{"id":"0b5ceb7c-6841-4cb4-9341-c70eb001e9ac","arxiv_id":"2501.06678","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"CLCS combines two-branch collaborative learning, class-adaptive curriculum thresholds, and a noise balance loss to train medical segmentation models from pixel-dependent noisy labels.","lead":"This paper presents CLCS, a training framework that separates clean from noisy pixel labels in medical images using two cooperating networks and a curriculum threshold that adapts per class. It reports more than 3% Dice and mIoU gains over prior noise-robust segmentation methods on two medical datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The class learning-status proxy in the curriculum threshold (Eqs. 2–4) is unvalidated; if σ_c(t) is miscalibrated under high per-class noise, the clean/noise partition that drives the reported gains becomes unreliable.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing point: the curriculum dynamic threshold relies on a self-referential proxy that could be miscalibrated early in training or under high per-class noise. This is the single most important vulnerability because the central claim (SOTA gains) depends on the clean/noise partition being correct: the clean set drives the main CE loss, and the noise set is used only with a secondary robust loss. If the partition is wrong, the reported improvements would be fragile and unlikely to transfer to other noise distributions. The paper provides indirect evidence (Fig. 10's clean-ratio plot and the fixed-vs-dynamic ablation in Sec. IV-F), but neither tests the proxy directly against ground truth. The concrete test I propose would settle this by comparing against an oracle threshold and measuring the proxy's correlation with true accuracy. I agree with the reader's CONDITIONAL verdict: the paper is methodologically sound in its ablations and statistics, but the unvalidated proxy and the secondary issues (upper-bound anomaly, hyperparameter reporting) justify requesting revisions rather than full acceptance. My concern does not move the verdict because the reader already asked for corrections/clarifications, and the proxy concern is best addressed by an additional experiment rather than outright rejection.","tokens_in":22220,"tokens_out":9619,"duration_ms":99765,"concrete_test":"During training on Endovis18 SFDA-Noise, replace the self-estimated thresholds Tc(t) with oracle thresholds computed from true per-class accuracy on a clean validation split, keeping all other components (two-branch, discrepancy loss, CCV, NBL) unchanged. If oracle thresholds improve average Dice by more than 1 point over the reported 67.51, the σ_c proxy is measurably miscalibrated and the curriculum mechanism is not the source of the gains. As a secondary check, compute the Spearman rank correlation between σ_c(t)/max_c σ_c(t) and true per-class accuracy at each epoch; a correlation below roughly 0.7 in the stable training phase would confirm the proxy is not a reliable learning-status estimate.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central mechanism for clean/noise separation is the Curriculum Dynamic Threshold (CDT), which sets per-class thresholds Tc(t) = [σ_c(t)/(2−σ_c(t))]·τ using σ_c(t) (Eq. 2), a count of the model's own confident predictions per class. This proxy is computed on noisy training labels and is used, through CCV (Eq. 6), to decide whether each pixel is trained with clean CE or with the Noise Balance Loss. If σ_c(t) does not track true per-class learning status, the entire partition is suspect. Under the reported noise levels (e.g., 44.3% noise for Clasper on Endovis18), confidently wrong predictions can inflate σ_c and raise Tc, excluding genuinely clean minor-class pixels; conversely, low confidence in under-learned minor classes lowers Tc and admits noisy pixels. The paper's Table IX and Fig. 10 show that the selected set has a higher clean ratio than the original labels, but this is an indirect check. It does not compare against an oracle threshold that uses true per-class accuracy, nor does it measure the correlation between σ_c(t) and true learning status over training. The fixed-threshold ablation (Sec. IV-F) uses a naive frequency-based baseline, so it cannot distinguish a well-calibrated curriculum from any dynamic adjustment. Because the claim is 'new state-of-the-art in different settings,' the lack of validation of this proxy makes the mechanism's contribution—and its generalizability—a load-bearing uncertainty.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CLCS, a two-branch collaborative learning framework for medical image segmentation with pixel-dependent noisy labels and class imbalance. The method combines (i) a boosted collaborative learning module with a cosine discrepancy loss, (ii) a curriculum dynamic threshold (CDT) that adapts per-class confidence thresholds based on the model's own learning-status proxy, (iii) a collaborative confidence voting (CCV) step that partitions pixels into clean and noisy sets, and (iv) a noise balance loss (NBL) that trains on the noisy set with a confidence-weighted mix of CE and RCE. Experiments on Endovis18 and RIGA under SFDA noise, SFDA+ED noise, and real annotation noise report state-of-the-art Dice/mIoU, with more than 3% average improvement over the best baselines in several settings.","tokens_in":22536,"tokens_out":7517,"duration_ms":68139,"significance":"If the empirical claims hold, the paper addresses a real and practical gap: existing noisy-label segmentation methods often assume class-dependent noise or use fixed thresholds that further imbalance minor classes. The paper's strengths are the fairness-oriented experimental setup (multiple baselines re-run, same backbone, ablations over components, p-values reported) and the release of code. The central risk is whether the curriculum selection mechanism, specifically the learning-status proxy in Eqs. (2)-(4), reliably separates clean from noisy pixels; this mechanism is the engine of the reported gains, and its validation is currently indirect.","major_comments":[{"comment":"The learning-status proxy σ_c(t) counts the model's own confident predictions on noisy labels and is used to set the per-class curriculum thresholds T_c(t). No evidence is provided that σ_c(t) tracks true per-class learning status: Fig. 10 shows that the selected clean set has a higher clean ratio than the original labels, but this is an indirect check, and the fixed-threshold ablation in Sec. IV-F is a weak baseline that cannot distinguish a well-calibrated dynamic schedule from any adaptive adjustment. Because the clean/noise partition drives both L_clean and NBL, this is a load-bearing uncertainty. Please provide a comparison against an oracle threshold using true per-class accuracy, or a direct correlation analysis between σ_c(t) and true per-class accuracy over training.","section":"III-B2, Eqs. (2)-(4) and Fig. 10"},{"comment":"On RIGA, CLCS under SFDA-Noise (89.38 Dice) and under Real-Noise (89.07 Dice) surpasses the Clean Upper Bound (88.52 Dice). The paper does not discuss this surprising result. If the clean upper bound is trained on clean labels with the same split and protocol, the fact that noisy-label training beats it requires an explanation (e.g., a regularization effect of the noise or a difference in how the clean labels are obtained). Without this clarification, the magnitude and interpretation of the reported improvements are unclear.","section":"Table III"},{"comment":"The implementation details state that the loss weights α and β are set to 1 and 0.01, respectively, but Table VII and Sec. IV-I report the optimal α as 0.01, and Table VIII and Sec. IV-H report the optimal β as 1. This is a direct internal inconsistency that makes the exact configuration used for the main results unreproducible. Please state the final hyperparameter values for each reported experiment and correct the conflicting sentences.","section":"IV-A vs. IV-H/I, Tables VII-VIII"},{"comment":"The algorithm uses a warmup stage (\"if iteration < warmup iteration\") but no warmup duration is reported in Sec. IV-A. The warmup length is a free parameter that affects how much early noisy-label fitting occurs and therefore influences the final results. Please specify the warmup length used in all experiments.","section":"Algorithm 1 and Sec. IV-A"}],"minor_comments":[{"comment":"The fixed-threshold baseline is described as setting the threshold for each class to the proportion of that class's pixels, but it is not explained how a class-pixel-proportion becomes a confidence threshold for sample selection; please clarify the conversion.","section":"IV-F"},{"comment":"In the SFDA-Noise block, the SCE row reports HD95 as \"10.96(±10.96)\", which is likely a typo (the error bar seems implausible); please verify and correct this value.","section":"Table IV"},{"comment":"The checkmark columns are ambiguous; it is not clear which rows use CCV with the convex mapping and which use CCV without it. Please label the rows or columns explicitly so that the comparison in Sec. IV-J can be verified.","section":"Table IX"},{"comment":"P-values are reported only for a subset of comparisons and the statistical test is not named. Report the p-values for all main comparisons and specify the test (e.g., paired t-test) and number of runs.","section":"IV-E"},{"comment":"The symbol ℓ_rce is introduced in Eq. (7) but is never defined in the text; please define the reverse cross-entropy loss explicitly or cite the exact equation from [25].","section":"III-C, Eq. (7)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of IEEE TMI and addresses a timely problem. The main technical uncertainty (the unvalidated learning-status proxy) is fixable with additional analysis, and the other major issues are textual or require re-reporting hyperparameters. The clean-upper-bound anomaly on RIGA should be resolved before publication, as it currently casts doubt on the interpretation of the results. I recommend major revision rather than rejection, as the core empirical claim is defensible and the required corrections are within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid, practice-oriented method paper for pixel-dependent noisy labels in medical segmentation. The main new idea is the class-adaptive curriculum threshold, which adjusts per-class selection thresholds based on the model's own confident-prediction counts. That, combined with two-branch collaborative voting and a noise balance loss (CE+RCE on the rejected set), is a reasonable and reasonably novel package. The experiments are mostly well done: baselines re-run on the same backbone, ablations over each component, p-values reported, code released, and two synthetic noise types plus real rater noise covered.\n\nThe soft spots are real but fixable. First, on RIGA the Clean Upper Bound (88.52 Dice) is lower than CLCS's SFDA-Noise score (89.38). A method trained on noisy labels should not beat a properly tuned clean-label model. This either means the upper bound is not a real upper bound or the clean-label training recipe is suboptimal. Either way, it needs an explanation because it undercuts confidence in the numbers. Second, the loss-weight reporting is contradictory: the implementation section gives α=1 and β=0.01, while the ablations in Tables VII and VIII show the best settings as α=0.01 and β=1. That is a direct inconsistency, and the stability claim over a 100× range of α is hard to square with Table VII, where α=1 drops Dice by about three points. Third, the base threshold τ is selected by scanning values across both datasets and noise types in Table V, which looks like test-set tuning. The SOTA claim and p-values should be read with that caveat.\n\nOn the stress-test point about the curriculum-threshold proxy σ_c(t): the concern is fair but not fatal. The paper shows the selected clean set has higher per-class clean ratios than the original labels, and the dynamic threshold beats a frequency-based fixed threshold. That is indirect evidence. An oracle-threshold comparison or a correlation plot of σ_c(t) against true per-class accuracy would strengthen the mechanism's validation. But the central empirical claim, that the full method outperforms the baselines, is supported by the re-run comparisons and ablations.\n\nBottom line: this deserves a serious referee. It could reduce annotation cost in medical segmentation if the gains replicate on other datasets. Send it to review, but the revision should address the upper-bound anomaly, fix the α/β contradiction, and add a more direct validation of the dynamic threshold.","headline":"Solid empirical method paper for pixel-dependent noisy labels; the class-adaptive dynamic threshold is the most novel piece, but an anomalous clean upper bound and an alpha/beta reporting contradiction need fixing.","tokens_in":23043,"tokens_out":4132,"would_cite":true,"duration_ms":30741,"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":"The paper claims that a collaborative two-branch framework with curriculum dynamic thresholds and a noise balance loss yields state-of-the-art medical image segmentation under pixel-dependent noisy labels, with more than 3% average Dice…","keywords":["medical image segmentation","pixel-dependent label noise","collaborative learning","curriculum dynamic threshold","noise balance loss","class imbalance","learning with noisy labels"],"falsifier":"Train a model with the proposed selection but reveal the oracle clean masks during evaluation, and compute the clean-label ratio inside the selected clean set per class at several epochs. If the selected set's clean ratio is not consistently higher than the original dataset's, or if the gains over a fixed-threshold variant vanish when per-class noise rates are roughly equal, the curriculum assumption is not what drives the reported improvements.","tokens_in":21994,"feed_emoji":"🩺","tokens_out":4231,"duration_ms":38348,"temperature":0.7,"pith_summary":"The paper argues that noisy labels in medical image segmentation should be treated as pixel-dependent rather than class-dependent, because annotation errors concentrate on boundaries and small objects, and that existing fixed-threshold denoising worsens class imbalance by dropping minority-class pixels. It proposes CLCS, a two-branch collaborative framework in which the two branches are pushed to disagree via a discrepancy loss, and pixels are voted clean or noisy using both branch predictions and the original label. A curriculum dynamic threshold adapts per class to learning progress, admitting more small-object pixels over time, while a noise balance loss keeps suspected noisy pixels in training instead of discarding them. The authors report state-of-the-art Dice and mIoU on Endovis18 and RIGA under SFDA noise, SFDA plus morphological noise, and real rater noise, with more than 3% average improvement over prior methods.","feed_headline":"Curriculum selection beats noisy-label segmentation by 3% Dice","feed_subtitle":"Adaptive per-class thresholds rescue small objects that fixed-threshold denoising drops, on surgery and retina benchmarks.","key_machinery":"The load-bearing object is the Curriculum Dynamic Threshold (CDT), defined class-wise as $T_c(t)= [\\hat{\\sigma}_c(t)/(2-\\hat{\\sigma}_c(t))]\\tau$, where $\\hat{\\sigma}_c(t)=\\sigma_c(t)/\\max_c \\sigma_c(t)$ and $\\sigma_c(t)$ counts pixels whose predicted confidence exceeds a base threshold $\\tau$ and whose predicted class is $c$. This turns the model's own confidence statistics into an adaptive, curriculum-like selection criterion: early in training low thresholds admit more minor-class pixels; as classes are learned, thresholds rise and selection becomes stricter. Around it, the Collaborative Confidence Voting (CCV) requires both branches' high-confidence predictions to equal the original label for a pixel to enter the clean set, and the Noise Balance Loss (NBL) applies confidence-weighted cross-entropy and reverse cross-entropy to the remaining pixels.","core_discovery":"The central discovery is that combining two-branch collaborative voting with a curriculum dynamic threshold and a noise balance loss can separate clean from noisy pixels well enough to improve segmentation under realistic pixel-level noise, even when noise rates differ sharply by class (17.0%, 32.1%, and 44.3% for Shaft, Wrist, and Clasper on Endovis18 SFDA noise). The two branches, kept divergent by minimizing cosine similarity between their feature maps, provide complementary evidence; a pixel is labeled clean only if both branches agree with the original label at high confidence. The per-class threshold is derived from the count of high-confidence predictions for that class, normalized by the maximum class count and shaped by a convex function, so small classes start with lower thresholds and are gradually included. Suspected noisy pixels are not dropped but trained with a confidence-weighted combination of cross-entropy and reverse cross-entropy. On Endovis18 SFDA noise average Dice rises from 64.50% for the best baseline to 67.51%, on RIGA SFDA noise from 85.49% to 89.38%, and on real rater noise from 86.45% to 89.07%.","pith_inferences":["A natural extension would be to apply the same per-class count-based threshold in semi-supervised segmentation, where unlabeled pixels could be pseudo-labeled with a curriculum that begins with easy classes; the paper's $\\sigma_c(t)$ statistic is a ready-made confidence-history measure for such a setting.","Because the threshold depends on the model's own confidence, I would expect the gains to shrink if noise is spatially correlated with object edges in a way that the model learns confidently wrong; testing on structured boundary noise would clarify the boundary of the claim.","The clean-ratio figures suggest the voting mainly helps small classes, so a diagnostic extension would be to report per-class Dice gain against per-class clean-ratio gain, checking whether the minority-class improvement is what drives the overall result."],"forward_implications":["On surgical instrument segmentation with SFDA noise, the reported average Dice is 67.51% versus 64.50% for the best previous method, with similar gains under SFDA plus morphological noise.","On retinal cup and disc segmentation with SFDA noise, the reported average Dice is 89.38% versus 85.49% for the best previous method, and on real rater noise it is 89.07% versus 86.45%.","The class-wise dynamic threshold increases most for minority classes (a 35% increase in selected Clasper pixels versus 7% for Shaft on Endovis18), directly countering the class-imbalance skew introduced by fixed-threshold denoising.","Because suspected noisy pixels are still used through Noise Balance Loss rather than discarded, the method avoids the information loss that comes from hard clean/noisy cuts, which suggests it will degrade more gracefully as the noise selector becomes less accurate.","The framework is end-to-end and evaluates with a single branch, so the improvements do not come from a larger model at test time."],"supporting_citations":[{"why":"Supplies the two-branch collaborative selection paradigm that the paper extends to segmentation with a discrepancy loss.","marker":"[9]"},{"why":"Provides the Co-Teaching+ baseline and the disagreement-based selection idea that CCV builds on.","marker":"[23]"},{"why":"Introduces symmetric cross entropy, the robust loss that Noise Balance Loss combines with cross-entropy, and serves as a baseline.","marker":"[25]"},{"why":"Motivates the curriculum-style adaptive threshold through confidence-based curriculum pseudo-labeling.","marker":"[24]"},{"why":"Defines the SFDA noise type used in the experiments and contributes a class-affinity loss correction baseline (JACS).","marker":"[3]"},{"why":"Supplies the morphological transformation noise that is combined with SFDA noise to create the SFDA+ED setting.","marker":"[4]"},{"why":"Provides the Endovis18 surgical instrument dataset and its instrument-part classes used for evaluation.","marker":"[30]"},{"why":"Provides the RIGA fundus dataset and the multi-rater annotations used as real noise and test data.","marker":"[32]"},{"why":"Serves as a discrepant collaborative training baseline and a pixel-wise denoising comparison.","marker":"[35]"}],"fun_headline_variants":["Pixel-dependent noise handled by collaborative voting and adaptive thresholds","Curriculum thresholds and noise balance lift medical segmentation Dice","Two-branch voting rescues minority classes from noisy labels","Adaptive per-class thresholds beat fixed filtering on noisy masks","Collaborative learning with curriculum selection ups Dice 3%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central premise is that the number of pixels a class can push above a confidence threshold is a faithful measure of how well that class has been learned, so lowering thresholds for low-count classes lets in clean minority-class pixels rather than just noisy ones.","fun_headline_variants_meta":{"raw":{"variants":["Pixel-dependent noise handled by collaborative voting and adaptive thresholds","Curriculum thresholds and noise balance lift medical segmentation Dice","Two-branch voting rescues minority classes from noisy labels","Adaptive per-class thresholds beat fixed filtering on noisy masks","Collaborative learning with curriculum selection ups Dice 3%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000445,"raw_usage":{"total_tokens":2301,"prompt_tokens":1048,"completion_tokens":1253,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":664,"completion_tokens_details":{"reasoning_tokens":1175}},"tokens_in":664,"tokens_out":1253,"duration_ms":10811,"temperature":1.0,"reasoning_tokens":1175,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:55:24.801319+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a model with the proposed selection but reveal the oracle clean masks during evaluation, and compute the clean-label ratio inside the selected clean set per class at several epochs. If the selected set's clean ratio is not consistently higher than the original dataset's, or if the gains over a fixed-threshold variant vanish when per-class noise rates are roughly equal, the curriculum assumption is not what drives the reported improvements.","supporting_citations":[{"cited_title":"How does disagreement help generalization against label corruption?,","cited_arxiv_id":null,"evidence_quote":"Provides the Co-Teaching+ baseline and the disagreement-based selection idea that CCV builds on."},{"cited_title":"Symmetric cross entropy for robust learning with noisy labels,","cited_arxiv_id":null,"evidence_quote":"Introduces symmetric cross entropy, the robust loss that Noise Balance Loss combines with cross-entropy, and serves as a baseline."},{"cited_title":"Flexmatch: Boosting semi-supervised learning with curriculum pseudo labeling,","cited_arxiv_id":null,"evidence_quote":"Motivates the curriculum-style adaptive threshold through confidence-based curriculum pseudo-labeling."},{"cited_title":"Joint class-affinity loss correction for robust medi- cal image segmentation with noisy labels,","cited_arxiv_id":null,"evidence_quote":"Defines the SFDA noise type used in the experiments and contributes a class-affinity loss correction baseline (JACS)."},{"cited_title":"Superpixel-guided iterative learning from noisy labels for medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the morphological transformation noise that is combined with SFDA noise to create the SFDA+ED setting."},{"cited_title":"Agreement among ophthalmologists in marking the optic disc and optic cup in fundus images,","cited_arxiv_id":null,"evidence_quote":"Provides the RIGA fundus dataset and the multi-rater annotations used as real noise and test data."},{"cited_title":"Learning from noisy labels via discrepant collaborative training,","cited_arxiv_id":null,"evidence_quote":"Serves as a discrepant collaborative training baseline and a pixel-wise denoising comparison."}],"review_version":1}