{"id":"7f9591a6-231c-4dee-b8ac-fda3c2e81f9e","arxiv_id":"2502.07302","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A consensus-aware self-corrective loss weighting method improves noisy-label cell segmentation on one real and two simulated glomerular datasets, but the false-positive correction claim is not consistently supported.","lead":"This paper introduces CASC-AI, a training method that uses agreement between an AI model and lay annotators to guide cell segmentation and suppress noisy labels. It reports improved Dice and F1 scores on one real and two simulated noisy pathology datasets, though one mechanism claim is contradicted by its own training-set table.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 2 contradicts the FP-correction claim: on the real dataset CASC-AI raises FP IoU (2.86→3.48), so the consensus-aware mechanism is not demonstrably correcting FP errors.","rationale":"The reader's verdict is CONDITIONAL, and the training-set table contradiction provides a concrete, evidence-based reason to maintain that conditionality. The reader's weakest assumption focused on confidence calibration; our concern is sharper: the paper's own Table 2 shows the FP-correction mechanism failing on the real dataset, directly contradicting Section 4.2's claim. This does not by itself overturn the empirical test-set improvements, since those could stem from other components (e.g., contrastive loss or loss weighting as implicit regularization). However, it undermines the paper's explanation of why CASC-AI works, which is central to its claimed contribution. A well-designed ablation would determine whether the consensus-aware supervision is necessary for the observed gains. If it is not, the paper's framing as a 'consensus-aware self-corrective' method is unsupported, though the empirical results may still hold value. The recommended verdict remains CONDITIONAL, exactly as the reader proposed, because the paper can address this by correcting Table 2, reporting FP/FN metrics more carefully, and adding the missing ablation.","tokens_in":12202,"tokens_out":5507,"duration_ms":50043,"concrete_test":"Recompute Table 2 for the real lay-annotated dataset, reporting FP IoU and FN IoU with standard deviations for Supervised and CASC-AI. If FP IoU is not lower for CASC-AI, run an additional ablation on the real dataset with ωc and ωsim fixed to 1 in Eq. 5 (i.e., standard supervised loss plus contrastive term only). If test-set Dice does not drop significantly relative to CASC-AI, the consensus-aware weighting is not the driver of the reported improvement, and the central claim would need revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that consensus-aware weighting identifies and corrects noisy annotations, specifically FP and FN errors, leading to improved segmentation. Section 4.2 states that CASC-AI 'reduces predictions in FP regions while increasing predictions in FN regions.' However, Table 2 (training-set performance) shows the opposite for the real lay-annotated dataset: FP IoU increases from 2.86% (Supervised) to 3.48% (CASC-AI), while FN IoU increases from 8.20% to 9.89%. Since FP IoU is a lower-is-better metric, CASC-AI worsens overlap with FP annotation regions on the very dataset for which the method is designed. This internally contradicts the claimed FP-correction mechanism. The likely cause is the paper's reliance on the model's own confidence map c (Eq. 1-2): if c is confidently wrong on noisy FP regions, those regions receive high weight in the disagreement terms and the exponential weighting ωc in Eq. 5 reinforces rather than corrects the error. This makes the reader's weakest assumption (confidence calibration) concrete and empirically supported. The test-set gains in Table 1 may therefore arise not from genuine self-correction, but from the added weighting acting as a generic regularizer or from the contrastive loss alone. The paper does not provide an ablation isolating the consensus-aware weighting, so the causal mechanism behind the reported improvements is unverified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes CASC-AI, a training-time loss-modification method for multi-class cell segmentation with noisy lay annotations. It uses the model's confidence map and the lay label to build a Consensus Matrix (CP, CN, DM, DH), distills cell and noise prototypes from high-agreement and disagreement pixels (Eqs. 1-4), reweights a Dice+BCE loss exponentially by consensus and feature similarity (Eq. 5), and adds a KL/MSE contrastive separation term (Eqs. 6-7). Experiments are conducted on one real lay-annotated glomerular dataset and two simulated FP/FN noisy datasets, with Dice and F1 on a test set and TP/FP/FN metrics on the training set. The authors report improved mean Dice and F1 over supervised training and several loss-correction baselines, and claim the method corrects false positives and false negatives during training.","tokens_in":12525,"tokens_out":5845,"duration_ms":52699,"significance":"If the test-set results are confirmed, CASC-AI would be a useful practical tool for training cell segmentation models from lay annotations without expert re-labeling, and the public release of code and annotations is a concrete contribution. The real-data test-set gains (e.g., mean Dice 73.00 vs. 66.90 for supervised training on the real dataset) are meaningful, and the comparison with multiple baselines is appropriate. However, the paper's explanatory claim that the gains arise from correcting FP and FN noise is not established: the direct training-set evidence for FP correction on the real dataset points in the opposite direction, and no ablation isolates the consensus-aware reweighting from the contrastive term. As presented, the causal mechanism behind the reported improvements remains uncertain.","major_comments":[{"comment":"The text states that CASC-AI 'reduces predictions in FP regions while increasing predictions in FN regions.' For the real lay-annotated dataset, Table 2 reports FP IoU (lower is better) increasing from 2.86 (Supervised) to 3.48 (CASC-AI), while FN IoU (higher is better) increases from 8.20 to 9.89. Thus, the FP-correction mechanism is not observed on the real dataset; in fact, the method yields more overlap with FP regions. Please reconcile this with the claim, state the exact definition of the FP IoU metric, and report whether this difference is statistically significant. This is not a presentation issue because the central 'self-corrective' claim in Section 5 relies on this training-set evidence.","section":"Section 4.2, Table 2"},{"comment":"No ablation isolates the components of L_consensus-aware. Appendix B reports an ablation on error-free labels, but it changes hyperparameters and backbones rather than testing the proposed components. Since the test-set gains in Table 1 could be due to the contrastive term alone or to the reweighting acting as a generic regularizer, please provide ablations on the noisy datasets: (a) full method, (b) Eq. (5) without Eq. (6), (c) Eq. (5) with omega_sim = 1, (d) Eq. (5) with omega_c = 1, and (e) the supervised baseline. Also report standard deviations and significance tests for Table 2, which currently lacks them.","section":"Section 2.3 and Eq. (7)"},{"comment":"The method uses the model's own confidence map c as the source of 'consensus,' but the calibration of c is not analyzed. The network providing c is trained on noisy labels, so if c is confidently wrong in DH regions, the weighting in Eq. (5) may reinforce systematic errors rather than correct them. Please report how well c identifies correct versus incorrect label regions against the pathologist-corrected labels (e.g., precision and recall of the CP, DH, and DM masks), and state whether any calibration or uncertainty estimate is applied before c is used to define consensus.","section":"Sections 2.1-2.2, Eqs. (1)-(5)"},{"comment":"The simulated FP noise in Algorithm 1 is generated from PAS-positive contours near existing cells, which aligns with the method's assumption in Eq. (3) that noise features are dissimilar to cell features. This makes the simulated FP dataset favorable to the proposed feature-similarity mechanism and limits how much it can validate the method on realistic annotation noise. Please discuss this limitation explicitly and, if possible, evaluate on a second real noisy dataset or on noise injected with a different distribution.","section":"Section 3.1 and Algorithm 1"}],"minor_comments":[{"comment":"The FP dataset row uses the baseline name 'RDice' while the real and FN rows use 'NRDice'; please unify the naming and verify that the same method was run in all three cases.","section":"Table 1"},{"comment":"The phrase 'Wilcoxon t-test' is not a standard statistical test name; please specify the exact test used (e.g., Wilcoxon signed-rank or rank-sum) and whether it is paired across the WSI-level splits.","section":"Section 3.3"},{"comment":"The contrastive loss weight is implicitly set to 1 in Eq. (7); please state whether this weight was tuned and report its sensitivity.","section":"Eq. (7)"},{"comment":"Selecting the top k/2 pixel indices assumes k is even, and the value of k is not justified or swept; please clarify the choice and its effect on the distilled features.","section":"Eq. (2)"},{"comment":"The figure uses quantities such as Sim_cell and Sim_noise that are not defined in the caption; adding a legend that maps the notation to Eqs. (1)-(4) would improve readability.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable empirical contribution with a useful public release, but the 'self-correction' narrative is stronger than the evidence. The contradictory FP IoU result in Table 2 on the real dataset and the absence of a component ablation are the main risks. If the authors can reconcile Table 2 with their claim and add ablations that isolate the consensus weighting from the contrastive term, the paper could become acceptable. I would also ask the editor to verify that the code repository and annotations are indeed publicly accessible as stated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: CASC-AI is a plausible combination of known noisy-label techniques—confidence-weighted loss, an agreement matrix, contrastive feature separation—applied to cell segmentation, with code and data released. But the paper's own Table 2 contradicts a core claim: on the real lay-annotated dataset, FP IoU gets worse (2.86 to 3.48), not better. So the 'self-correcting FP/FN' story is only half supported.\n\nWhat's new: the specific Consensus Matrix for cell segmentation, the similarity-weighted noise-feature distillation, and the contrastive separation between cell and noise features. I don't see these exact pieces in the cited prior work. The test-set Dice/F1 gains are modest (roughly +1.8 mean Dice over the strongest baseline) but consistent across three datasets, with standard deviations and Wilcoxon tests reported. That's decent evidence for a small real effect. Public code and the lay-annotation dataset are a plus.\n\nWhere it's soft: the FP contradiction is the main issue. In Section 4.2 they claim the method 'reduces predictions in FP regions,' but Table 2 shows the opposite on the real dataset. That suggests the mechanism may not be doing what they think—the gains could come from generic regularization or the contrastive loss alone. The design is also partially circular: the model's own confidence map c decides what is trustworthy, then the same model is trained with weights derived from that c. If c is confidently wrong early on, the weighting can reinforce errors. The simulated noise datasets are generated with assumptions aligned to the method (PAS-intensity-based FP injection, random contour removal for FN), so they are not strong independent validation. And the ablation study (Table 5) only varies hyperparameters and backbones; there is no ablation isolating the consensus-aware weighting or the contrastive loss. The causal mechanism is unverified.\n\nMissing modern baselines is a minor weakness—GCE, NCE+NMAE, NRDice are from 2018–2020. A recent noisy-label segmentation method would make the comparison more convincing. But this is fixable.\n\nWho this is for: researchers working on label noise in medical image segmentation, especially cell/nuclei segmentation. The released annotations alone could be useful. The paper deserves a serious referee, but it needs revision: reinterpret or fix the FP numbers, add component ablations, validate on an external noisy benchmark.\n\nRecommendation: engage with it—send to peer review with expectations of major revision.","headline":"Plausible noisy-label cell segmentation method, but the FP-correction claim is contradicted by the paper's own Table 2; worth a revised peer review.","tokens_in":13083,"tokens_out":3452,"would_cite":true,"duration_ms":29007,"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":"Training a cell-segmentation model with a consensus-aware loss that weights human-AI agreement and separates cell from noise features improves Dice and F1 on noisy lay-annotated glomerular images compared with supervised training and…","keywords":["cell segmentation","noisy labels","consensus matrix","self-corrective learning","contrastive learning","digital pathology","whole slide images","lay annotation"],"falsifier":"Take a noisy training set, train with the method, and compare the predictions in disagreement regions against pathologist ground truth per pixel. If the IoU or Dice of those disagreement-region predictions is not higher than the original noisy labels, or if reversing the weighting direction (rewarding disagreement rather than consensus) leaves performance unchanged, then the claimed consensus-aware mechanism is not doing the work.","tokens_in":12025,"feed_emoji":"🔬","tokens_out":9298,"duration_ms":74778,"temperature":0.7,"pith_summary":"This paper proposes a training strategy for cell segmentation when pixel labels are noisy because they were produced by lay annotators rather than expert pathologists. The central idea is to use the AI model's own confidence map to build a Consensus Matrix that separates pixels where model and annotator agree from pixels where they disagree. Agreement regions receive stronger supervision, while disagreement regions are down-weighted or re-weighted according to how similar their features are to high-confidence cell features; a contrastive loss then pushes noise features away from cell features. On a real lay-annotated glomerular dataset and two simulated noisy datasets, the method is reported to achieve higher Dice and F1 than direct supervised training and several loss-correction baselines. If the result stands, it means annotation noise can be turned into a learnable signal instead of an error to be ignored.","feed_headline":"Self-corrective model beats supervised training on noisy cell labels","feed_subtitle":"Weights human-AI agreement, separates cell and noise features, and beats supervised training on noisy pathology images","key_machinery":"The Consensus Matrix is the central object: a pixel-level partition of the training image into CP, CN, DM, and DH, formed by comparing the model's confidence map $c$ with the lay annotation $y_l$. From the top-$k$ consensus-positive pixels the method distills a cell feature $f_{cell}$; from disagreement pixels, weighted by cosine similarity to $f_{cell}$, it distills a noise feature $f_{noise}$. These prototypes produce similarity maps $\\mathrm{sim}_{cell}$ and $\\mathrm{sim}_{noise}$, which enter the weighted loss $L_{supervise} = (L_{Dice}+L_{BCE}) \\cdot \\exp(c \\cdot y_l + (1-c)\\cdot(1-y_l)) \\cdot \\exp(\\mathrm{sim}_{cell} - \\mathrm{sim}_{noise})$. A contrastive term $L_{contrastive} = (L_{KL}+L_{MSE})(\\mathrm{norm}(f_{cell}), \\mathrm{norm}(f_{noise}))$ separates the prototypes. The mechanism's job is to let the model decide, pixel by pixel, whether a disagreement is a missed cell or an over-annotated region, and to train itself from that decision.","core_discovery":"On its own terms, the paper shows that a cell-segmentation model can be trained on noisy lay annotations by treating the intersection of the annotator's label and the model's softmax confidence as reliable ground truth. The method defines four pixel classes: consensus positives (both model and annotator say cell), consensus negatives (both say non-cell), and two disagreement classes (model-positive/human-negative and human-positive/model-negative). The supervised loss is weighted by an exponential term that rewards consensus and by the difference between each pixel's similarity to distilled cell features and distilled noise features; a contrastive term then maximizes separation between those feature prototypes. The reported effect is that false positives in the label are suppressed and false negatives are recovered during training, yielding a mean Dice of 73.00 against 66.90 for supervised training on the real dataset, with similar gains on the two simulated noisy datasets. The paper's conclusion is that consensus-aware self-corrective supervision enhances segmentation reliability in the presence of noisy annotations.","pith_inferences":["Editorial inference: the same two-source agreement construction could transfer to other dense prediction tasks with noisy labels, such as gland, nuclei, or lesion segmentation, whenever a second weak label source is available.","Editorial inference: disagreement regions that remain low-confidence after training could be routed to an expert for targeted review, turning the method into a selective annotation tool rather than a pure loss reweighting.","Editorial inference: a testable extension is to use the consensus matrix as an offline label-correction step by relabeling disagreement pixels according to the learned prototypes, then training on the corrected labels, which would separate the correction mechanism from the loss design."],"forward_implications":["Lay annotators can replace expert raters for training cell segmentation models, with the model itself correcting false positives and false negatives during training.","The consensus-aware loss should improve not only test Dice but also training-set true-positive Dice and F1 while reducing predicted false-positive IoU and increasing false-negative IoU.","On the three reported datasets, the gains over generalized cross-entropy, normalized losses, and the previous contrastive baseline indicate that combining pixel-level and feature-level agreement is more effective than loss reweighting alone.","Because the method only needs one noisy label map and the model's own prediction, it can be applied to other partially labeled dense prediction tasks without extra supervision."],"supporting_citations":[{"why":"supplies the real lay-annotated glomerular dataset and the molecular-empowered partial-label annotation protocol the method trains on.","marker":"Deng et al., 2023"},{"why":"provides the token-based residual U-Net backbone selected by the ablation study and used in all comparisons.","marker":"Deng et al., 2024b"},{"why":"defines the generalized cross-entropy loss baseline the method must beat on all three datasets.","marker":"Zhang and Sabuncu, 2018"},{"why":"defines the NCE+NMAE normalized-loss baseline in the comparison table.","marker":"Ma et al., 2020"},{"why":"defines the NRDice noise-robust Dice baseline in the comparison table.","marker":"Wang et al., 2020"}],"fun_headline_variants":["CASC-AI self-corrects noisy labels for cell segmentation","Consensus-aware learning beats supervised on noisy cells","Self-corrective AI improves noisy cell segmentation","Noisy annotations? AI consensus fixes false positives"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The model's confidence map, produced by a network trained on noisy labels, is accurate enough that high-confidence agreements are truly correct and high-confidence disagreements are truly wrong; if the network is confidently wrong, the weighting will amplify the noise instead of correcting it.","fun_headline_variants_meta":{"raw":{"variants":["CASC-AI self-corrects noisy labels for cell segmentation","Consensus-aware learning beats supervised on noisy cells","Self-corrective AI improves noisy cell segmentation","Noisy annotations? AI consensus fixes false positives"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000211,"raw_usage":{"total_tokens":1446,"prompt_tokens":1007,"completion_tokens":439,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":623,"completion_tokens_details":{"reasoning_tokens":376}},"tokens_in":623,"tokens_out":439,"duration_ms":4534,"temperature":1.0,"reasoning_tokens":376,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T13:10:24.139818+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a noisy training set, train with the method, and compare the predictions in disagreement regions against pathologist ground truth per pixel. If the IoU or Dice of those disagreement-region predictions is not higher than the original noisy labels, or if reversing the weighting direction (rewarding disagreement rather than consensus) leaves performance unchanged, then the claimed consensus-aware mechanism is not doing the work.","supporting_citations":[{"cited_title":"Democratizing pathological image segmentation with lay annotators via molecular-empowered learning","cited_arxiv_id":null,"evidence_quote":"supplies the real lay-annotated glomerular dataset and the molecular-empowered partial-label annotation protocol the method trains on."}],"review_version":1}