + β·CE(p, y_pseudo) with α+β=1, where α (trust in the observed label) and β (trust in the pseudo-label) are optimized in an outer loop against a clean validation set, while the model parameters are optimized in the inner loop. (2) The Dynamic Center Distance (DCD): using γ=α+β as a pixel reliability weight, the method builds feature centers c_fg, c_bg, c_bd for high-confidence, low-confidence, and boundary pixels; each boundary pixel gets DCD = ‖h−c_fg‖·‖h−c_bg‖/(‖h−c_bd‖+ε), and softmax-normalized DCD values weight an extra","core_discovery":"The central claim is that the bottleneck in learning from noisy medical labels is spatial heterogeneity: pixels near anatomical boundaries are disproportionately corrupted, and treating all pixels equally—via global confidence thresholds or instance-level reweighting—overfits either noise or background. MetaDCSeg instead meta-learns, for each pixel, a complementary pair of weights α and β that balance the noisy ground-truth label against the model's own pseudo-label, using a small expert-clean validation set to guide the update. Then, using the resulting reliability map γ, it computes feature centers for foreground, background, and boundary regions and defines a composite distance DCD = ‖h−c","pith_inferences":["An unstated corollary is that the DCD boundary term is not logically tied to meta-learning: it could be attached to any segmentation loss whose confidence map supplies reliability weights, suggesting a cheaper adapter for existing robust-training pipelines.","A testable extension is to inject spatially uniform label noise (random flips of interior pixels) alongside the boundary-targeted protocol; the paper's mechanism predicts a much smaller gain in that regime, since DCD's value depends on noise concentrating at transitions.","The boundary center c_bd can be read as a feature-space 'transition centroid', implying DCD measures how strongly a pixel's representation mixes the two classes—an interpretation that could yield uncertainty maps for active learning or clinician QA without new training.","Since the online approximation performs a single-step gradient unroll, the method's practical behavior may depend on meta-learning rate and batch composition; replaying the same protocol with different unroll steps (k-step meta-gradients) is a natural robustness probe the paper does not run."],"forward_implications":["At 60% synthetic boundary noise, MetaDCSeg retains 86.56% mIoU on MSD Heart and 75.68% on PROMISE12, outperforming prior methods by up to 13.96 mIoU points; the gains grow with noise level.","The method does not sacrifice clean data: it exceeds baseline performance on uncorrupted labels (e.g., 93.89% mIoU on MSD Heart), so noise-robust components do not impose a clean-set penalty.","Ablations place the meta-learning module as the main driver (removing it costs 4.22 mIoU at 40% noise), with DCD contributing boundary-specific gains and Dice loss enforcing global structural consistency.","A clean meta-validation set as small as 2% of the data captures most of the benefit; scaling to 10% adds only 0.28 mIoU while roughly doubling time and memory cost.","Inference time is essentially unchanged by the added modules (≈45.2 ms), so the robustness gains do not slow deployment."],"fun_headline_variants":["Pixel-wise meta-learning silences noisy labels in segmentation","Per-pixel weights beat boundary noise in medical imaging","MetaDCSeg: learning pixel trust to fix noisy medical labels","Harnessing pixel heterogeneity to clean medical segmentation","Meta-learned pixel weights cut noise harm in segmentation"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The performance claims rest on the assumption that the synthetic corruption pipeline—rotations, morphological erosion/dilation, and ellipse replacement—produces label noise resembling how real annotators actually make boundary errors, and that a small expert-clean meta-validation set is available during training.","fun_headline_variants_meta":{"raw":{"variants":["Pixel-wise meta-learning silences noisy labels in segmentation","Per-pixel weights beat boundary noise in medical imaging","MetaDCSeg: learning pixel trust to fix noisy medical labels","Harnessing pixel heterogeneity to clean medical segmentation","Meta-learned pixel weights cut noise harm in segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00016,"raw_usage":{"total_tokens":1052,"prompt_tokens":709,"completion_tokens":343,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":453,"completion_tokens_details":{"reasoning_tokens":265}},"tokens_in":453,"tokens_out":343,"duration_ms":4941,"temperature":1.0,"reasoning_tokens":265,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T20:35:23.897891+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run MetaDCSeg on natural noisy labels, e.g., masks derived from multiple expert annotators with known inter-observer disagreement, without synthetic corruption; if mIoU gains over baselines largely disappear compared to the synthetic-noise experiments, the noise model is the load-bearing element. Conversely, remove the clean meta-set (e.g., replace it with noisy labels) and check whether performance collapses toward baseline; a large drop would show the clean-set requirement is not just an implementation convenience.","supporting_citations":[],"review_version":1}