{"id":"dd67c679-90e4-46bd-afca-fae272eab42d","arxiv_id":"2501.01658","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"An annotation method using two nested polygons plus a confidence-guided consistency learner achieves state-of-the-art weakly-supervised segmentation on two medical datasets and slightly beats fully supervised baselines.","lead":"This paper introduces a weakly-supervised medical image segmentation method where a lesion is marked with two polygons, one inside and one outside the boundary, and a training framework that uses the uncertain band between them. It claims to beat both other weak-label methods and fully-supervised models on two public datasets while using less than 20% of annotation time.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"BPAnno masks in Table I are generated from ground-truth masks (§IV-A2), so the claimed superiority over fully supervised training may reflect leaked boundary information rather than the weak annotation method.","rationale":"Good-faith reading: the authors propose a reasonable weak-annotation idea and a consistency-learning framework; the confidence-generator equations 10-11 are indeed garbled and the reader is right that the method is underspecified. However, the single most load-bearing condition for the central claim is that the weak annotation is actually what the paper says it is: 'simply labeling two polygons.' The experiments violate that condition by generating the polygons from GT masks. This is not a stylistic quibble: the entire quantitative advantage over fully supervised methods (Dice 86.60 vs 86.25 on ISIC2017; 89.88 vs 89.64 on Kvasir-SEG) is within the range that could be explained by boundary information inherited from GT. A manual 'refinement' step makes the leakage worse, and the single-annotator cost study cannot bridge the gap. A concrete rerun with human-drawn polygons would settle whether the method works as claimed. Because this concern undermines the headline independently of the Eq. 10-11 issue, the reader's rejection is appropriate; no verdict change is needed.","tokens_in":15991,"tokens_out":5574,"duration_ms":57447,"concrete_test":"Re-run Table I with BPAnno masks produced independently by human annotators who see only the images (no access to ground-truth masks), using the same EAUWSeg training recipe on ISIC2017 and Kvasir-SEG. Also compute Dice/IOU between each annotator's polygons and the GT-derived BPAnno masks on the same images. If human-drawn masks differ substantially from GT-derived masks, or if EAUWSeg's Dice drops below the fully supervised baselines in Table I, the central claim is not supported. Ideally use at least three annotators and report inter-annotator agreement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing flaw is the evaluation protocol for the central claim, not the pseudo-label reliability. Section IV-A2 says the bounded polygons are created by dilation-erosion of the available ground-truth masks, followed by manual refinement. Thus the BPAnno masks used to train the models in Table I are not independent weak annotations: they are derived from the exact dense masks used by the fully supervised baselines, and the refinement step can only tighten them toward the GT boundary. The paper's headline comparison therefore does not test 'two polygons drawn by a clinician'; it tests a GT-informed envelope. The annotation-cost study in Section IV-E uses a single dermatologist on 20 images and never checks whether those manually drawn polygons match the GT-derived ones used in training. Even if the Eq. 10-11 confidence-generator formulas were corrected, the empirical claim that EAUWSeg with BPAnno outperforms fully supervised training with <20% workload would remain unsupported because the training masks in the comparison are not produced by the claimed annotation workflow.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BPAnno, a weakly-supervised annotation scheme in which each lesion is annotated by two polygons, an inscribed-like polygon and an envelope-like polygon, so that the region between them is treated as uncertain. The accompanying EAUWSeg framework explicitly treats the two polygons as separate supervision signals, and adds a classification-guided confidence generator (CCG) and a confidence-auxiliary consistency learner (CCL) to produce pseudo-supervision for the uncertain band. Experiments on ISIC2017 and Kvasir-SEG compare EAUWSeg with scribble-, box-, and fully-supervised baselines, and the paper reports that EAUWSeg outperforms fully supervised training with less than 20% of the annotation workload. The paper also includes ablations, a semi-supervised comparison, a cross-dataset generalization study on ISIC2018, and an annotation-cost study.","tokens_in":16206,"tokens_out":7014,"duration_ms":70936,"significance":"The bounded-polygon annotation idea is potentially attractive: it is a natural middle ground between dense masks and boxes, and the proposed framework is modular and applicable to standard segmentation backbones. The paper provides extensive experiments, ablations across components, backbone generalization, and a cross-dataset generalization check, which are useful strengths. However, the central empirical claim is not supported by the current evaluation protocol: the BPAnno masks used for training are generated from the ground-truth masks, so the comparison against fully supervised training does not test the claimed manual two-polygon annotation workflow. In addition, the confidence-generator equations are internally inconsistent, so the core mechanism is underspecified. If both issues were fixed, the method could be of interest, but as submitted the headline contributions are not established.","major_comments":[{"comment":"The central empirical claim is undermined by the annotation-generation protocol. Section IV-A2 states that the bounded-polygon annotations are initially generated by dilation and erosion of the available ground-truth masks, then simplified with the Douglas-Peucker algorithm and manually refined. Thus the BPAnno masks used to train the models in Table I are not independent weak annotations drawn by a clinician; they are a ground-truth-informed envelope around the exact masks used by the fully-supervised baselines. The abstract and conclusion claim that EAUWSeg 'outperforms fully-supervised counterparts' with 'less than 20% of the annotation workload,' but the comparison cannot separate the method's contribution from leaked boundary information in the supervision masks. The annotation-cost study in Section IV-E uses one dermatologist on twenty images and never verifies that manually drawn polygons agree with the ground-truth-derived polygons used in the actual training experiments, so it does not repair this gap.","section":"§IV-A2, Table I, §IV-E"},{"comment":"The classification-guided confidence generator is not internally consistent. Eq. (10) defines U_c = argmax(P(y=0|f_S(x), Θ_c), P(y=2|f_S(x), Θ_c)) ⊙ M_u, which takes the argmax over only the background and foreground classes. Consequently, it cannot ever mark a pixel as belonging to class 1, even though the text says that pixels with prediction of 1 for the multi-class classification task are treated as solid uncertain. Moreover, argmax returns a class index, not a confidence value. The situation is compounded in Eq. (11): U = min(U_c + 2U_e, -1) ⊙ M_u, where U_c ∈ {0,2} and U_e ∈ {0,-1}; with these ranges the expression collapses to -1 or -2 for nearly all combinations, making the intended distinction between 'solid uncertain' and 'certain' pixels ill-defined. Because the confidence-auxiliary consistency learner in Section III-D selects pseudo-labels from this map, the central mechanism is not reproducibly specified as written.","section":"§III-C, Eqs. (10)-(11)"},{"comment":"The pseudo-label construction inherits the inconsistency of the confidence map. In Eq. (12), for pixels in the uncertain band (M_u=1), the pseudo-label becomes U, which can take value -2 according to Eq. (11); however, the text only describes -1 as the label for solid uncertain pixels. The pixel-wise contrastive loss in Eq. (13) then treats these values as class labels for positive/negative sampling, but no rule is given for interpreting -2 or for deciding whether such pixels should be included, excluded, or reweighted. This makes the sample-selection criterion for the contrastive learner incomplete.","section":"§III-D, Eq. (12)"}],"minor_comments":[{"comment":"Eq. (8) declares U_e to be in R^{C×H×W}, but the right-hand side assigns scalar values of -1 and 0; please clarify the intended shape and how the uncertainty map relates to the channel dimension.","section":"§III-C, Eq. (8)"},{"comment":"The figure caption and diagram use terms such as 'solid classes,' 'solid error predictions,' and 'Store the embedding in solid classes' without defining 'solid' in the text; please define the terminology consistently.","section":"Fig. 2"},{"comment":"The abstract and conclusion state that EAUWSeg 'outperforms fully-supervised counterparts,' but Section IV-C5 and Table V report that the cross-dataset generalization performance is 'comparable' rather than superior; the wording should be aligned with the actual results.","section":"Abstract and Conclusion"},{"comment":"The scribble annotations for the comparison are generated by connecting random endpoints sampled from the ground-truth mask, which may not reflect how clinicians draw scribbles in practice; this should be acknowledged as a limitation of the comparison protocol.","section":"§IV-A2"}],"recommendation":"reject","confidential_remarks":"The grounds for this recommendation are the two load-bearing problems: the training masks in the headline comparison are derived from the ground-truth masks rather than from the claimed manual two-polygon workflow, and the confidence-generator equations in Section III-C are internally inconsistent. The first issue makes the paper's central claim about outperforming fully supervised training with a low-cost annotation protocol unsupported, and the second issue means the proposed mechanism is not adequately specified even for a reader who accepts the evaluation protocol. A revision that replaces the experiments with genuinely independent manual BPAnno annotations and repairs Eqs. (10)-(12) would be needed before the contribution can be assessed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe short version: the paper has a genuinely usable idea — two nested polygons as weak supervision, with a contrastive learner that uses the band between polygons as an uncertainty region — but the empirical case for it is built on annotations derived from the very ground-truth masks it claims to beat, and the core confidence formulas in Eqs. 10–11 don't do what the text says. The central claims, as submitted, don't hold.\n\nWhat's new: the BPAnno protocol is a clean twist on trimap-style weak supervision, and pairing it with uncertainty-guided pixel contrastive learning is a reasonable design. The ablations show consistent improvements when CCL/CCG are added, and the cross-dataset generalization to ISIC2018 is a nice check. The authors also include an annotation-cost study, which is the right kind of evidence to collect. The paper is clearly written and the method is described in enough detail that a reader can spot the problems — no code or data release, though.\n\nWhere it falls apart. First, section IV-A2 says the BPAnno masks are produced by dilation/erosion of the available ground-truth masks, then refined. That means the 'weak' annotations in Table I know the exact GT boundary; the comparison against fully supervised training is not testing 'two polygons drawn by a clinician.' The 18% workload figure comes from one dermatologist on 20 images, and there's no check that those manual polygons match the GT-derived ones used in training. So the headline 'outperforms fully supervised with <20% workload' is unsupported.\n\nSecond, the confidence generator is internally inconsistent. Eq. 10 takes argmax over P(y=0) and P(y=2), so it can never output the class-1 'uncertain' prediction that the text says marks solid uncertain pixels. Eq. 11 is min(U^c + 2U^e, −1), which collapses to −1 for every value in the uncertain band. Either the equations or the text is wrong, and the method is unreproducible as written.\n\nThe margins over fully supervised baselines are also tiny (0.3–0.6 Dice) with overlapping standard deviations and no significance testing, so even the measured gap is fragile.\n\nWho this is for: a reader working on weakly supervised medical segmentation will find the BPAnno idea worth thinking about, and the framework is a plausible starting point. But this version shouldn't be published. I'd send it to reviewers — the idea deserves a careful look — but I'd expect rejection on the current evidence. If the authors fix the formulas and re-run the comparison with genuinely manual two-polygon annotations on a larger set, it could become a solid paper.","headline":"Good idea, broken evidence: the BPAnno masks are GT-derived and the core confidence formulas are internally inconsistent.","tokens_in":16710,"tokens_out":5132,"would_cite":false,"duration_ms":43859,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that labeling a lesion with two rough polygons—one inside, one outside the true boundary—and training with a confidence-guided contrastive consistency learner lets weakly supervised segmentation match or exceed fully…","keywords":["weakly-supervised segmentation","bounded polygon annotation","medical image segmentation","contrastive learning","consistency learning","annotation uncertainty","skin lesion segmentation","polyp segmentation"],"falsifier":"Measure the CCG's three-class predictions in the $\\Omega_\\Delta$ band against dense ground-truth masks on a held-out set: if the assigned 'certain' labels in the band agree with ground truth no better than chance, then the confidence-guided contrastive learning is reinforcing noise and the reported gains must come from some other component.","tokens_in":15787,"feed_emoji":"🩺","tokens_out":6033,"duration_ms":54839,"temperature":0.7,"pith_summary":"EAUWSeg sets out to close the gap between weakly supervised and fully supervised medical image segmentation by changing the annotation itself: instead of a dense pixel mask, a specialist draws two rough polygons around a lesion, one inscribed and one envelope-like. Because the pixels between the two polygons are genuinely uncertain, the paper argues that treating the two polygons as separate adversarial labels and then using a classification-guided confidence generator plus a contrastive consistency learner can turn most of that uncertain band into reliable supervision. On ISIC2017 and Kvasir-SEG the resulting models match or beat fully supervised counterparts while requiring roughly 18% of the annotation time, which is why the claim matters: it promises near-full supervision accuracy at a fraction of the labeling cost.","feed_headline":"Two rough polygons match full pixel masks in segmentation","feed_subtitle":"Skin-lesion and polyp models trained on two-polygon labels beat fully supervised baselines at one-fifth the annotation time.","key_machinery":"The load-bearing object is the bounded polygon annotation (BPAnno), which splits the image into certain foreground $\\Omega_I$ (inside the inscribed polygon), certain background $\\Omega_O$ (outside the envelope polygon), and an uncertain band $\\Omega_\\Delta$ between them. The framework then uses: (1) two dice losses treating the two polygons as separate labels, so the band provides an adversarial training signal; (2) a classification-guided confidence generator (CCG) that predicts a three-class label per pixel and combines classification confidence with predictive entropy into a confidence map $U$; and (3) a confidence-auxiliary consistency learner (CCL) that runs pixel-wise contrastive learning using only certain pixels and confident uncertain pixels, storing certain-class embeddings in a memory queue. The contrastive loss is the mechanism that actually propagates reliable supervision into the uncertain band; CCG decides which uncertain pixels deserve to be treated as certain.","core_discovery":"The central discovery is that a binary lesion segmentation can be supervised by two nested polygon annotations that bracket the true boundary, and the uncertainty they create can be mostly resolved during training rather than by extra annotation. The paper's EAUWSeg framework trains a segmentation network with two dice losses, one for the inscribed polygon and one for the envelope polygon, so pixels between them receive contradictory labels that push the model to learn boundary-invariant features. A multi-class classification head (background, uncertain band, foreground) then estimates the confidence of each uncertain pixel, and a pixel-wise contrastive loss with a memory queue pulls same-class uncertain pixels toward certain-class features while pushing different classes apart. Reported results on ISIC2017 and Kvasir-SEG show Dice scores of 86.60% and 89.88% with TransUNet, above the fully supervised baselines in the same table, and the modules used to generate confidence are discarded at inference so runtime is unchanged.","pith_inferences":["As an extension, the two-polygon annotation could be re-used as a general weak-label format: any scribble or bounding-box annotation can be expanded into an inner and outer region, and the same CCG/CCL machinery may convert the uncertain band into supervision.","The 18% annotation-time figure comes from a single dermatologist on 20 ISIC2017 images; a multi-rater study with varied lesion shapes and image qualities would show whether the workload saving holds in practice.","If the confidence generator is reliable, EAUWSeg should combine naturally with semi-supervised learning on unlabeled images; the remaining gap to the best semi-supervised method suggests that unlabeled data could push it past fully supervised performance.","A straightforward testable extension is to apply the same bounded-polygon supervision to multi-class segmentation; the paper itself notes that an envelope polygon containing multiple foreground classes is an unsolved challenge."],"forward_implications":["Specialists can annotate skin lesions and polyps by drawing two rough polygons instead of a dense mask, cutting annotation time to roughly 18% while keeping or improving Dice scores.","Because CCL and CCG are removed at inference, the trained model runs at the same cost as the base segmentation network such as UNet or TransUNet.","The bounded-annotation strategy transfers across backbone architectures and, when trained on ISIC2017, generalizes to ISIC2018 without fine-tuning at nearly the same Dice as fully supervised models.","On both datasets, EAUWSeg outperforms scribble-, box-, and rectangle-supervised baselines, so the two-polygon annotation is a stronger weak-label format than those alternatives."],"supporting_citations":[{"why":"Supplies the feature-similarity assumption that similar pixels in feature space prefer consistent category predictions, which motivates the confidence generator.","marker":"[10]"},{"why":"ISIC2017 is the primary skin-lesion segmentation dataset used for evaluation and comparison.","marker":"[11]"},{"why":"Kvasir-SEG is the polyp segmentation dataset used for the second main evaluation.","marker":"[12]"},{"why":"Prior work injecting prediction uncertainty into pixel contrastive sampling, which the confidence-guided contrastive learner builds on.","marker":"[24]"},{"why":"UNet is a backbone network used in the experiments and as a fully supervised baseline.","marker":"[25]"},{"why":"TransUNet is the backbone used for the main reported results and the ablation studies.","marker":"[27]"},{"why":"Partial cross-entropy loss represents the scribble-supervised baseline that EAUWSeg must outperform.","marker":"[33]"},{"why":"Douglas-Peucker algorithm is used to generate polygon contours from masks for the BPAnno annotation process.","marker":"[40]"},{"why":"CASSL serves as the semi-supervised comparison method in the table showing EAUWSeg's performance with only bounded polygon labels.","marker":"[43]"}],"fun_headline_variants":["Two polygon labels surpass full masks in segmentation","Double-polygon annotation beats pixel-wise supervision","Nested polygons resolve weak-label uncertainty","Bounded polygons deliver supervised-level segmentation","Two-polygon labels cut annotation time by 80% with better Dice"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the classification head and the entropy filter can reliably identify which pixels in the band between the two polygons are true lesion and which are background, so that the contrastive learner is pulling correct samples together rather than amplifying mislabeled pixels.","fun_headline_variants_meta":{"raw":{"variants":["Two polygon labels surpass full masks in segmentation","Double-polygon annotation beats pixel-wise supervision","Nested polygons resolve weak-label uncertainty","Bounded polygons deliver supervised-level segmentation","Two-polygon labels cut annotation time by 80% with better Dice"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000594,"raw_usage":{"total_tokens":2792,"prompt_tokens":964,"completion_tokens":1828,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":580,"completion_tokens_details":{"reasoning_tokens":1756}},"tokens_in":580,"tokens_out":1828,"duration_ms":14460,"temperature":1.0,"reasoning_tokens":1756,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:22:22.763102+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the CCG's three-class predictions in the $\\Omega_\\Delta$ band against dense ground-truth masks on a held-out set: if the assigned 'certain' labels in the band agree with ground truth no better than chance, then the confidence-guided contrastive learning is reinforcing noise and the reported gains must come from some other component.","supporting_citations":[{"cited_title":"Sparsely annotated semantic segmentation with adaptive gaussian mixtures,","cited_arxiv_id":null,"evidence_quote":"Supplies the feature-similarity assumption that similar pixels in feature space prefer consistent category predictions, which motivates the confidence generator."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"ISIC2017 is the primary skin-lesion segmentation dataset used for evaluation and comparison."},{"cited_title":"Kvasir-seg: A segmented polyp dataset,","cited_arxiv_id":null,"evidence_quote":"Kvasir-SEG is the polyp segmentation dataset used for the second main evaluation."},{"cited_title":"Uncertainty-guided pixel contrastive learning for semi-supervised medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Prior work injecting prediction uncertainty into pixel contrastive sampling, which the confidence-guided contrastive learner builds on."},{"cited_title":"Transunet: Rethinking the u-net architecture design for medical image segmentation through the lens of transformers,","cited_arxiv_id":null,"evidence_quote":"TransUNet is the backbone used for the main reported results and the ablation studies."},{"cited_title":"Normal- ized cut loss for weakly-supervised cnn segmentation,","cited_arxiv_id":null,"evidence_quote":"Partial cross-entropy loss represents the scribble-supervised baseline that EAUWSeg must outperform."},{"cited_title":"Algorithms for the reduction of the number of points required to represent a digitized line or its caricature,","cited_arxiv_id":null,"evidence_quote":"Douglas-Peucker algorithm is used to generate polygon contours from masks for the BPAnno annotation process."},{"cited_title":"Consistency and adversarial semi-supervised learning for medical image segmentation,","cited_arxiv_id":null,"evidence_quote":"CASSL serves as the semi-supervised comparison method in the table showing EAUWSeg's performance with only bounded polygon labels."}],"review_version":1}