{"id":"3c863a11-3a60-4881-ae75-caf9b28af119","arxiv_id":"2412.02314","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"LoCo, a mean-teacher semi-supervised segmentation framework with low-contrast contrastive learning and a confidence-based dynamic pseudo-label filter, reports state-of-the-art results on laryngeal cancer and polyp datasets.","lead":"Researchers propose LoCo, a semi-supervised training method for endoscopic image segmentation that uses contrastive learning to focus on low-contrast pixels, such as faint tumor boundaries and similar-looking tumor types, plus a dynamic threshold to make better use of unlabeled images. The method reports higher segmentation accuracy than several prior semi-supervised approaches on a proprietary laryngeal cancer dataset and two public polyp datasets, with source code released.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline SOTA claim is not statistically supported: LoCo trails AugSeg on CVC-ClinicDB at 10% labels, and no error bars or significance tests are reported anywhere.","rationale":"The reader's verdict is already CONDITIONAL, and the reader's rationale mentions missing error bars and the CVC-ClinicDB result. However, the reader's stated weakest assumption is the ICE prototype reliability. I do not dispute that assumption, but I see the most load-bearing concern as narrower and more direct: the empirical superiority claim itself lacks statistical support. The reported margins are small, one comparison goes the wrong way, and no variance information is given. This affects the central claim regardless of whether ICE is theoretically sound. The concrete test I propose would settle whether the claimed superiority is real or within noise. Because the reader already conditioned the verdict on error bars and independent replication, my read does not move the verdict; it reinforces the existing CONDITIONAL status. I mark agreement as 'partial' because I am pointing to the statistical robustness issue as the key concern rather than the ICE prototype mechanism, though both are related and both appear in the reader's rationale.","tokens_in":16216,"tokens_out":7283,"duration_ms":85453,"concrete_test":"Run LoCo and AugSeg (the strongest baseline) for 5 independent training runs on the 10% labeled split of CVC-ClinicDB and Kvasir-SEG, using identical data splits and evaluation code. Report mean ± std for mIoU, DSC, and NSD, and compute paired bootstrap 95% confidence intervals on the per-image metric differences. If the CVC-ClinicDB difference is not significant or favors AugSeg, the abstract's 'significantly outperforming' claim is unsupported and should be softened to a qualified claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that LoCo 'significantly outperforms' previous methods, but the evidence is a single run per setting. The reported margins are small (1-3 mIoU) and one cell contradicts the claim: on CVC-ClinicDB with 10% labels (Table III), LoCo mIoU is 83.81 versus AugSeg's 84.13, and LoCo is also lower on DSC and NSD. The paper acknowledges this as 'on par', yet the abstract claims significant superiority. Without multiple seeds, standard deviations, or paired tests, a 2-3 point difference—especially on a 62-image test set—cannot be distinguished from training noise. This is load-bearing because the ablation (Table IV) attributes most of the gain to ICE (M5 vs M2), and ICE relies on per-batch class prototypes computed from only 8 labeled images; under a 10% label partition these prototypes may be unstable, so the reported advantage could be an artifact of a particular initialization or batch composition. The proprietary FAHSYU-LC dataset further prevents external verification of the largest claimed gains. The method may well work, but the strength of the claim is not matched by the statistical evidence.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LoCo, a mean-teacher framework for semi-supervised endoscopic image segmentation. The method combines a supervised cross-entropy loss with an unsupervised consistency loss on pseudo-labels filtered by a confidence-based dynamic filter (CDF), and adds a low-contrast-enhanced contrastive loss (LCC) built from two selection strategies: inter-class contrast enhancement (ICE), which picks pixels with the lowest similarity to their class prototype, and boundary contrast enhancement (BCE), which picks boundary pixels with the highest local feature similarity. Experiments are reported on one proprietary laryngeal-cancer dataset (FAHSYU-LC) and two public polyp datasets (Kvasir-SEG, CVC-ClinicDB) under 10%, 30%, and 50% labeled-data partitions, against FixMatch, U2PL, UniMatch, AugSeg, and CorrMatch. The paper claims state-of-the-art performance, significantly outperforming previous methods, and provides an ablation study on FAHSYU-LC at 10% labels.","tokens_in":16414,"tokens_out":3488,"duration_ms":39320,"significance":"The problem addressed is relevant and timely: semi-supervised endoscopic segmentation with low-contrast boundaries and minority tumor classes is practically important, and the paper tackles an underexplored inter-class contrast issue. The proposed components are clearly motivated, the framework is simple to combine with existing mean-teacher pipelines, and the authors provide source code, which strengthens reproducibility. If the reported gains are reproducible, the method would be a meaningful advance, particularly for low-contrast and minority-class pixels. However, the evidence as presented is not sufficient to support the strength of the headline claim, because the comparisons rest on single runs and at least one reported cell contradicts the claim of consistent superiority. The largest gains are on a proprietary dataset, which limits external verification.","major_comments":[{"comment":"The central claim that LoCo 'significantly outperforms previous methods' is not statistically supported. Tables I-III report a single run per setting with no error bars, no standard deviations, and no significance tests. Moreover, on CVC-ClinicDB at 10% labels (Table III), LoCo is below AugSeg on all three metrics: mIoU 83.81 vs. 84.13, DSC 76.31 vs. 78.49, and NSD 79.67 vs. 81.81; the text acknowledges this as 'on par', which is inconsistent with the abstract and conclusion. Because the margins elsewhere are 1-3 mIoU points, a paired test or multiple-seed reporting is needed to distinguish the claimed improvement from training noise. Please provide mean and standard deviation over at least three seeds and a paired significance test, and temper the claims accordingly.","section":"IV-B, Tables I-III"},{"comment":"The ICE selection rule relies on class embeddings computed by class-wise average pooling over the labeled batch (Eq. 9), and the batch size is 8 (Section IV-A3). Under a 10% label partition, the benign class may be represented by very few pixels per batch, making the prototype unstable and the top-k% lowest-similarity selection potentially driven by noise or mislabeled features. Since the ablation in Table IV attributes the largest gain to ICE (M5 vs. M2: mIoU 61.64 vs. 58.05), this stability concern is load-bearing. Please provide an analysis of prototype sensitivity, such as varying k, reporting performance across seeds, or showing that the selected pixels are consistently low-contrast rather than noise.","section":"III-B1, Eqs. (9) and (12)"},{"comment":"The ablation study is reported without variance, yet several component contributions are small and have mixed directions. For example, M4 (adding BCE to M2) increases mIoU by 0.65 but decreases IoU(M) from 70.42 to 69.91, and M3 (adding CDF to M2) decreases IoU(M) from 70.42 to 70.18. Without multiple runs, it is impossible to determine whether these differences are real or within noise. Please report standard deviations or significance tests for the ablation, and discuss which contributions are robust.","section":"IV-C, Table IV"}],"minor_comments":[{"comment":"The description of weak perturbation says 'including random scale, random scale, and random flip'; the duplicate 'random scale' appears to be a typo, and the intended augmentation list should be stated precisely.","section":"III-A, Eq. (3) and surrounding text"},{"comment":"In Eq. (14), the confidence term max(ˆy^u_{i,j}) uses the student prediction on the strongly augmented image, while the mask 1[yp_{i,j}=c] is derived from the teacher pseudo-label. Please clarify whether the confidence is intentionally taken from the student output; if so, justify this choice, and if not, correct the equation to use the teacher prediction.","section":"III-C, Eqs. (14) and (17)"},{"comment":"The definition of the low-contrast feature embedding set H is informal. It is not specified whether the top k% selection in ICE and BCE is performed per class, per image, or globally across the batch, nor whether the same k is applied independently to labeled and unlabeled features. Please clarify the exact selection procedure.","section":"III-B, III-B1, III-B2"},{"comment":"The ablation table would be easier to interpret if it also reported the supervised-only baseline with the same pseudo-label and contrastive components removed stepwise; currently M2 already includes Lu, and the contributions of the individual losses relative to M1 are conflated with the effect of unlabeled data.","section":"IV-C, Table IV"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a competent empirical study with a plausible method, but the statistical gap between the claims and the evidence is the main obstacle. The single-run comparison, the one contradicting cell on CVC-ClinicDB at 10% labels, and the mixed ablation results mean the paper needs multiple-seed experiments and significance testing before the SOTA claim can be accepted. The reliance on a proprietary dataset for the largest gains also limits reproducibility; I would encourage the editor to request the authors to either release a public subset or provide additional public-dataset validation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read the paper. My take: the contribution is real but modest, and the writing overstates it. LoCo isn't a new paradigm; it's a thoughtful combination of known components: density-guided hard-pixel contrastive selection from Hunting Sparsity, boundary-pixel selection, and a FreeMatch-style dynamic confidence threshold applied to pseudo-labels. The specific combination is new in this niche, and the three-class proprietary laryngeal dataset with pixel-level annotations is genuinely useful. The ablation separating CDF, BCE, ICE is one of the better parts of the paper—it makes the attribution clear. The code link is a plus. Related-work coverage is solid, too.\n\nThe soft spots are real. Most important: not a single standard deviation, error bar, or significance test in Tables I-III. The margins over AugSeg are mostly 1–3 mIoU, and on a 62-image test set one run can't distinguish that from training noise. The abstract's 'significantly outperforming' is not supported by what's reported, especially since Table III at 10% labels has LoCo at 83.81 mIoU and AugSeg at 84.13—the text concedes 'on par' at that cell. That's not a fatal flaw, but it is a mismatch between claim and evidence that needs to be fixed.\n\nSecond, the ablation attributes most of the improvement to ICE (M5 vs M2, 61.64 vs 58.05 mIoU). ICE computes class prototypes by average-pooling over the labeled batch. At 10% labels with batch size 8, a given class may have very few pixels; the paper doesn't say how missing or tiny classes are handled. Prototype instability under those conditions is a plausible source of noise, and there is no sensitivity analysis over k, h, gamma, or tau. This is worth addressing in a revision rather than being fatal.\n\nThe hyperparameter values are given, but no sensitivity; the proprietary dataset is not available, so the largest claimed gains can't be independently checked. Those are limitations, not crimes.\n\nWho's it for: people working on semi-supervised endoscopic or medical segmentation will find the selection rules worth knowing, and the visualizations and ablations are helpful. It deserves a serious referee rather than a desk reject, but the referee should ask for multi-seed runs, error bars, sensitivity analysis, and a claim rewrite. I'd take it to a reading group only if someone in the group is actively working on SSL in medical imaging.","headline":"LoCo is a sensible, incremental combination of density-guided contrastive learning and adaptive pseudo-label filtering for semi-supervised endoscopic segmentation, but the SOTA claim outruns the statistics: no error bars anywhere and one public-dataset cell actually trails AugSeg.","tokens_in":16987,"tokens_out":5973,"would_cite":true,"duration_ms":59746,"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 deliberately mining low-contrast pixels for contrastive learning is what pushes semi-supervised endoscopic segmentation past previous methods, especially for minority tumor classes.","keywords":["semi-supervised semantic segmentation","contrastive learning","endoscopic image segmentation","low-contrast pixels","pseudo-label filtering","mean-teacher","minority class","medical image segmentation"],"falsifier":"Run LoCo on the laryngeal cancer dataset under a 10% label partition but replace the ICE pixel-selection rule with random pixel selection at the same $k\\%$, keeping BCE and CDF fixed; the paper's claim predicts a large drop in benign-class IoU relative to the full model, while a noise-sensitivity alternative predicts little change. A second check is to compare the selected low-similarity pixels against expert-annotated low-contrast regions: if the selection is doing its job, most selected pixels should fall inside or immediately beside regions experts mark as low contrast.","tokens_in":15974,"feed_emoji":"🔬","tokens_out":9821,"duration_ms":92183,"temperature":0.7,"pith_summary":"Semi-supervised endoscopic image segmentation struggles when the boundary between tumor and normal tissue is faint and when benign and malignant tumors look alike. This paper argues that the pixels the model finds hardest to classify—low-contrast pixels—should be the primary source of contrastive supervision, not all pixels equally. It proposes LoCo, a mean-teacher framework whose low-contrast-enhanced contrastive loss (LCC) is built from two selection rules: inter-class contrast enhancement (ICE) picks pixels whose features sit farthest from their class average, and boundary contrast enhancement (BCE) picks boundary pixels whose features are most similar to their neighbours. A confidence-based dynamic filter (CDF) then adjusts pseudo-label thresholds per class so rare classes such as benign tumors contribute more unlabeled supervision. On a proprietary laryngeal cancer dataset and two public polyp datasets, LoCo reports the highest mIoU among the compared methods at 10%, 30%, and 50% labeled partitions, with the largest margins on the minority benign class.","feed_headline":"Picking the hardest pixels lifts semi-supervised endoscopy AI","feed_subtitle":"LoCo targets low-contrast tumor pixels and minority classes, beating prior methods with 10–50% labeled data.","key_machinery":"The load-bearing machinery is the low-contrast-enhanced contrastive loss (LCC), a per-pixel contrastive objective whose samples are chosen by two selection rules rather than drawn uniformly. Inter-class contrast enhancement (ICE) uses class embeddings $\\bar{z}_c$ obtained by class-wise average pooling of labeled features (Eq. 9) and keeps the top $k\\%$ of pixels with lowest cosine similarity to their own class embedding (Eq. 12). Boundary contrast enhancement (BCE) defines boundary pixels by a neighbourhood label check and keeps the top $k\\%$ with highest boundary feature similarity (Eq. 13). The contrastive loss (Eq. 11) then pulls each selected embedding toward its class embedding and pushes it away from the other class embeddings, and the confidence-based dynamic filter (CDF) sets per-class pseudo-label thresholds through EMA-updated global and local confidence estimates (Eqs. 14–18). This selection-plus-contrast mechanism is what the ablation study credits for most of the reported improvement over the unlabeled-supervision baseline.","core_discovery":"On the paper's own terms, the central discovery is that explicitly selecting low-contrast pixels for contrastive learning yields a large and consistent gain in semi-supervised endoscopic segmentation. ICE computes a class embedding $\\bar{z}_c$ for each class by average-pooling labeled feature embeddings, then selects the top $k\\%$ of pixels with the lowest cosine similarity to their own class embedding; BCE selects boundary pixels with the highest minimum similarity to their $h$ neighbours. The low-contrast-enhanced contrastive loss (Eq. 11) pulls those selected embeddings toward the correct class embedding and pushes them away from the other class embeddings. CDF replaces a fixed pseudo-label threshold with an EMA-updated global threshold scaled by per-class local thresholds, increasing the number of pseudo-labels kept for minority classes. Together the components improve mIoU on the laryngeal cancer dataset by 2.55, 3.41, and 2.75 points over the previous best method at 10%, 30%, and 50% labels, and the benign-class IoU by 4.13, 5.29, and 3.58 points; the full model also reports the best mIoU on both polyp datasets.","pith_inferences":["The ICE selection rule implicitly assumes each class has a single compact prototype in embedding space; if a lesion class is visually heterogeneous, the lowest-similarity pixels may be legitimate subclusters rather than hard negatives, and pulling them to one prototype could distort the feature space. Testing class-conditional clustering before selection would clarify this.","Because CDF's thresholds are driven by teacher confidence, early-training confidences could let noisy pseudo-labels through for minority classes before the model stabilises; a confidence-calibration or warm-up schedule is a natural extension the paper does not explore.","The same hard-pixel selection idea should transfer to other low-contrast dense prediction tasks such as skin-lesion or surgical-scene segmentation, but this is an extrapolation: the paper only evaluates endoscopy.","If the gain comes from selection rather than contrastive loss per se, then replacing ICE's low-similarity pixels with randomly chosen pixels at the same $k\\%$ should erase most of the reported improvement, a comparison the paper does not run."],"forward_implications":["Under a 10% label partition on the laryngeal cancer dataset, LoCo reports a benign-tumor IoU of 52.96%, a 4.13-point gain over the previous best method, showing that minority-class pixels benefit the most from hard-pixel contrastive learning.","On Kvasir-SEG, LoCo reports mIoU gains over the second-best method of 2.57, 1.61, and 1.83 points at 10%, 30%, and 50% labels, indicating the benefit persists when labels are more plentiful.","The ablation study (M5 vs M2) attributes the largest single contribution to ICE rather than BCE, suggesting that inter-class hard-pixel mining is the more critical of the two selection rules on this data.","CDF raises pseudo-label utilization for the minority benign class relative to a fixed 0.95 threshold, implying that fewer rare-class labels are discarded during unlabeled training.","The authors position LoCo as applicable beyond laryngeal cancer to polyp segmentation, where the same framework reports the best mIoU on both public datasets."],"supporting_citations":[{"why":"Provides the FixMatch consistency-and-confidence baseline that LoCo extends and reports beating on all three datasets.","marker":"[13]"},{"why":"Provides the AugSeg strong-augmentation baseline that is the closest competitor on Kvasir-SEG and a key comparison point.","marker":"[10]"},{"why":"Provides the UniMatch weak-to-strong consistency baseline used in all comparison tables.","marker":"[11]"},{"why":"Provides the CorrMatch correlation-matching baseline against which LoCo is compared.","marker":"[12]"},{"why":"Supplies the U2PL unreliable-pseudo-label contrastive baseline, the prior method that also uses contrastive learning on all pixels.","marker":"[46]"},{"why":"Inspires the self-adaptive thresholding mechanism that CDF adapts with per-class local thresholds for minority classes.","marker":"[50]"},{"why":"Supplies the mean-teacher EMA update rule that forms the student-teacher backbone of LoCo.","marker":"[48]"},{"why":"Defines the polyp dataset split protocol used to evaluate LoCo on Kvasir-SEG and CVC-ClinicDB.","marker":"[17]"}],"fun_headline_variants":["Low-contrast pixels fuel semi-supervised endoscopy segmentation","Target low-contrast pixels to boost endoscopy AI","LoCo: contrastive learning for tricky endoscopy pixels","Sharpening low-contrast regions lifts endoscopy segmentation","Spotting hard pixels improves semi-supervised endoscopy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the pixels whose learned features sit farthest from their class's average feature are genuinely hard low-contrast pixels worth extra training, rather than noise or mislabeled examples—an assumption that becomes fragile when only 10% of images are labeled and those class averages come from very few images.","fun_headline_variants_meta":{"raw":{"variants":["Low-contrast pixels fuel semi-supervised endoscopy segmentation","Target low-contrast pixels to boost endoscopy AI","LoCo: contrastive learning for tricky endoscopy pixels","Sharpening low-contrast regions lifts endoscopy segmentation","Spotting hard pixels improves semi-supervised endoscopy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00059,"raw_usage":{"total_tokens":2827,"prompt_tokens":1060,"completion_tokens":1767,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":676,"completion_tokens_details":{"reasoning_tokens":1696}},"tokens_in":676,"tokens_out":1767,"duration_ms":12163,"temperature":1.0,"reasoning_tokens":1696,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:35:55.513136+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run LoCo on the laryngeal cancer dataset under a 10% label partition but replace the ICE pixel-selection rule with random pixel selection at the same $k\\%$, keeping BCE and CDF fixed; the paper's claim predicts a large drop in benign-class IoU relative to the full model, while a noise-sensitivity alternative predicts little change. A second check is to compare the selected low-similarity pixels against expert-annotated low-contrast regions: if the selection is doing its job, most selected pixels should fall inside or immediately beside regions experts mark as low contrast.","supporting_citations":[{"cited_title":"Revisiting weak- JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 10 to-strong consistency in semi-supervised semantic segmentation,","cited_arxiv_id":null,"evidence_quote":"Provides the UniMatch weak-to-strong consistency baseline used in all comparison tables."},{"cited_title":"Freematch: Self-adaptive thresholding for semi-supervised learning,","cited_arxiv_id":null,"evidence_quote":"Inspires the self-adaptive thresholding mechanism that CDF adapts with per-class local thresholds for minority classes."}],"review_version":1}