{"id":"da0eab9b-bff6-479b-9bf0-db0dfd24ad55","arxiv_id":"2412.06530","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"HES-UNet, combining Haar-wavelet downsampling, attention, and multi-scale fusion, reports a Dice score of 89.21% for hepatic echinococcosis lesion segmentation on a private 268-patient CT dataset.","lead":"HES-UNet, a new U-Net-style neural network, reports the best scores on a private CT dataset for hepatic echinococcosis lesion segmentation. The paper matters because it targets a neglected disease in low-resource pastoral areas, but its results rest on a dataset and code that are not public.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim collapses under an internal inconsistency: Table I reports HES-UNet at 89.21% DSC while Table II reports the same architecture at 88.13%, nearly erasing the claimed 1.09% margin over TransUNet.","rationale":"The reader's verdict is CONDITIONAL, and I agree with that verdict, but the reason should be sharpened. The reader's weakest assumption was slice-level data splitting, which is a legitimate external concern about generalization. However, the paper contains a more direct, internal threat to the headline claim: the same full model is reported with two different DSC values in the two main tables. The claimed lead over TransUNet is 1.09%, and the discrepancy between the two reported HES-UNet scores is 1.08%, so the SOTA claim hinges entirely on accepting the Table I number. The ablation baseline row matches Table I's U-Net exactly, which suggests the same evaluation protocol was used, making the inconsistency harder to dismiss as a mere protocol difference. If Table II is correct, the main result is not SOTA; if Table I is correct, the ablation table is wrong. Either way, the paper as written does not support the headline. A public release of code/data or an explicit reconciliation would resolve this. No issue with the architecture itself was found; the concern is about the empirical claim, so a conditional verdict with this specific condition is appropriate.","tokens_in":8990,"tokens_out":3297,"duration_ms":32554,"concrete_test":"Obtain from the authors the exact split indices, preprocessing, and training configuration used for Table I and Table II. Then rerun the full HES-UNet with the Table I configuration and compute DSC on the identical test set. If the result reproduces 88.13% (Table II) rather than 89.21% (Table I), the claimed 1.09% SOTA margin over TransUNet is an artifact. If the authors respond that the two tables used different test partitions or different model variants, request that both tables be recomputed on a single shared, patient-disjoint split and report the reconciled value.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing problem is internal to the paper. In Table I, HES-UNet is reported at DSC 89.21%, precision 88.14%, and recall 89.60%, and this value is used in the abstract and conclusion to claim a 1.09% improvement over TransUNet (88.12%). In Table II, the row with all three modules enabled (MDB ✓, MUB ✓, MAB ✓) represents the same full HES-UNet architecture, yet it is reported at DSC 88.13%, precision 87.84%, and recall 88.37%. The difference between the two reported values for the same model is 1.08 percentage points, almost exactly the claimed margin over TransUNet. The paper offers no explanation for this discrepancy. If the Table II value is the correct one, HES-UNet is only 0.01% above TransUNet, and the central SOTA claim collapses. This is not an external assumption like leakage; it is a visible contradiction between two tables that purport to evaluate the same model on the same dataset. Moreover, the baseline U-Net row in Table II (86.58% DSC) exactly matches the U-Net entry in Table I, suggesting the ablation was run under the same evaluation protocol; this makes the full-model mismatch even harder to attribute to a benign protocol difference. The claimed SOTA advantage therefore rests entirely on accepting the Table I number, and the paper currently provides no way to know which number is trustworthy.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes HES-UNet, a U-Net-style architecture for segmenting hepatic echinococcosis (HE) lesions in CT images. The architecture combines convolutional and attention components, introducing three modules: a multi-directional downsampling block (MDB) based on Haar wavelets, a multi-scale aggregation block (MAB), and a multi-scale upsampling block (MUB) that feeds a global feature filtering module and a progressive fusion decoder with deep supervision. The authors collected a private CT dataset from 268 HE patients (137 CE, 131 AE), performed a slice-level 80/10/10 train/validation/test split, and report that HES-UNet achieves a DSC of 89.21%, claiming a 1.09% improvement over TransUNet (88.12%). They include comparisons with U-Net, U-Net++, TransUNet, Swin-UNet, Res-UNet, and EGE-UNet, plus an ablation study over the three proposed modules.","tokens_in":9266,"tokens_out":4311,"duration_ms":43749,"significance":"Hepatic echinococcosis segmentation is a clinically relevant problem, especially in resource-limited pastoral regions, and a well-performing automated method would be of practical value. The paper proposes a reasonably motivated architecture that combines wavelet-based downsampling, multi-scale aggregation, global attention filtering, and deep supervision; if the reported results were credible, the work would provide a useful reference for applying such hybrid CNN-attention designs to HE lesions. The paper also supplies a project page and describes hyperparameters and the training protocol, which aids reproducibility. However, the central performance claim is currently not substantiated because of an internal inconsistency between the two main result tables, and the evaluation protocol has design issues (slice-level split, no statistical tests) that affect the validity of the claimed state-of-the-art margin.","major_comments":[{"comment":"The headline claim is undermined by an internal inconsistency between the two result tables. Table I reports HES-UNet at DSC 89.21%, precision 88.14%, recall 89.60%, and the 1.09% margin over TransUNet is derived from these numbers. Table II, which ablates the proposed modules on the same dataset, reports the full model (MDB ✓, MUB ✓, MAB ✓) at DSC 88.13%, precision 87.84%, recall 88.37%. The difference of 1.08 percentage points is almost exactly the claimed margin over TransUNet. Notably, the U-Net baseline row is identical in both tables (86.58 DSC), suggesting the same evaluation protocol was used. The authors provide no explanation for which number is correct. If Table II is correct, the claimed state-of-the-art advantage collapses to a 0.01% difference over TransUNet. This is a load-bearing discrepancy that must be resolved before the central claim can be accepted; the authors should report the source of each number, the protocol used, and ideally per-run or per-patient variation.","section":"Abstract / Section V / Tables I and II"},{"comment":"The data are split at the slice level, not the patient level: 'we saved the CT slice data, using 80% for the training set and 10% each for the testing and validation sets.' In a multi-slice CT dataset, slices from the same patient will appear in both training and test sets. This introduces potential patient-specific leakage: the model may memorize appearance or lesion characteristics that are shared across slices of the same patient, inflating the reported DSC and making the comparison measure intra-patient slice prediction rather than generalization to new patients, which is the clinically relevant task. The authors should perform a patient-level split (or otherwise ensure that no patient contributes slices to both training and test), and they should report the number of patients and slices in each split.","section":"Section IV.A (Dataset and Implementation Details)"},{"comment":"No error bars, standard deviations, or statistical significance tests are reported for any metric. The claimed improvements over baselines are small (1.09% over TransUNet, 2.63% over U-Net, 2.45% over U-Net++), and with a single evaluation run on a private dataset, these differences could easily be within run-to-run or patient-subset variation. At minimum, the authors should report the mean and standard deviation over multiple training seeds and a paired statistical test (e.g., Wilcoxon signed-rank at the patient level, or a bootstrap over patients) to establish that the differences are not due to chance.","section":"Tables I and II / Section IV.B"}],"minor_comments":[{"comment":"The sentence reporting improvements is inaccurate: 'HES-UNet achieved improvements of 2.63%, 1.56%, and 1.09% in the DSC metric compared to CNN-based segmentation methods (U-Net and U-Net++), the Transformer-based Swin-UNet, and the hybrid architecture TransUNet, respectively.' The U-Net++ difference is 2.45%, not 1.56%; the three listed percentages correspond to U-Net, Swin-UNet, and TransUNet, so the model list and values do not align.","section":"Section IV.B"},{"comment":"The caption contains a typo: 'THE COMPARISON RESULTS BETWEEN OUR HES-UN ET AND OTHER MODELS' has an unwanted space in 'HES-UNET' and inconsistent capitalization.","section":"Table I caption"},{"comment":"There is a formatting issue: 'H = 512and W = 512represent' should be 'H = 512 and W = 512 represent'.","section":"Section III.A"},{"comment":"The manuscript states that patient-identifying metadata were removed but does not mention ethics approval or institutional review board clearance for the collection and use of patient CT data. For a medical-imaging study, an explicit statement about ethics approval (or a justification for its absence) is expected.","section":"Section IV.A"},{"comment":"The MDB is described in the contributions as 'lossless feature downsampling,' but the module includes a 1x1 convolution after concatenation of wavelet sub-bands, which is not invertible in general. The claim is overstated; consider using a term such as 'detail-preserving' or 'information-preserving.'","section":"Section III.B"},{"comment":"The design of MUB and GAM is inspired by two self-citations ([19], [20]) from prior speech-separation work. The authors should state more explicitly how these modules are adapted from the cited works and how the adaptation differs, to help readers assess novelty.","section":"References [19], [20]"},{"comment":"The deep supervision loss weights λ_i are set to {0.1, 0.2, 0.3, 0.4, 0.5, 1} without any sensitivity analysis or description of how these values were chosen. A brief justification or a sensitivity check would strengthen the presentation, although this is not a blocking issue.","section":"Section III.E"}],"recommendation":"major_revision","confidential_remarks":"The internal inconsistency between Table I and Table II is severe: it directly affects the main claim of state-of-the-art performance and must be fixed before the paper can be considered. The slice-level split and lack of statistical testing compound the concern, because the reported margins are small. I would advise the editor to require a revised version with a patient-level evaluation, per-run variation, a clear reconciliation of the two tables, and a statement of ethics approval. The architectural contribution itself is plausible and the clinical problem is important, so the appropriate path is major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nQuick take: this is a U-Net variant for hepatic echinococcosis segmentation. The architecture is a plausible recombination of published components – Haar wavelet downsampling, GHPA attention, CBAM – and the authors have collected a private CT dataset from 268 patients, which is real work. The writing is clear and the ablation is thoughtful.\n\nBut the headline claim doesn't survive contact with the paper's own tables. Table I reports HES-UNet at 89.21% DSC. Table II, the row with all three modules enabled, reports the same architecture at 88.13% DSC. That's a 1.08-point gap, almost exactly the 1.09-point margin they claim over TransUNet. If the Table II number is the right one, they're 0.01% above TransUNet and the 'state-of-the-art' claim collapses. The paper offers no explanation, and the baseline U-Net row matches across both tables, so it's hard to blame a protocol change.\n\nOther issues: the data split is at slice level, not patient level, so slices from the same patient may appear in both training and test sets. That can inflate DSC via memorization. There are no error bars or significance tests, and no code or data released. The novelty is moderate – each module is borrowed, and the contribution is the assembly and the application. That's legitimate, but it doesn't change the fact that the empirical evidence is the weakest link.\n\nAll that said, this is not a silly paper. The problem is real, the method is described in enough detail to reproduce, and the application to a neglected disease in low-resource settings is worth pursuing. The internal inconsistency is the sort of thing a careful referee could catch and force the authors to fix. If they can reconcile the two numbers and clarify the split, the paper would be much stronger.\n\nI'd send it to peer review, but with instructions that the discrepancy between Table I and Table II must be resolved before acceptance. Without that, the central claim is unverified. For a reader in the hepatic echinococcosis imaging niche, this is worth a look; for anyone else, it's a reminder that ablation tables should agree with main tables.\n\nRecommendation: engage with it, but demand the reconciliation.","headline":"A plausible U-Net variant for hepatic echinococcosis segmentation, but the headline SOTA claim is contradicted by the paper's own ablation table.","tokens_in":9844,"tokens_out":2742,"would_cite":false,"duration_ms":27465,"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":"HES-UNet, a U-Net variant with multi-directional downsampling and multi-scale aggregation, segments hepatic echinococcosis lesions in CT scans with a Dice Similarity Coefficient of 89.21%, outperforming six baselines including TransUNet…","keywords":["hepatic echinococcosis","medical image segmentation","U-Net","computed tomography","multi-scale feature fusion","attention mechanism","deep supervision","Dice similarity coefficient"],"falsifier":"A concrete check is to re-split the dataset by patient (e.g., 80% of the 268 patients for training, 10% validation, 10% test) and rerun the identical HES-UNet training protocol; if the DSC drops well below 89.21%, the reported state-of-the-art result is an artifact of slice-level leakage rather than a property of the model.","tokens_in":8762,"feed_emoji":"🩻","tokens_out":5643,"duration_ms":48586,"temperature":0.7,"pith_summary":"This paper aims to establish that a U-Net variant called HES-UNet, built around three new feature-fusion blocks, can segment hepatic echinococcosis lesions in abdominal CT scans more accurately than six existing segmentation models. On a private dataset of CT slices from 268 patients, HES-UNet reports a Dice Similarity Coefficient of 89.21%, outperforming the strongest baseline, TransUNet, by 1.09 percentage points. The authors argue that the gain comes from preserving high- and low-frequency detail during downsampling, aggregating features across all encoder levels, and letting each decoder stage receive multi-scale global features. If the result holds, it would give a practical tool for automated lesion delineation in regions where echinococcosis is common but specialized medical staff are scarce.","feed_headline":"HES-UNet beats TransUNet on liver parasite lesion segmentation","feed_subtitle":"A multi-scale U-Net with wavelet downsampling reaches 89.21% Dice on liver parasite CT scans, beating six baselines.","key_machinery":"The load-bearing mechanism is a set of three feature-fusion blocks grafted onto the U-Net skeleton. The multi-directional downsampling block (MDB) applies a Haar wavelet transform to split each feature map into low-frequency and three directional high-frequency parts, then concatenates and compresses them with a 1x1 convolution, so downsampling does not discard boundary detail. The multi-scale aggregation block (MAB) uses adaptive average pooling to resize all encoder outputs to one resolution, adds them, and passes the deepest scale through a GHPA attention module, producing a single global feature tensor. The multi-scale upsampling block (MUB) expands that tensor with group-wise attention and pixel shuffle, generating multi-scale global features that are fused with encoder features and intermediate predictions inside five global attention modules (GAMs). These modules, plus deep supervision, carry the claimed accuracy gain.","core_discovery":"The central claim is that HES-UNet, a symmetric encoder-decoder with 32 to 512 channels, achieves state-of-the-art segmentation of hepatic echinococcosis lesions on a collected CT dataset, reaching a DSC of 89.21%, precision of 88.14%, and recall of 89.60%. These numbers beat U-Net, U-Net++, TransUNet, Swin-UNet, Res-UNet, and EGE-UNet on the same test slices. The paper attributes the improvement to three components added to the U-Net backbone: a multi-directional downsampling block using Haar wavelet decomposition to keep detail, a multi-scale aggregation block that pools encoder features of all five scales into a global representation, and a multi-scale upsampling block with pixel shuffle that feeds multi-scale global features into skip connections. A deep supervision module computes Dice plus binary cross-entropy losses at five decoder stages. The paper concludes that the architecture is an effective auxiliary diagnostic tool for echinococcosis in resource-poor areas.","pith_inferences":["Because the dataset split is by slice rather than by patient, slices from the same patient likely appear in both training and test sets; the reported 89.21% DSC is therefore probably optimistic for unseen-patient generalization, and the comparison against baselines may be biased in the model's favor if it memorizes patient-specific appearance.","The 1.09-point lead over TransUNet is small relative to typical slice-level variance, so the practical significance of the improvement would need confirmation on a patient-level split or an external dataset.","The proposed modules are architecture-agnostic, so one could test them on public datasets such as liver or spleen segmentation to see whether the gain generalizes beyond this private echinococcosis collection.","If the authors release slice-level metadata with patient identifiers, independent groups could quantify the leakage and re-run the comparison with a grouped split; without that, the state-of-the-art claim remains tied to a single site's data."],"forward_implications":["If the reported DSC is reproducible, HES-UNet could serve as a computer-aided detection tool for hepatic echinococcosis, reducing the time and expertise needed to outline lesions on CT.","The MDB's wavelet-based downsampling could be applied to other segmentation tasks with fine boundaries and low contrast, such as liver tumors or kidney stones, where max-pooling loses detail.","The MAB-MUB pairing suggests a general recipe for feeding multi-scale global context back into skip connections, potentially improving any U-Net-style architecture without a full Transformer.","Deep supervision at five decoder stages provides a training signal that may make the model converge faster and more stably on small medical datasets."],"supporting_citations":[{"why":"Defines the U-Net backbone that HES-UNet modifies and is the primary baseline (DSC 86.58%).","marker":"[6]"},{"why":"TransUNet is the strongest baseline (DSC 88.12%); HES-UNet claims a 1.09-point improvement over it.","marker":"[7]"},{"why":"Swin-UNet is a transformer-based baseline (DSC 87.65%) that HES-UNet compares against.","marker":"[9]"},{"why":"U-Net++ is a baseline (DSC 86.76%) that HES-UNet surpasses.","marker":"[12]"},{"why":"Haar wavelet downsampling is the basis for the proposed multi-directional downsampling block (MDB).","marker":"[14]"},{"why":"Supplies the GHPA attention module used in encoder and decoder blocks, and EGE-UNet serves as a baseline.","marker":"[17]"},{"why":"CBAM attention extracts channel and spatial attention in the encoder blocks.","marker":"[18]"},{"why":"Provides the global attention module design used in the MGF module to filter redundant features.","marker":"[20]"}],"fun_headline_variants":["HES-UNet beats six baselines on liver echinococcosis lesions","Wavelet downsampling boosts U-Net for liver parasite segmentation","HES-UNet hits 89.21% Dice on echinococcosis CT scans","Multi-scale U-Net edges TransUNet on liver parasite lesions","New U-Net variant improves liver parasite lesion segmentation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The CT slices were randomly split 80/10/10 without grouping by patient, so slices from the same patient may appear in both training and test sets, meaning the reported Dice could reflect memorization of patients rather than generalization to new patients.","fun_headline_variants_meta":{"raw":{"variants":["HES-UNet beats six baselines on liver echinococcosis lesions","Wavelet downsampling boosts U-Net for liver parasite segmentation","HES-UNet hits 89.21% Dice on echinococcosis CT scans","Multi-scale U-Net edges TransUNet on liver parasite lesions","New U-Net variant improves liver parasite lesion segmentation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000403,"raw_usage":{"total_tokens":2145,"prompt_tokens":1038,"completion_tokens":1107,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":654,"completion_tokens_details":{"reasoning_tokens":1008}},"tokens_in":654,"tokens_out":1107,"duration_ms":10647,"temperature":1.0,"reasoning_tokens":1008,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T19:32:18.838873+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A concrete check is to re-split the dataset by patient (e.g., 80% of the 268 patients for training, 10% validation, 10% test) and rerun the identical HES-UNet training protocol; if the DSC drops well below 89.21%, the reported state-of-the-art result is an artifact of slice-level leakage rather than a property of the model.","supporting_citations":[{"cited_title":"Haar wavelet downsampling: A simple but effective downsampling module for semantic segmentation,","cited_arxiv_id":null,"evidence_quote":"Haar wavelet downsampling is the basis for the proposed multi-directional downsampling block (MDB)."},{"cited_title":"Ege-unet: an efficient group enhanced unet for skin lesion segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the GHPA attention module used in encoder and decoder blocks, and EGE-UNet serves as a baseline."},{"cited_title":"Cbam: Convolutional block attention module,","cited_arxiv_id":null,"evidence_quote":"CBAM attention extracts channel and spatial attention in the encoder blocks."}],"review_version":1}