{"id":"fc63ac93-041c-4c96-999e-587077fe6ea1","arxiv_id":"1908.07669","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A weakly-supervised endoscopic lesion segmentation method that transfers semantic knowledge via class-balanced pseudo-labels, adversarial alignment, and feature centroids obtains higher mIoU than prior methods on a new dataset and on urban benchmarks.","lead":"This paper presents a model that outlines lesions in endoscopy images using only weak image-level labels, by transferring knowledge from a similar fully-labeled dataset. It matters because pixel-level medical labels are expensive, and the method also introduces a new 3,659-image endoscopic dataset.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The pseudo-label loop is never validated: Algorithm 1 thresholds uncalibrated softmax scores, and Algorithm 3 treats all selected labels as ground truth, so the main target-supervision signal may be dominated by confident errors.","rationale":"The reader's weakest_assumption is exactly the most load-bearing point: the entire target-side supervision chain depends on softmax confidence thresholds selecting predominantly correct pseudo-labels, and the paper never quantifies pseudo-label accuracy. I agree with this identification. The concern is not a formal contradiction but a correctness risk: the pseudo-label loss in Eq. (3), the centroid loss in Eq. (7), and the exponential centroid updates in Algorithm 3 all inherit whatever systematic errors the thresholded softmax outputs contain. If the initial model is miscalibrated on enteroscopy images, the method could reinforce confident mistakes rather than recover true lesion structure. The ablations show that removing pseudo-labels hurts performance, but they do not show that the pseudo-labels themselves are accurate, so the mechanism claimed in the abstract remains unverified. The absence of error bars and test-set hyperparameter tuning are secondary but real limitations; they do not change the verdict because the conditional acceptance already rests on adding such evidence. The paper does include useful independent support: ablation tables, parameter sweeps, a new dataset, and experiments on GTA/Cityscapes and SYNTHIA/Cityscapes all point in a consistent direction. Thus the reader's conditional verdict is appropriate, and the concern should be addressed by measuring pseudo-label quality directly rather than by rejecting the paper outright.","tokens_in":16957,"tokens_out":6681,"duration_ms":108269,"concrete_test":"Use the 390 held-out enteroscopy test images, which already have pixel ground truth, to measure pseudo-label quality without touching the training loop: after each training epoch, run Algorithm 1 and Algorithm 2 on those test images, compute per-class pixel IoU and precision of the generated pseudo-labels against ground truth, and track the error rate of newly admitted pixels as p increases from 25% to 55%. If pseudo-label precision for the lesion class is not substantially higher than the class prior, or if the precision of pixels newly admitted by raising p is no better than chance, then the pseudo-label loop is not a reliable source of target supervision and the reported PL gain cannot be attributed to correct semantic transfer. This evaluation should be performed only for diagnosis, not for model selection, to avoid contaminating the training procedure.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The method's only source of pixel-level target supervision is the pseudo-label generator. In Algorithm 1, each target pixel is assigned class k when its softmax score exceeds a per-class threshold e^{-lambda_k} (Eq. 5); lambda_k is read from the (1-p) quantile of the model's own confidence distribution, with p increasing from 25% to 55%. This is a confidence threshold, not an accuracy threshold. For a DeepLab-v3 trained on gastroscope images and then adapted to enteroscopy, softmax scores are uncalibrated under the target domain shift, so the selected top-p% pixels are not guaranteed to be mostly correct. The paper reports no pseudo-label error rate, no calibration analysis, and no study of how error changes as p grows. Because Algorithm 3 computes source/target centroids from these pseudo-labels and Eq. (3) trains S2 directly on them, any systematic confident-error pattern, such as lesions predicted only in their easy central region or normal tissue labeled as lesion, is reinforced by self-training. The class-balance and superpixel steps reduce spatial fragmentation and class imbalance, but they do not correct wrong high-confidence predictions. The ablation Ours-woPL shows that PL helps, but not whether the help comes from correct pseudo-labels or merely from adding target pixels to the segmentation loss; if the pseudo-labels were largely wrong, one would expect the opposite, so this is the load-bearing assumption of the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a weakly-supervised semantic segmentation method for endoscopic lesions. The method trains on a fully-labeled source domain (gastroscope images) and an image-level-labeled target domain (enteroscopy images), combining three main components: a pseudo-label generator with class-balanced confidence thresholds and superpixel spatial refinement; adversarial alignment of segmentation outputs across domains; and a semantic representation transfer loss that aligns exponentially weighted feature centroids for each class. The model is evaluated on a newly collected endoscopic dataset and on two standard synthetic-to-real benchmarks (SYNTHIA to Cityscapes, GTA to Cityscapes), with ablation studies for each component. The central claim is that the combination of these components improves weakly-supervised target-domain segmentation compared with prior domain adaptation and weakly supervised segmentation methods.","tokens_in":17209,"tokens_out":4344,"duration_ms":47351,"significance":"If the reported gains are reliable, the paper offers a practically useful direction: it reduces the need for pixel-level annotations in endoscopy by transferring knowledge from a related fully-labeled source domain. The release of a new endoscopic dataset is itself a valuable contribution, and the ablations across three transfer settings provide a systematic view of the method's components. However, the empirical support is currently weakened by the complete absence of error bars, significance tests, or repeated-run statistics, and by the fact that the pseudo-label generator—the sole source of pixel-level target supervision—is never validated for label accuracy or calibration. The central claim is plausible but is not yet established at the standard needed for a strong journal publication.","major_comments":[{"comment":"All reported results appear to come from a single training run: no error bars, standard deviations, or significance tests are provided anywhere. This matters because several key differences are small in absolute terms (e.g., Table 1: 63.96 vs. 61.42 mIoU; Table 2: Ours 63.96 vs. BL+PL+SRT 63.58; Table 3: Ours 42.9 vs. Ours-woSP 42.3). Without repeated runs and variance estimates, the claimed improvements over the nearest baselines cannot be distinguished from run-to-run noise. I request the authors to report means and standard deviations over at least three seeds, and, where possible, a paired significance test for the main comparisons.","section":"Section 4, Tables 1–4"},{"comment":"The pseudo-label generator is the only source of pixel-level supervision on the target domain, but its output quality is never measured. Algorithm 1 sets class thresholds at the (1-p) quantile of the model's softmax confidence, and Eq. (5) assigns a label to every pixel whose confidence exceeds the threshold. These are confidence thresholds, not accuracy thresholds; under domain shift, a model's softmax scores may be miscalibrated, and the selected top-p% pixels may contain a substantial share of confident errors. Because Eq. (3) trains the segmenter directly on these labels and Algorithm 3 computes target centroids from them, systematic errors (e.g., predicting only the easy central region of a lesion) are self-reinforced. The ablation Ours-woPL shows that adding pseudo-labels improves mIoU, but it does not distinguish correct pseudo-labels from merely adding target pixels to the segmentation loss. I request a pseudo-label error analysis: report per-class precision and recall of pseudo-labels against a small manually labeled target subset as a function of p, and include calibration curves for the target domain.","section":"Section 3.2, Algorithm 1, Eq. (5), Algorithm 3"},{"comment":"The exponentially weighted centroid update is written as C_k = sum_{x=1}^n C_x * gamma^{n-x} with gamma = 0.7. As written, this is an unnormalized cumulative sum, so the magnitude of the centroid grows with the iteration count n and the quantity is not a moving average; this could make the LSRT loss unstable as training proceeds. The authors should clarify whether the weights are normalized (e.g., divided by sum of gamma^{n-x}) and how centroids are initialized or handled for classes that do not appear in a given batch.","section":"Section 3.2, Algorithm 3, lines 9–10"},{"comment":"The dataset description states that 3659 images were collected from more than 1100 volunteers, but it does not state whether the 390 test enteroscopy images and the 300 training enteroscopy images come from disjoint volunteers. If images from the same patient appear in both training and test sets, the reported medical segmentation performance could be inflated by patient-level leakage. The authors should confirm that the train/test split is patient-disjoint and, ideally, report how many volunteers are used in each split.","section":"Section 4.1, Medical Endoscopic Dataset"}],"minor_comments":[{"comment":"Equation (3) uses a scalar weight lambda in the l1-norm term, while Eq. (4) and Algorithm 1 use class-dependent lambda_k; the relationship between the two formulations should be clarified.","section":"Section 3.2, Eqs. (3) and (4)"},{"comment":"The hyperparameter sensitivity plots in Figure 5 are difficult to read from the printed text; the authors should provide the numeric values or ensure the curves and axis labels are legible.","section":"Section 4.2, Figure 5"},{"comment":"There are several typos, including 'previously-leaned experience' in Section 3.2 and 'gasteroscope' in the caption of Figure 3; a careful proofread is needed.","section":"Throughout"},{"comment":"Reference [11] is cited as 'CVPR, 2019' for the ImageNet paper; the correct venue is CVPR 2009.","section":"References"},{"comment":"The text states that removing pseudo-label selection or semantic representation transfer decreases mIoU by 0.38% to 4.42%, but Table 2 shows several ablations within this range; please state explicitly which pairs the 0.38% and 4.42% refer to.","section":"Section 4.2, Ablation Study"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the journal's scope and the new dataset is a useful resource. The main risk is that the empirical claims may not be robust: there are no repeated runs, and the pseudo-label quality—the core mechanism for target supervision—is not validated. I recommend requesting the pseudo-label error analysis and statistical reporting described in the major comments before further consideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThis is a standard but credible empirical paper on weakly-supervised domain-adaptive segmentation for endoscopic images. The genuinely new pieces are the 3,659-image dataset (gastroscope/enteroscopy) and the specific assembly of class-balanced pseudo-labels with superpixel refinement, adversarial output alignment, and exponentially weighted centroid alignment. The ablations on the medical dataset and on GTA/SYNTHIA to Cityscapes consistently show the full model ahead of its parts and ahead of several recent baselines, so the authors are likely onto something real.\n\nThe soft spots are the usual ones for this area, plus one that is more specific. First, no error bars or repeated runs: every comparison is a single run, and the margins over the best baseline are 1–3 mIoU points, which is within the noise you’d expect from one seed. The hyperparameters (mu, eta, alpha, gamma, p schedule) are chosen empirically from the test set, with no validation split described. That weakens the comparison. Second, the stress-test concern is legitimate: Algorithm 1 picks the top-p% most confident pixels per class, which is a confidence threshold, not an accuracy threshold. If the model is miscalibrated under domain shift, the pseudo-labels can be confidently wrong, and Algorithm 3 then pulls the centroids in that wrong direction. The paper gives no pseudo-label error rate or calibration analysis. The ablation showing that PL helps suggests the pseudo-labels are mostly useful, but it doesn’t tell us whether the gain comes from correct labels or merely from adding target pixels to the loss. This is a real gap, though not a fatal one—the direction of improvement is consistent across datasets, which would be unlikely if the pseudo-labels were mostly garbage.\n\nThe dataset is a useful contribution, and the method is a sensible assembly of known parts. I’d like to see repeated runs with seeds, validation-based selection of hyperparameters, and some estimate of pseudo-label quality. The paper deserves a serious referee; it shouldn’t be desk-rejected. If the authors can provide those numbers, the paper becomes solid. As is, it’s conditionally acceptable.\n\nRecommendation: send to peer review, and ask for the missing statistics.","headline":"A credible assembly of known domain-adaptation pieces with a useful new endoscopic dataset; single-run results and an unquantified pseudo-label loop make it clearly conditional, but it deserves a real review.","tokens_in":17761,"tokens_out":1919,"would_cite":false,"duration_ms":30486,"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":"A semantic-transfer model with class-balanced pseudo-labels and feature-centroid alignment improves weakly-supervised endoscopic lesion segmentation, reaching 63.96 mIoU on a new 3659-image endoscopic dataset.","keywords":["weakly supervised learning","semantic segmentation","endoscopic lesion segmentation","pseudo-label generation","domain adaptation","feature centroid alignment","adversarial learning","superpixel priors"],"falsifier":"A direct test: run the pseudo-label generator on images whose lesion masks are known and measure pixel-level precision at the starting 25% threshold and at 55%; if precision is low, or if raising the selected fraction does not improve final mIoU, the self-training loop is not the reason for the reported gains.","tokens_in":16729,"feed_emoji":"🩺","tokens_out":8648,"duration_ms":506465,"temperature":0.7,"pith_summary":"A weakly supervised lesion segmenter that sees only image-level labels on the target endoscopic dataset can be made substantially more accurate by transferring knowledge from a fully labeled source dataset of related gastrointestinal images. The proposed model combines a class-balanced, superpixel-regularized pseudo-label generator with adversarial output alignment and per-class feature-centroid alignment built on exponentially weighted historical features. On the authors' new endoscopic dataset it reports 63.96 mean intersection-over-union (mIoU), versus 61.42 for the closest prior method, and it also improves synthetic-to-real street-scene segmentation, supporting the claim that the transfer mechanism is general rather than task-specific. If correct, the method would let hospitals build usable lesion segmenters from image tags alone, without expensive pixel-level annotation on each target domain.","feed_headline":"Confident pseudo-labels lift weakly-supervised lesion segmentation","feed_subtitle":"Borrowing fully labeled gastroscope knowledge, the model needs only image tags on target endoscopies.","key_machinery":"The load-bearing mechanism is a self-training loop built from three interacting components. First, a pseudo-label generator ranks all target pixels by predicted class probability, sets a per-class threshold $e^{-\\lambda_k}$ so that the top $p$ fraction of confident pixels are selected ($p$ growing from 25% to 55% over training), and then fills unlabeled pixels by voting among spatial neighbors that share the same superpixel region. Second, an adversarial discriminator on segmentation outputs pushes source and target output distributions together. Third, per-class feature centroids are computed from the pseudo-labeled target pixels and the ground-truth source pixels, exponentially averaged over past training iterations, and aligned between domains with an $\\ell^2$ plus $\\ell^1$ loss. The pseudo-labels do double duty: they supervise target segmentation and define the target centroids.","core_discovery":"The paper's central claim is that weakly supervised lesion segmentation can be substantially improved by transferring knowledge from a fully labeled source domain to an image-label-only target domain through a specific combination: class-balanced pseudo-labels refined by superpixel continuity, adversarial alignment of segmentation output distributions, and alignment of per-class feature centroids computed from exponentially weighted historical features. On the authors' collected endoscopic dataset, this reaches 63.96 mean intersection-over-union, against 61.42 for the closest comparison, and the same model reaches 47.2 on synthetic-to-real street-scene transfer versus 44.5 for the best prior method. Ablation results show that removing pseudo-labels, class balance, superpixel refinement, or semantic centroid transfer each lowers performance.","pith_inferences":["Editorial inference: because pixel-annotated source images are available, the pseudo-label generator could be validated directly on them; the paper does not report this, and that measurement would separate pseudo-label quality from the domain-alignment losses.","Editorial inference: a natural extension is to make per-class thresholds depend on class prevalence instead of a shared percentile schedule, which should improve rare-class recall if the class-balance assumption holds.","Editorial inference: an oracle experiment that replaces target pseudo-labels with human pixel labels on a small subset would quantify the ceiling of the self-training loop, something the paper leaves implicit."],"forward_implications":["Image-level labels plus a related fully labeled dataset can train a lesion segmenter that beats prior transfer methods; pixel-level annotation of the target domain is not required.","The same pipeline transfers from synthetic street scenes to real street scenes, so it is not tied to gastrointestinal endoscopy.","Expanding the selected confident-pixel fraction during training gradually adds harder target samples, so segmentation quality should improve as training progresses rather than plateauing early.","Class-balanced thresholds and superpixel voting specifically protect rare and spatially scattered classes, so the method's main gains should appear on hard-to-transfer categories."],"supporting_citations":[{"why":"provides the superpixel spatial prior used to fill in and regularize pseudo-labels","marker":"[1]"},{"why":"supplies the segmentation backbone and the baseline network against which all components are ablated","marker":"[4]"},{"why":"a cross-city adaptation method using soft pseudo-labels and class-wise statistics, used as comparison and motivation for class balance","marker":"[5]"},{"why":"a cycle-consistent adversarial adaptation baseline for synthetic-to-real segmentation transfer","marker":"[15]"},{"why":"a conditional GAN domain-adaptation baseline compared on both medical and street-scene transfers","marker":"[17]"},{"why":"contributes the self-paced-learning idea that iterative inclusion of confident pseudo-labeled samples rests on","marker":"[21]"},{"why":"the output-space adversarial adaptation approach this paper extends and the closest prior on the medical dataset","marker":"[33]"},{"why":"a curriculum adaptation baseline for semantic segmentation that motivates hard-sample inclusion","marker":"[42]"}],"fun_headline_variants":["Gastroscope knowledge transfer sharpens weakly supervised endoscopy segmentation","Transfer from gastroscope data lifts weakly labeled endoscopy lesion segmentation","Semantic transfer hits 63.96 mIoU for weakly supervised endoscopy segmentation","Gastroscope to endoscopy: knowledge transfer sharpens weak supervision","Weak supervision plus source data: better endoscopy lesion segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the pixels the model is most confident about are mostly correct, so that building pseudo-labels from the top percent of confident pixels strengthens rather than poisons training.","fun_headline_variants_meta":{"raw":{"variants":["Gastroscope knowledge transfer sharpens weakly supervised endoscopy segmentation","Transfer from gastroscope data lifts weakly labeled endoscopy lesion segmentation","Semantic transfer hits 63.96 mIoU for weakly supervised endoscopy segmentation","Gastroscope to endoscopy: knowledge transfer sharpens weak supervision","Weak supervision plus source data: better endoscopy lesion segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000675,"raw_usage":{"total_tokens":3070,"prompt_tokens":940,"completion_tokens":2130,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":2036}},"tokens_in":556,"tokens_out":2130,"duration_ms":173549,"temperature":1.0,"reasoning_tokens":2036,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:00:19.541489+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test: run the pseudo-label generator on images whose lesion masks are known and measure pixel-level precision at the starting 25% threshold and at 55%; if precision is low, or if raising the selected fraction does not improve final mIoU, the self-training loop is not the reason for the reported gains.","supporting_citations":[{"cited_title":"Slic superpix- els compared to state-of-the-art superpixel methods","cited_arxiv_id":null,"evidence_quote":"provides the superpixel spatial prior used to fill in and regularize pseudo-labels"},{"cited_title":"Encoder-decoder with atrous separable convolution for semantic image segmentation","cited_arxiv_id":null,"evidence_quote":"supplies the segmentation backbone and the baseline network against which all components are ablated"},{"cited_title":"No More Discrimina- tion: Cross City Adaptation of Road Scene Segmenters","cited_arxiv_id":null,"evidence_quote":"a cross-city adaptation method using soft pseudo-labels and class-wise statistics, used as comparison and motivation for class balance"},{"cited_title":"CyCADA: Cycle-consistent adversarial domain adapta- tion","cited_arxiv_id":null,"evidence_quote":"a cycle-consistent adversarial adaptation baseline for synthetic-to-real segmentation transfer"},{"cited_title":"Conditional generative adversarial network for struc- tured domain adaptation","cited_arxiv_id":null,"evidence_quote":"a conditional GAN domain-adaptation baseline compared on both medical and street-scene transfers"},{"cited_title":"Self- paced learning for latent variable models","cited_arxiv_id":null,"evidence_quote":"contributes the self-paced-learning idea that iterative inclusion of confident pseudo-labeled samples rests on"},{"cited_title":"Learning to adapt structured output space for semantic seg- mentation","cited_arxiv_id":null,"evidence_quote":"the output-space adversarial adaptation approach this paper extends and the closest prior on the medical dataset"},{"cited_title":"Curricu- lum domain adaptation for semantic segmentation of urban scenes","cited_arxiv_id":null,"evidence_quote":"a curriculum adaptation baseline for semantic segmentation that motivates hard-sample inclusion"}],"review_version":1}