{"id":"1b7ac425-d8e9-4021-a2e3-261eba3c1902","arxiv_id":"2411.14250","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"CP-UNet adds contour probability modeling and a channel-shift downsampling module to UNet, reporting small and inconsistent metric differences of about 0.1-0.2 percentage points versus TransUNet on three ultrasound datasets.","lead":"This paper presents CP-UNet, a deep learning model for outlining lesions in ultrasound images, adding a new downsampling module and a probabilistic contour model. Its reported accuracy is close to existing methods, with small and mixed differences, on three breast and thyroid ultrasound datasets.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The state-of-the-art claim rests on single-run metric differences of at most ~0.002 in Dice, with no error bars, multiple seeds, or significance tests; this is the load-bearing weakness.","rationale":"The reader identified the same load-bearing concern I would: the empirical superiority claim is supported only by single-run point estimates whose differences are negligible compared with expected training noise. My pass through the manuscript found no independent evidence—no seeds, no error bars, no code, no released private dataset—that would convert these differences into a reliable effect. The architectural proposal (MgCSD, CPM, GF) is coherent as an incremental idea, but the paper's acceptance hinges on the claim of outperforming SOTA, and that claim is not secured by the reported numbers. I therefore agree with the REJECT verdict. The proposed concrete test is deliberately minimal: repeat the comparison with multiple seeds and paired statistics on the two public datasets. If the result survives, the claim could be revisited. No ad hominem is intended; the issue is entirely with the evidential basis of the headline empirical claim.","tokens_in":8165,"tokens_out":2133,"duration_ms":22457,"concrete_test":"Run CP-UNet and pretrained TransUNet on BUSI and DDTI with at least 5 random seeds, identical data splits, preprocessing, and evaluation code. Report mean ± std of per-image IoU and Dice, plus paired bootstrap confidence intervals or a paired significance test (e.g., Wilcoxon signed-rank on per-image Dice). If the 95% CI of the paired difference includes zero, or if CP-UNet's mean Dice is not higher than TransUNet's on both public datasets while mean IoU is also not lower, the claim of consistent state-of-the-art improvement is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—'consistently improves... outperforming the strong baseline and state-of-the-art methods'—is supported only by Table I, which reports one run per method. Against the strongest competitor, pretrained TransUNet, CP-UNet is lower in IoU on BUSI (0.6445 vs 0.6454) and DDTI (0.8136 vs 0.8147), and higher in Dice by only 0.0015 on BUSI (0.7827 vs 0.7812) and 0.0009 on DDTI (0.8972 vs 0.8963). On TUI the gains are similarly tiny: +0.0015 IoU, +0.0017 Dice. These differences are well within typical run-to-run stochastic variation for deep segmentation models. The paper provides no standard deviations, no multiple seeds, no statistical tests, and no code, so there is no way to tell whether the reported ordering is systematic or noise. The ablation study (Table II) has the same problem: the full model beats the next-best configuration by 0.0122 IoU, but again with one run and no variance estimate. The private TUI dataset is also not described or released, which further weakens verifiability. The conclusion 'outperforms state-of-the-art' therefore cannot be accepted on the evidence presented; the weakest load-bearing assumption is that the reported single-run metric differences are real and reproducible.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CP-UNet, a U-Net variant for medical ultrasound image segmentation. Three components are introduced: a multi-group channel-shift downsampling module (MgCSD) that creates global-local connections, a contour probabilistic modeling module (CPM) that uses a Gaussian mixture model to represent contour feature distributions and aligns them to ground-truth mask features via a KL divergence, and a gating-based feature filtering module (GF) that fuses contour features in the decoder. The model is trained with a combination of BCE, Dice, and KL losses. Experiments are reported on the public BUSI and DDTI datasets and a private TUI dataset, comparing against a range of U-Net, transformer, and attention-based baselines, with the claim that CP-UNet consistently outperforms state-of-the-art methods.","tokens_in":8409,"tokens_out":8233,"duration_ms":73572,"significance":"If the claimed improvements were properly supported, CP-UNet would be a useful contribution to ultrasound segmentation, specifically the idea of guiding contour representation through distributional alignment rather than a single contour loss. The paper retrains all baselines on the same datasets, which is good practice. However, the absence of error bars, the small and partially inconsistent differences with the pretrained TransUNet baseline, the uncharacterized private dataset, and the underspecified GMM formulation currently prevent the central claim from being assessed. The paper also provides no code, further limiting reproducibility.","major_comments":[{"comment":"The central claim that CP-UNet 'consistently improves' segmentation and outperforms state-of-the-art methods is not supported by the reported numbers. On BUSI, CP-UNet's IoU is 0.6445 versus 0.6454 for pretrained TransUNet, and on DDTI the IoU is 0.8136 versus 0.8147; only on TUI is the IoU higher (0.8496 versus 0.8481). The Dice differences are all below 0.002. Because Table I reports only a single run per method and provides no standard deviations, confidence intervals, or significance tests, these differences are statistically indistinguishable from training variability. This is a load-bearing weakness for the paper's main claim, and the authors must provide multiple-seed results with error bars and a statistical comparison, or substantially revise the claim.","section":"§III-B, Table I"},{"comment":"The TUI dataset, on which the largest relative improvements are reported, is described only as a private thyroid ultrasound dataset with 15,233 images. No information is provided about image acquisition, annotation protocol, patient demographics, inclusion/exclusion criteria, or the train/validation/test split, and the dataset is not released. This makes the TUI results impossible to verify or reproduce, and the reported gains may be specific to an uncharacterized data distribution. Please provide a detailed dataset description and release the data or at least the evaluation split.","section":"§III-A, private TUI dataset"},{"comment":"The Gaussian mixture modeling is not implemented as a Gaussian mixture. Equation (3) samples a single z from N(0,1) and forms G = Ω(zσ_A + μ_A), which is a deterministic linear combination of Gaussian variables; a true mixture sample would first select a component index according to component weights and then sample from that component's Gaussian. The KL divergence in Eq. (4) is given for a single Gaussian pair, but the text says μ_A,σ_A and μ_B,σ_B contain K sets of Gaussians; no summation over K is specified. As a result, the training loss in Eq. (7) is underspecified, and the claim that the model 'samples from the mixed Gaussian distribution' is inaccurate. Please clarify the actual sampling and loss computation.","section":"§II-B, Eq. (3)"},{"comment":"The paper does not state how the datasets are split into training/validation/test sets, how many random seeds are used, or whether data augmentation and preprocessing are applied. It is also unclear whether the CPM sampling in Eq. (3) is performed at test time with a fixed z or averaged over multiple samples. The ablation study in Table II is based on a single run; the full model's improvement over MgCSD+CPM is only 0.0025 IoU (0.6445 versus 0.6420), which is within the range of random variation expected from the issues in Major Comment 1. These details are essential for any of the reported comparisons to be interpretable.","section":"§III-A, experimental protocol"}],"minor_comments":[{"comment":"The caption refers to 'CMP, Contour Probabilistic Modeling' but the module is consistently called CPM in the text; please unify the abbreviation.","section":"Fig. 2 caption"},{"comment":"The sentence '⊕ and ⊗ are element-wise operation' should say 'element-wise operations'.","section":"§II-A, Eq. (1)"},{"comment":"The main text refers to 'Table III-C' but the table is labeled 'TABLE II'; correct the cross-reference.","section":"§III-C, Table II"},{"comment":"The symbol L is used both for the number of encoder layers and for the globally enhanced feature F_L in Eq. (1); this is confusing and should be disambiguated.","section":"§II-B"},{"comment":"The term 'adaptive homoscedastic resampling' is used for Eq. (3) but not defined; please explain what is adaptive and what is homoscedastic about the resampling.","section":"§II-B"}],"recommendation":"major_revision","confidential_remarks":"The main empirical claim is currently unsupported because the reported single-run differences are tiny and partly contradict the claim. The method section also leaves the Gaussian mixture formulation underspecified. These issues are severe but could be addressed by a careful rerun with multiple seeds, a corrected and clarified model description, and either a release of the private dataset or removal of its results. I recommend that acceptance be contingent on these revisions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Punchline: the architecture is a reasonable incremental contribution, but the empirical case for it is built on single-run differences of a few tenths of a percent, and the paper's own table shows CP-UNet does not beat pretrained TransUNet on IoU on two of three datasets. The right response is not panic, but: the method is plausible, the evaluation is not.\n\nWhat's new: CP-UNet combines three modules: a channel-shift downsampling (MgCSD), a Gaussian-mixture contour modeling module with KL alignment to the mask's contour features, and a gating fusion module. The individual ideas are familiar—channel attention, spatial shift, reparameterized Gaussian sampling, gating—but the specific combination for contour-aware ultrasound segmentation is not something I've seen. The ablation is honest: MgCSD alone gives the big jump (~10 IoU points), CPM alone gives very little, and the full model is best. That pattern is internally consistent and suggests the authors understand their own modules.\n\nThe soft spots: the central claim 'outperforming state-of-the-art' is not supported by Table I. Against pretrained TransUNet, CP-UNet is lower in IoU on BUSI and DDTI, and the Dice gains are 0.0015 and 0.0009. On TUI the gains are 0.0015 IoU and 0.0017 Dice. These are within run-to-run noise for deep segmentation; without multiple seeds, error bars, or significance tests, the ordering is not established. The private TUI dataset is not described or released, so those numbers are unverifiable. The probabilistic contour model is also under-specified: the KL term is written for a single Gaussian but applied to K components, and the adaptive resampling weight Ω is described loosely. Minor issues: the Fig. 2 caption has a typo ('CMP'), and the paper says 'homoscedastic' in one place and 'homologous' in another.\n\nWho it's for: someone working on ultrasound lesion segmentation who wants to try a contour-distribution prior. A serious referee should see it, mostly to push for proper statistical evaluation and a released/described private dataset. My own verdict would be reject as-is, but the idea is not dead; with multi-seed runs and error bars, or a more careful claim, it could be a decent workshop or short paper.","headline":"A plausible incremental architecture for contour-aware ultrasound segmentation, but the state-of-the-art claim is built on single-run differences within noise.","tokens_in":8988,"tokens_out":1784,"would_cite":false,"duration_ms":16810,"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":"Contour-aware CP-UNet sharpens ultrasound lesion segmentation","keywords":["medical ultrasound segmentation","contour probabilistic modeling","Gaussian mixture model","downsampling","breast nodule","thyroid nodule","U-Net","boundary uncertainty"],"falsifier":"Run CP-UNet and the strongest baseline (pretrained TransUNet) at least five times each with different random seeds on the same BUSI, DDTI, and TUI splits, compute Dice and IoU confidence intervals, and check whether the intervals separate; if they overlap, the central claim does not hold.","tokens_in":7912,"feed_emoji":"🩺","tokens_out":3303,"duration_ms":29183,"temperature":0.7,"pith_summary":"The paper claims that explicitly modeling lesion boundaries as a probability distribution improves segmentation of blurred ultrasound contours. It proposes CP-UNet, which adds a contour probabilistic modeling module to a U-Net-style encoder-decoder, and reports higher Dice and IoU than strong convolutional, transformer, and attention baselines on three breast and thyroid ultrasound datasets. A sympathetic reader would care because blurred boundaries are a common obstacle in ultrasound-based diagnosis, and the method is designed to capture boundary uncertainty rather than ignore it.","feed_headline":"Contour-aware CP-UNet sharpens ultrasound lesion segmentation","feed_subtitle":"Modeling blurry boundaries as a Gaussian mixture beats strong baselines on breast and thyroid ultrasound scans.","key_machinery":"Three modules carry the argument. Multi-group Channel Shift Downsampling (MgCSD) splits features into channel groups, cyclically shifts most groups spatially, and re-fuses them with a global-pooling-weighted branch to give the encoder both local detail and global context. Contour Probabilistic Modeling (CPM) maps the downsampled features to K Gaussian mean-variance pairs, samples them via the reparameterization trick, and weights the samples with adaptive learnable parameters to approximate sampling from a Gaussian mixture. A gating-based feature filtering module (GF) then fuses upsampled, skip, and contour-sampled features at each decoding stage.","core_discovery":"CP-UNet guides the decoding stage with contour features sampled from a Gaussian mixture distribution fitted to the encoder's global-local features. The paper reports IoU/Dice of 0.6445/0.7827 on BUSI, 0.8136/0.8972 on DDTI, and 0.8496/0.9183 on TUI, with the largest margins over convolution baselines and results roughly comparable to a pretrained TransUNet while better following irregular contour zigzags. The authors state that their approach consistently improves segmentation accuracy for breast and thyroid nodules, outperforming strong baselines and state-of-the-art medical image segmentation methods.","pith_inferences":["The ablation results suggest most of the gain comes from MgCSD rather than the probabilistic contour modeling, since adding MgCSD alone raises IoU from 0.5439 to 0.6410 while CPM alone adds little; a tighter test would compare MgCSD-only variants against the full model on multiple seeds.","The claimed advantage over pretrained TransUNet is small in table metrics and rests on visual contour quality; a reader could test whether the reported boundary improvements translate to clinically meaningful contour errors.","A natural extension is to output the mixture variance as an uncertainty map for each predicted boundary and measure whether radiologists trust low-variance boundaries more under reading-time pressure."],"forward_implications":["If validated, CP-UNet offers a concrete way to reduce contour blurring in ultrasound segmentation tasks, improving boundary accuracy for lesion measurement and follow-up.","The probabilistic contour representation could provide not just a mask but a per-boundary uncertainty estimate, useful in clinical review workflows.","MgCSD is a drop-in replacement for standard downsampling and may improve other U-shaped segmentation networks on imaging modalities with noise artifacts.","The method's smaller-data gains on BUSI suggest the contour prior helps most when training data are limited."],"supporting_citations":[{"why":"Supplies the U-Net encoder-decoder baseline that CP-UNet extends and compares against.","marker":"[1]"},{"why":"Provides the BUSI breast ultrasound dataset used for training and evaluation.","marker":"[13]"},{"why":"Provides the DDTI thyroid ultrasound dataset used for training and evaluation.","marker":"[14]"},{"why":"Serves as an attention-based baseline and supplies the channel self-attention design that MgCSD adapts.","marker":"[15]"},{"why":"Provides the reparameterization trick that lets CPM sample from the learned Gaussian distributions while preserving gradients.","marker":"[16]"},{"why":"TransUNet is the strongest compared baseline, and its pretrained variant is the closest competitor in the reported metrics.","marker":"[19]"}],"fun_headline_variants":["CP-UNet: contour probability model sharpens ultrasound segmentation","Gaussian mixture contours improve ultrasound lesion segmentation","Contour-aware probabilistic model beats baselines on ultrasound","CP-UNet models uncertain contours via Gaussian mixture","Probability-based contour guidance improves thyroid and breast scans"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported metric differences come from a single training run per method, with no standard deviations or statistical tests, so the claim of consistent improvement assumes those small differences are not random training noise.","fun_headline_variants_meta":{"raw":{"variants":["CP-UNet: contour probability model sharpens ultrasound segmentation","Gaussian mixture contours improve ultrasound lesion segmentation","Contour-aware probabilistic model beats baselines on ultrasound","CP-UNet models uncertain contours via Gaussian mixture","Probability-based contour guidance improves thyroid and breast scans"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001057,"raw_usage":{"total_tokens":4367,"prompt_tokens":805,"completion_tokens":3562,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":421,"completion_tokens_details":{"reasoning_tokens":3487}},"tokens_in":421,"tokens_out":3562,"duration_ms":23280,"temperature":1.0,"reasoning_tokens":3487,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:23:04.767424+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run CP-UNet and the strongest baseline (pretrained TransUNet) at least five times each with different random seeds on the same BUSI, DDTI, and TUI splits, compute Dice and IoU confidence intervals, and check whether the intervals separate; if they overlap, the central claim does not hold.","supporting_citations":[{"cited_title":"Dataset of breast ultrasound images,","cited_arxiv_id":null,"evidence_quote":"Provides the BUSI breast ultrasound dataset used for training and evaluation."}],"review_version":1}