{"id":"ec615548-c6a3-49cc-b752-7c9332752bd1","arxiv_id":"2502.06650","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"PCCS improves semi-supervised medical image segmentation by contrasting prototypes built from signed distance maps, with uncertainty weighting and a teacher-student prototype update.","lead":"A new training method, PCCS, uses distance-from-boundary prototypes and uncertainty weighting to improve semi-supervised medical image segmentation with few labeled images. On three datasets, including a new private breast MRI set, it reports higher Dice scores than several published baselines.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation 12's prototype update is not a valid update: the coefficients sum to 2.0, so the teacher prototype is amplified by 1.989 per step and no renormalization is specified.","rationale":"The paper's primary claim is that PCCS achieves state-of-the-art semi-supervised medical image segmentation, with the largest gains reported on a private dataset. For that claim to hold, the training procedure described in the paper must be mathematically sound and reproducible. Equation 12 is not sound as written: it is not a moving-average or convex combination, since the coefficients sum to 2.0, and the teacher prototype term is amplified rather than averaged. Because this update is the paper's newly proposed 'prototype guide prototype' mechanism and feeds directly into the auxiliary pixel-prototype loss and the uncertainty estimates, an invalid update would invalidate the reported experimental evidence. The reader's conditional verdict already calls for independent verification and flags the prototype update equation as suspicious; this concern makes that condition concrete and testable. The released code provides a direct way to check whether the implementation matches the text. If the code disagrees with Eq. 12, the manuscript must be corrected and the experiments re-run; if the code matches, the method is unstable. Therefore the appropriate verdict remains conditional, pending the specific check proposed above.","tokens_in":17178,"tokens_out":14414,"duration_ms":131877,"concrete_test":"Run the released code (https://github.com/comphsh/PCCS) with a debug hook after every prototype update and record ||p2_c|| over the first 1,000 training iterations. If the norms grow at the ~1.989^k rate or overflow, Eq. 12 is implemented literally and the method is unstable. If the code renormalizes p2_c after each update or uses a different coefficient combination (e.g., µγ p2_c + (1−µ)p1_c + (1−γ)v_c), the paper must be corrected and the main experiments re-run with the stated equation. Either outcome settles whether the described mechanism produced the reported numbers.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central 'Prototype Guide Prototype' mechanism is internally inconsistent. Equation 12 defines the teacher prototype update as p2_c = (µ+γ)p2_c + (1−µ)p1_c + (1−γ)v_c, with µ=0.99 and γ=0.999. The coefficients sum to 2.0 (1.989 + 0.01 + 0.001), so the stored teacher prototype is multiplied by about 1.989 at every update. In the absence of a stated renormalization, ||p2_c|| grows roughly as 1.989^t over training iterations, which would overflow and destabilize training. The only normalization mentioned in the paper applies to the inputs of the contrastive loss in Eq. 7, not to the update in Eq. 12 or to the teacher prototypes used in Eq. 10 and Eq. 13. If the implementation literally follows Eq. 12, the described training procedure cannot be what produced the reported SOTA Dice values; if the implementation instead normalizes or uses a different convex update, the paper's description of the core mechanism is not reproducible from the text.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes PCCS, a semi-supervised medical image segmentation method built on a student–teacher framework with three modules: uncertainty-weighted prototype contrastive learning using prototypes defined by signed distance maps, a 'prototype guide prototype' update that combines student and teacher prototypes with historical features, and an uncertainty-consistency loss that penalizes predictive uncertainty in both branches. The total loss combines supervised cross-entropy and Dice with the three auxiliary terms. Experiments are reported on BUSI, a private breast MRI dataset (BML), and ACDC, with Dice/Jaccard (and for ACDC, 95HD/ASSD) metrics under 10% and 20% labeled-data settings. The authors claim state-of-the-art performance, with the largest reported gains on the private BML dataset (e.g., 10.85% Dice improvement over SLC-Net at 20% labeled data).","tokens_in":17460,"tokens_out":3416,"duration_ms":31267,"significance":"If the results hold, the paper contributes a useful idea: using signed-distance-based boundary prototypes and uncertainty weighting to make prototype contrastive learning more robust in semi-supervised medical image segmentation. The reported gains on the private BML dataset are large and, if reproducible, would be practically valuable. The paper also ships code, which is a strength. However, the evidence is currently weakened by (i) an internally inconsistent prototype update equation that, as written, cannot be the mechanism used in training; (ii) hyperparameters selected on the test set with contradictory values across sections; and (iii) single-run results without error bars, with the largest gains coming from a private dataset. The central claim is plausible but not yet convincingly supported.","major_comments":[{"comment":"The prototype update in Eq. (12) is not a valid convex combination: the coefficients (µ+γ) + (1−µ) + (1−γ) sum to 2.0, and with µ=0.99, γ=0.999 the teacher prototype is multiplied by 1.989 each step. Since no renormalization is stated for the teacher prototypes used in Eqs. (10) and (13), the stored prototypes would grow roughly as 1.989^t and destabilize training. If the implementation instead normalizes or uses a different update, the paper's description of the core 'prototype guide prototype' mechanism is not reproducible. This must be corrected and clarified.","section":"3.3, Eq. (12)"},{"comment":"Hyperparameters are tuned on the test set: Table 5 reports Dice on the BUSI test data for different λpc, λu, and λc, rather than on a validation split, which inflates the reported performance and is a methodological concern. The values are also internally inconsistent with Section 3.1, which states λpc=0.1 and λu=0.01, while the text of Section 4.4 concludes λpc=1, λc=10^-1, and λu between 10^-3 and 10^-2. Additionally, Table 5 is corrupted (e.g., '62.3660.5561.62' and '42.763.1951.11'), making it impossible to verify the claimed optima. A proper validation-based selection protocol and a clean table are needed.","section":"4.4, Table 5"},{"comment":"All quantitative results are reported as single numbers with no error bars, no multiple seeds, and no statistical significance tests. Given that the largest claimed improvements (10.85% Dice on BML at 20% labeled data, Table 3) are on a private dataset that is not released, the reader cannot assess whether these differences are reliable or idiosyncratic to one run. At minimum, the paper should report mean±std over multiple runs and, ideally, confirm the main conclusions on a public dataset beyond BUSI and ACDC.","section":"Tables 2, 3, 6"},{"comment":"The abstract and Section 4.3 state that PCCS 'significantly outperforms' state-of-the-art methods, but Table 6 shows that on the ACDC RV class, U2PL achieves 86.07% Dice versus PCCS's 85.07%, and the average Dice gain over U2PL is only about 0.51 points. The claim is technically about average performance, but the word 'significantly' is not supported by any statistical test. The authors should either temper the claim or provide significance evidence.","section":"4.5, Table 6"}],"minor_comments":[{"comment":"The text referring to BML visualizations says 'Fig. 5 showcases the visualization outcomes ... on the BUSI dataset,' but the BML results are shown in Fig. 6; the figure reference is incorrect.","section":"4.3.2"},{"comment":"The caption contains a typo: 'Singed distance map' should be 'Signed distance map.'","section":"Figure 8"},{"comment":"The captions use 'unlabled' instead of 'unlabeled.'","section":"Tables 2 and 3 captions"},{"comment":"Equation (18) contains LaTeX remnants ('vuut') and the square-root symbols are not rendered; the formula should be rewritten cleanly.","section":"3.4, Eq. (18)"},{"comment":"References [23] and [41] appear to be the same paper (Zhang et al., 'Self-aware and cross-sample prototypical learning...') listed twice; also, 'SLC-Net' and 'SCL-Net' are used inconsistently in the text.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses a relevant problem and the code availability is commendable, but the current version has a load-bearing technical inconsistency in Eq. (12) that, taken literally, describes an unstable procedure. If this is a typographical error and the implementation uses a normalized or convex update, the paper may be salvageable after a thorough revision. However, the test-set hyperparameter tuning and lack of error bars are serious methodological issues that the editor should require the authors to address before any acceptance decision. I also note that the largest gains are on a private dataset, so the reproducibility of the headline result depends on the authors' willingness to release the data or provide a public-dataset equivalent."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe short version: PCCS is a plausible prototype-contrastive method for semi-supervised medical segmentation, but the description of its core update mechanism, Eq. 12, cannot be right. The coefficients sum to 2.0 (µ+γ = 1.989, plus 0.01 and 0.001), so the teacher prototype is multiplied by ~1.989 each step. No renormalization is stated, meaning ||p2_c|| would explode. Either the implementation differs from the text or the equation is a typo. Either way, the paper as written is not reproducible.\n\nWhat is actually new and good: the signed-distance-map-indexed prototypes are a reasonable way to focus contrastive learning on boundary structure, and the uncertainty weighting is sensible. The experiments are extensive: three datasets, ablations, visualizations, comparison with eleven baselines. On the private BML dataset the gains over SLC-Net are large (10.85 Dice at 20% labels). On ACDC the improvement over U2PL is modest (~0.5 Dice), consistent with an incremental advance.\n\nSoft spots, in order of importance. First, the Eq. 12 problem above. It is load-bearing because 'prototype guide prototype' is a stated contribution. Second, Table 5 reports hyperparameter sweeps on the test set and the chosen values contradict earlier text: Section 3.1 says λ_pc = 0.1, but Table 5 concludes λ_pc = 1 is optimal; λ_u also shifts between the 10% and 20% settings. Third, no error bars or multiple runs; single-run comparisons against strong baselines are hard to interpret. Fourth, the largest gains are on a private dataset, which limits independent verification. The self-training circularity (pseudo-labels define prototypes, and the model's own uncertainty weights them) is conventional for this field and not itself a flaw.\n\nWho should read this: anyone working on semi-supervised medical segmentation with prototype contrastive learning. It deserves a serious referee, because the core idea is plausible and the empirical scope is substantial, but it needs major revision: fix or clarify Eq. 12, align hyperparameters, add statistical rigor, and make code/data available. I would not cite it in its current form, but would revisit after a corrected version.","headline":"PCCS has a plausible pipeline and broad experiments, but Eq. 12's prototype update sums to ~2.0, so the mechanism as written cannot be what ran; needs major revision before it is reproducible.","tokens_in":17932,"tokens_out":3283,"would_cite":false,"duration_ms":27670,"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":"PCCS claims a prototype contrastive consistency method that beats prior semi-supervised medical segmentation approaches, with reported Dice gains up to 10.85% on a private breast-MRI dataset.","keywords":["semi-supervised learning","medical image segmentation","prototype contrastive learning","signed distance map","uncertainty weighting","consistency learning","student-teacher architecture","breast tumor segmentation"],"falsifier":"Take a public segmentation dataset with ground-truth masks and train PCCS with progressively corrupted pseudo-labels, for example by flipping a growing fraction of boundary pixels; if the uncertainty-weighted prototype loss stops outperforming the unweighted version under those corruptions, the claim that uncertainty-weighted boundary prototypes drive the reported gains would be falsified.","tokens_in":17008,"feed_emoji":"🩻","tokens_out":6930,"duration_ms":53143,"temperature":0.7,"pith_summary":"The paper proposes Prototype Contrastive Consistency Segmentation (PCCS), a semi-supervised method that uses unlabeled medical images to sharpen segmentation boundaries. Instead of contrasting individual pixels, it defines prototypes as the mean feature of pixels at each fixed signed distance from the predicted object boundary, and applies an uncertainty-weighted contrastive loss that pulls same-class prototypes together and pushes different-class prototypes apart. A student-teacher scheme updates prototypes using both student features and accumulated teacher history, while an uncertainty-consistency loss aligns the two branches and lowers prediction uncertainty. The central claim is that this combination segments breast ultrasound, breast MRI, and cardiac MRI more accurately than state-of-the-art semi-supervised methods, with reported Dice gains such as 10.85% over the best prior method on the private breast-MRI dataset under 20% labeled training data.","feed_headline":"Contrastive prototypes lift semi-supervised segmentation Dice 10.85%","feed_subtitle":"Distance-from-boundary prototypes and uncertainty weighting cut the amount of labeled medical imaging needed","key_machinery":"The load-bearing object is the signed-distance-map prototype, defined as the mean projected feature of all pixels whose signed distance to the predicted boundary equals a fixed value $j$, so each class contributes a sequence of prototypes ordered from boundary inward. These prototypes anchor an uncertainty-weighted prototype contrastive consistency loss whose per-prototype weight is set by the entropy of its predicted class distribution, giving more influence to confident prototypes and less to uncertain ones. The prototype-updating mechanism combines the student prototype, the teacher prototype, and accumulated teacher history in one update so the teacher does not collapse onto the student, and the uncertainty-consistency loss aligns the student and teacher probability maps while minimizing their average predictive entropy.","core_discovery":"On the paper's own terms, the discovery is that the signed distance from each pixel to the predicted boundary is a more informative coordinate for building contrastive prototypes than the raw pseudo-label class alone, because it exposes the boundary structure that is hardest to segment. Prototypes indexed by distance level make boundary pixels explicit contrastive anchors, and weighting the prototype contrastive loss by the entropy of each prototype's predicted class distribution suppresses the influence of wrong pseudo-labels. With these distance-indexed prototypes, a prototype-updating mechanism that preserves teacher history, and an uncertainty-consistency loss, PCCS reports the best average Dice on BUSI, BML, and ACDC under both 10% and 20% labeled-data settings, including a 10.85% Dice improvement over SLC-Net on BML.","pith_inferences":["An implication the paper leaves implicit is that the method's gain depends on the pseudo-label boundary being roughly correct; a natural extension would corrupt boundary pseudo-labels by a controlled amount and measure how quickly the reported advantage shrinks.","Because the BML dataset used for the headline 10.85% gain is private, a public re-run under the same protocol would be needed before treating that number as a general benchmark.","The uncertainty weighting presumes that prototype entropy is a faithful proxy for pseudo-label correctness; if the model is poorly calibrated, the weighting could suppress high-confidence but wrong boundary prototypes.","The signed-distance construction should carry over to 3D segmentation by replacing the 2D distance transform with a surface distance transform, which would make the method directly applicable to volumetric medical images."],"forward_implications":["If the reported gains hold, semi-supervised medical segmentation can use 10–20% labeled images and still match or exceed the accuracy of methods that need many more labels.","Boundary-focused prototypes should help most on lesions with irregular or low-contrast boundaries, which are exactly the cases where earlier methods show the largest errors.","The same distance-indexed prototype construction can be applied to any dense prediction task that can produce a signed distance map from its predictions.","Uncertainty weighting provides a defense against noisy pseudo-labels, so the method should be less prone to error accumulation during long training."],"supporting_citations":[{"why":"Supplies the student-teacher weight-averaging update that the paper's prototype-updating mechanism adapts.","marker":"[28]"},{"why":"Provides the uncertainty-aware consistency baseline and the 0.99 update coefficient used in prototype updating.","marker":"[29]"},{"why":"Cross Pseudo Supervision, a pseudo-label semi-supervised baseline that PCCS extends and compares against.","marker":"[10]"},{"why":"Shows how unreliable pseudo-labels can be handled; PCCS's uncertainty weighting targets the same problem and reports better results.","marker":"[19]"},{"why":"Earlier prototypical contrastive learning that treats same-class pixels as prototypes; PCCS replaces this with distance-indexed prototypes.","marker":"[22]"},{"why":"Provides the BUSI breast ultrasound dataset used for the 10% and 20% labeled-data experiments.","marker":"[36]"},{"why":"Provides the ACDC cardiac MRI dataset used to demonstrate multi-class segmentation.","marker":"[37]"},{"why":"The SLC-Net baseline that PCCS reports beating by 10.85% Dice on the private BML dataset.","marker":"[40]"}],"fun_headline_variants":["Distance-based prototypes boost semi-supervised segmentation Dice","Boundary-aware contrastive prototypes shrink labeled data needs","Uncertainty-weighted prototype contrast lifts segmentation Dice","Signed-distance prototypes improve semi-supervised medical segmentation","Prototype contrastive consistency cuts labeling for segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's gains depend on the model's pseudo-labels and uncertainty estimates being accurate enough that distance-indexed prototypes represent true semantic classes, especially near object boundaries, even when only 10% of training images are labeled.","fun_headline_variants_meta":{"raw":{"variants":["Distance-based prototypes boost semi-supervised segmentation Dice","Boundary-aware contrastive prototypes shrink labeled data needs","Uncertainty-weighted prototype contrast lifts segmentation Dice","Signed-distance prototypes improve semi-supervised medical segmentation","Prototype contrastive consistency cuts labeling for segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000448,"raw_usage":{"total_tokens":2276,"prompt_tokens":975,"completion_tokens":1301,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":591,"completion_tokens_details":{"reasoning_tokens":1229}},"tokens_in":591,"tokens_out":1301,"duration_ms":9194,"temperature":1.0,"reasoning_tokens":1229,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T14:46:26.137668+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a public segmentation dataset with ground-truth masks and train PCCS with progressively corrupted pseudo-labels, for example by flipping a growing fraction of boundary pixels; if the uncertainty-weighted prototype loss stops outperforming the unweighted version under those corruptions, the claim that uncertainty-weighted boundary prototypes drive the reported gains would be falsified.","supporting_citations":[{"cited_title":"Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation","cited_arxiv_id":null,"evidence_quote":"Provides the uncertainty-aware consistency baseline and the 0.99 update coefficient used in prototype updating."},{"cited_title":"Semi-supervised semantic segmentation with cross pseudo supervision","cited_arxiv_id":null,"evidence_quote":"Cross Pseudo Supervision, a pseudo-label semi-supervised baseline that PCCS extends and compares against."},{"cited_title":"Semi-supervised semantic segmentation using unreliable pseudo-labels","cited_arxiv_id":null,"evidence_quote":"Shows how unreliable pseudo-labels can be handled; PCCS's uncertainty weighting targets the same problem and reports better results."},{"cited_title":"Semi-supervised semantic segmentation via prototypical contrastive learning","cited_arxiv_id":null,"evidence_quote":"Earlier prototypical contrastive learning that treats same-class pixels as prototypes; PCCS replaces this with distance-indexed prototypes."},{"cited_title":"Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved? IEEE Transactions on Medical Imaging, 37:2514–2525, 2018","cited_arxiv_id":null,"evidence_quote":"Provides the ACDC cardiac MRI dataset used to demonstrate multi-class segmentation."},{"cited_title":"Semi-supervised medical image segmentation using cross-model pseudo-supervision with shape awareness and local context constraints","cited_arxiv_id":null,"evidence_quote":"The SLC-Net baseline that PCCS reports beating by 10.85% Dice on the private BML dataset."}],"review_version":1}