{"id":"355cb3be-9879-41e0-afb8-fdc35d5be084","arxiv_id":"2508.02844","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A weakly supervised cardiac segmentation method that combines positive and complementary coarse annotations with per-pixel transition matrices achieves near-fully-supervised Dice on ACDC, MSCMRseg, and UK Biobank.","lead":"This paper trains a heart MRI segmentation model using only rough outlines and background strokes, not exact pixel-by-pixel labels. The model learns to clean up the rough labels and reports accuracy close to fully supervised methods, potentially reducing the cost of preparing medical training data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central claim that per-pixel transition matrices let the model 'infer the true segmentation distribution' is not established: Eq. 8 is compatible with many (prediction, transition-matrix) decompositions, and no identifiability analysis or ablation tests this premise.","rationale":"I read the paper in good faith. The experimental results are promising: RefineSeg reports higher average Dice than prior weakly supervised baselines and approaches fully supervised nnU-Net on ACDC and MSCMRseg, and the code is publicly available. Those are real positives. However, the paper's central scientific claim is about mechanism: that jointly training on positive and negative coarse annotations with per-pixel transition matrices and identity regularization lets the model 'infer the true segmentation distribution.' That claim is considerably stronger than 'this loss works well empirically,' and the manuscript provides no direct evidence for the mechanism. The objective in Eq. 8 is underdetermined: for each pixel, the transition matrix T has L×L free parameters, and the positive loss is computed only on positive coarse pixels, so many (p, T) pairs can fit the observed labels. Identity regularization biases the solution toward T = I, but the regularization coefficient is not reported and no analysis shows that the learned T are close to identity in meaningful regions or that deviations from identity correspond to real annotation noise. The negative-channel construction in Eq. 6 also introduces additional transition matrices that are not covered by the stated regularizer. The missing ablation is particularly telling: if fixing T to identity (i.e., training directly on coarse labels with partial losses) already achieves the reported Dice, then the transition-matrix mechanism is not responsible for the gains, and the paper should be read as an engineering contribution rather than a validated inference framework. I therefore do not see a reason to change the reader's CONDITIONAL verdict; the concern is exactly the weakest assumption the reader identified, and the suggested synthetic test would settle it decisively.","tokens_in":7246,"tokens_out":3593,"duration_ms":50488,"concrete_test":"Run a synthetic identifiability audit: generate images with known ground-truth segmentations p*, corrupt them with known per-pixel transition matrices T_true to produce positive and negative coarse labels, then train RefineSeg exactly as described. Compare recovered (p, T) against (p*, T_true): if the recovered T diverges from T_true while the loss remains low, or if setting T to identity (λ → ∞) yields nearly the same Dice as the learned T, the claim that transition matrices recover the true distribution fails. Also run the ablation with only positive annotations and only negative annotations to test whether the dual-channel refinement contributes anything beyond each channel alone.","verdict_should_be":"UNCHANGED","load_bearing_attack":"RefineSeg's headline contribution rests on the mechanism in Section 2: per-pixel, per-image transition matrices T^(r) map the predicted segmentation p to the observed coarse label distribution via Eq. 1, and joint training with positive and negative coarse labels plus an identity regularization (Eq. 7) is said to progressively refine the prediction toward the true segmentation distribution. The load-bearing premise is that the optimization in Eq. 8, together with the identity prior, forces T to represent annotator noise rather than absorbing the segmentation signal. This premise is untested and, as stated, mathematically fragile: because T^(t_i) is independent per pixel and per image, the positive loss (Eqs. 2-4) can be driven toward zero for essentially any predicted distribution p by choosing T to put probability mass from p onto the observed coarse label. The negative channel (Eqs. 5-6) and the Frobenius regularization toward identity do not remove this degeneracy; they only bias the solution, and the regularization weight λ is not specified in the text. Moreover, Eq. 7 regularizes only T^(t_i), while the negative-channel matrices in Eq. 6 appear unregularized. There is no identifiability analysis, no synthetic experiment with known transition matrices, and no ablation that fixes T to identity or removes one coarse-annotation channel. Without such evidence, the mechanistic claim that the model 'infers the true segmentation distribution' is not supported, even if the reported Dice scores are reproducible.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes RefineSeg, a weakly supervised segmentation framework that trains a 2D U-Net from coarse positive (target) and complementary (negative) annotations. The method introduces per-pixel, per-image transition matrices to model annotation noise, an identity regularization term on the positive transition matrices, and a joint loss combining positive and negative cross-entropy/Dice terms. The authors evaluate on ACDC, MSCMRseg, and UK Biobank, reporting average Dice values of 0.901 and 0.852 on the first two datasets, which they interpret as surpassing existing weakly supervised methods and approaching fully supervised nnU-Net (0.920 and 0.902). The central mechanistic claim is that simultaneous training on both coarse channels 'infers the true segmentation distribution' through the transition-matrix formulation.","tokens_in":7557,"tokens_out":6530,"duration_ms":80294,"significance":"If the empirical results withstand scrutiny, the contribution is relevant: it offers a fully weakly supervised pipeline that uses only coarse target and complementary annotations, which are cheaper to obtain than pixel-level masks, and it releases code. The reported Dice numbers on ACDC and MSCMRseg are competitive with stronger supervision. However, the paper's mechanistic claim about inferring the true segmentation distribution is not currently backed by identifiability analysis, synthetic validation, or ablations of the transition-matrix mechanism. The statistical evidence is also incomplete: no standard deviations, confidence intervals, or p-values are reported despite a strong statistical claim in Table 1. The significance paragraph is therefore conditional on the authors addressing these load-bearing gaps.","major_comments":[{"comment":"The table footnote claims that bold results are statistically better than other methods at p<0.01 by paired t-tests, but no p-values, standard deviations, or confidence intervals are reported anywhere. In addition, the comparisons with ScribFormer and CycleMix appear to use numbers reported in those original papers rather than re-running those methods under the same 80/20 split and U-Net backbone. Since the headline claim is 'surpasses the state-of-the-art weakly supervised methods,' please report per-class Dice as mean±std, provide the actual paired p-values for the key comparisons, and either retrain all baselines under the identical protocol or state clearly which numbers are transferred from the original publications.","section":"Table 1 / Section 3.2"},{"comment":"The load-bearing premise that the identity regularizer forces the per-pixel, per-image transition matrices T^(t_i) to represent annotator noise rather than to absorb the segmentation signal is not established. Because T is independent per pixel and per image and the loss in Eq. (8) is minimized jointly over p and T, many (p, T) decompositions can fit the observed coarse labels; the identity regularization only biases the solution. No identifiability analysis, no synthetic experiment with known transition matrices, and no ablation with T fixed to identity is provided. Please add such experiments, or explicitly weaken the claim from 'infers the true segmentation distribution' to 'learns a useful refinement under the imposed regularization.'","section":"Section 2.4, Eqs. (7)-(8)"},{"comment":"The negative-channel formulation is under-specified and appears unregularized. The notation switches between T and T^(c_i), and q_n is never defined. More importantly, L_reg in Eq. (7) regularizes only T^(t_i), not the matrices used in the negative channel in Eq. (6). If the negative-channel transition matrices are independent and unregularized, the degeneracy concern in the previous comment is amplified. Please define all variables, clarify the relationship between T and T^(c_i), and either regularize both channels or justify why the negative channel needs no regularization.","section":"Section 2.3, Eqs. (5)-(6) and Eq. (7)"},{"comment":"The UK Biobank evaluation, which is highlighted in the abstract and in Section 1 as a contribution, is presented only through qualitative figures (Figures 3 and 4). No Dice numbers are given for UKBB. Since the paper claims to provide the first weakly supervised benchmark on this dataset, a quantitative table with per-class Dice, averages, and standard deviations is necessary.","section":"Section 3.3"}],"minor_comments":[{"comment":"For ACDC and MSCMRseg, the coarse annotations are generated by eroding ground-truth masks; this is not equivalent to real non-expert coarse drawings. Please discuss how this may affect the external validity of the results, and clarify how the manually annotated UKBB coarse labels were verified.","section":"Section 3.1"},{"comment":"The displayed values of the weight parameters α_i, β_i, and λ are not legible in the manuscript (the symbols appear as blank placeholders). Please state these hyperparameters explicitly in the text.","section":"Section 2.2, Eq. (2)"},{"comment":"Please specify whether the transition matrices are column-stochastic or row-stochastic and whether any non-negativity or column-sum constraints are imposed, since this matters for the probabilistic interpretation of Eq. (1).","section":"Section 2.1, Eq. (1)"},{"comment":"There is a small numerical inconsistency: the text reports an average Dice of 0.902 for the proposed method on ACDC, while Table 1 lists 0.901. Please correct this.","section":"Section 3.2 / Table 1"},{"comment":"References [5] and [6] appear to be the same Bayeseg paper; please deduplicate and update the citation list.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is a plausible empirical contribution, and the reader's skepticism about identifiability is well-founded: the transition-matrix mechanism needs synthetic validation and ablations before the 'infers the true segmentation distribution' claim can be accepted. The missing statistical reporting is also a serious but fixable issue. I would not reject the paper, because the benchmark numbers are promising and the authors can address the concerns within the scope of a revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: RefineSeg is a genuinely new combination of dual coarse annotations and per-pixel transition matrices for weakly supervised segmentation, and the reported gains over ScribFormer are plausible. But the paper's central mechanistic claim—that the model infers the true segmentation distribution—is not supported by the evidence as presented.\n\nWhat's actually new: the framework simultaneously trains on positive and complementary coarse labels, with per-pixel, per-image transition matrices and an identity regularizer. That specific combination doesn't appear in the cited prior work, and the UKBB weakly supervised benchmark is a useful addition to the literature. The code is public, and the experiments cover two well-known public cardiac datasets plus UKBB. Those are real strengths.\n\nWhere it's soft: Table 1 claims statistical superiority via paired t-tests at p<0.01 but shows no standard deviations, confidence intervals, or p-values. That's a reporting gap that matters. More importantly, there is no ablation isolating the contribution of the negative-label channel, the transition matrix, or the regularization. The stress-test concern about identifiability is legitimate: Eq. 8 is compatible with many (prediction, transition-matrix) decompositions, and the identity regularization on only the positive-channel matrices is a weak bias, not a guarantee. Without a synthetic experiment with known transition matrices or an ablation that fixes T to identity, the statement that the model 'infers the true segmentation distribution' overreaches. The method might still work for reasons the authors haven't isolated, but the mechanism as described is not established.\n\nThe empirical direction is worth taking seriously. The authors are not hiding their method; they provide code and a clear enough description that someone could reimplement and test the ablations. That is exactly what peer review is for. This paper deserves a serious referee, but it needs major revisions before the claims can be trusted. I'd tell the authors to add ablations, error bars, and an identifiability analysis, and to soften the mechanistic language.\n\nBottom line: send it to review, but with the expectation that the central claim will be tested and likely scaled back. I'd cite it in related work on weakly supervised medical segmentation, and I'd bring it to a reading group interested in annotation-efficient learning.","headline":"A plausible weak-supervision method with a real empirical gain, but the paper's central mechanistic claim is undersupported and needs substantial strengthening.","tokens_in":8084,"tokens_out":1555,"would_cite":true,"duration_ms":19962,"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":"This paper claims that training a U-Net on only coarse target and complementary drawings, with per-pixel transition matrices modeling the label noise, yields cardiac MRI segmentations that surpass other weakly supervised methods and…","keywords":["medical image segmentation","weakly supervised learning","coarse label","transition matrix","complementary labels","cardiac MRI","U-Net","label noise"],"falsifier":"On synthetic images with known per-pixel transition matrices, train the framework using only coarse labels generated from those matrices, then compare the learned matrices with the ground-truth ones; the central claim is falsified if the matrices are not recovered even when the predicted segmentations are accurate.","tokens_in":7058,"feed_emoji":"🫀","tokens_out":7196,"duration_ms":75200,"temperature":0.7,"pith_summary":"This paper tries to establish that pixel-level medical image segmentation can be trained from coarse annotations alone, without any fine-grained ground-truth masks. Its framework uses per-pixel transition matrices to model the noise in both 'target' and 'complementary' coarse drawings, and a joint loss that progressively refines the output toward the true segmentation distribution. If the claim holds, expert-level annotation could be replaced by cheaper rough drawings made by non-experts, lowering the barrier to large-scale medical segmentation. On ACDC and MSCMRseg cardiac MRI data, the method reports average Dice scores of 0.901 and 0.852, close to the 0.920 and 0.902 achieved by a fully supervised nnU-Net.","feed_headline":"Coarse labels alone rival full supervision for cardiac MRI","feed_subtitle":"A transition-matrix model learns fine heart segmentations from rough target and complementary drawings, hitting 0.90 average Dice.","key_machinery":"The load-bearing object is the pixel-wise transition matrix $T^{(r)}_{uv}$, whose entry $T^{(r)}_{uv,mk}$ is the probability that the true class $k$ appears as coarse class $m$ at pixel $(u,v)$ under annotation strategy $r$; it converts the predicted fine label distribution into the coarse label distribution that is compared with the observed coarse labels. The network learns these matrices jointly with the segmentation, using a positive-channel loss that supervises only pixels inside the coarse positive regions and a negative-channel loss that aligns the coarse prediction via the transposed transition matrix. A Frobenius-norm regularization term $\\sum_i \\|T^{(t_i)} - I\\|_F^2$ keeps the positive transition matrices near the identity, which the paper argues preserves structural integrity while still allowing refinement.","core_discovery":"The central discovery is that jointly optimizing two annotation-specific transition matrices—one mapping the true label distribution to positive coarse labels and one to complementary (negative) coarse labels—lets a U-Net recover fine segmentation masks from coarse-only supervision. Each image gets its own pixel-wise transition matrix, and the matrices are regularized toward identity to stop them from absorbing all the learning signal. Trained with a hybrid cross-entropy and Dice partial loss on annotated pixels for the positive channel and a transposed-matrix alignment for the negative channel, the model reaches average Dice scores of 0.901 on ACDC and 0.852 on MSCMRseg, surpassing prior weakly supervised methods and approaching full supervision.","pith_inferences":["A natural extension the paper leaves unexplored is whether the identity regularization alone is sufficient; replacing it with a learned prior or per-class constraints could be tested on datasets where the annotation noise is spatially biased.","The coarse labels on ACDC and MSCMRseg are produced by eroding ground-truth masks, so the noise is spatially uniform; real non-expert coarse drawings may violate this structure, and the framework's robustness to such mismatch is a testable open question.","If the transition matrices are identifiable in practice, the method could be iterated: refine coarse labels with a first pass, then retrain on the refined labels, turning the coarse-to-fine learning into a self-training loop."],"forward_implications":["If coarse target and complementary drawings suffice, the cost of annotating cardiac MRI drops from expert pixel-level tracing to rough boundary drawing that non-experts can perform.","The method outperforms current weakly supervised baselines, including scribble-based ScribFormer and CycleMix, on both ACDC and MSCMRseg, and its gap to full supervision is small (1.9 points of average Dice on ACDC, 5.0 on MSCMRseg).","The work provides the first weakly supervised versus fully supervised benchmark on UK Biobank, giving downstream cardiac studies a reference point for what coarse-label training can achieve.","Because the whole pipeline trains end-to-end from coarse labels only, the same loss could be applied to roughly labeled institutional data, potentially enabling large medical segmentation models to be trained with minimal manual labeling effort."],"supporting_citations":[{"why":"Supplies the pixel-wise transition matrix modeling of annotation noise that the framework adapts.","marker":"[22]"},{"why":"Fully supervised nnU-Net baseline that defines the performance target the method approaches.","marker":"[8]"},{"why":"Strongest weakly supervised scribble-based baseline that the method must surpass.","marker":"[10]"},{"why":"ACDC dataset whose ground-truth masks are eroded to generate the coarse annotations.","marker":"[1]"},{"why":"MSCMRseg dataset used for the second main evaluation.","marker":"[24]"},{"why":"UK Biobank dataset on which the first weakly supervised benchmark is run.","marker":"[13]"},{"why":"Partial loss that restricts supervision to annotated pixels, basis of the positive-channel loss.","marker":"[18]"},{"why":"Transition-matrix approach for complementary labels that inspires the negative-channel learning.","marker":"[19]"}],"fun_headline_variants":["Rough labels train fine heart segmentation","Coarse-only supervision nears full precision","Transition matrices turn noise into fine masks","Weakly supervised heart MRI closes gap","Coarse annotations achieve near-perfect Dice"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework assumes that per-pixel transition matrices, regularized only toward the identity, can recover the true segmentation distribution from coarse target and complementary labels without any fine-grained supervision.","fun_headline_variants_meta":{"raw":{"variants":["Rough labels train fine heart segmentation","Coarse-only supervision nears full precision","Transition matrices turn noise into fine masks","Weakly supervised heart MRI closes gap","Coarse annotations achieve near-perfect Dice"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000104,"raw_usage":{"total_tokens":979,"prompt_tokens":840,"completion_tokens":139,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":456,"completion_tokens_details":{"reasoning_tokens":76}},"tokens_in":456,"tokens_out":139,"duration_ms":2601,"temperature":1.0,"reasoning_tokens":76,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T04:51:20.229940+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On synthetic images with known per-pixel transition matrices, train the framework using only coarse labels generated from those matrices, then compare the learned matrices with the ground-truth ones; the central claim is falsified if the matrices are not recovered even when the predicted segmentations are accurate.","supporting_citations":[{"cited_title":"PLoS medicine 12(3), e1001779 (2015)","cited_arxiv_id":null,"evidence_quote":"UK Biobank dataset on which the first weakly supervised benchmark is run."},{"cited_title":"In: International conference on machine learning","cited_arxiv_id":null,"evidence_quote":"Partial loss that restricts supervision to annotated pixels, basis of the positive-channel loss."}],"review_version":1}