{"id":"3a3d349c-028b-49f5-be53-50388873dbf7","arxiv_id":"2509.08860","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"USEANet is a 3.64M-parameter U-Net variant for ultrasound segmentation whose reported advantage over baselines is inconsistent with its own results on DDTI and TN3K.","lead":"A new lightweight deep network, USEANet, is proposed for ultrasound image segmentation, with multi-branch processing for noise, edges, and contrast. The paper claims it outperforms all baselines on five ultrasound datasets, though its own results show a transformer baseline winning on two of them.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1 contradicts the central claim: XboundFormer beats USEANet on BUS-BRA, DDTI, and TN3K, so 'superior across all five' fails as printed.","rationale":"I read the paper in good faith: USEANet is a plausible lightweight architecture and the code is released, which is a positive. However, the headline contribution is an empirical performance claim, and that claim is internally contradicted by the paper's own Table 1. Because the contradiction appears in the published numbers, it is more decisive than the possible patient-split issue: even under a clean split, the table does not support 'superior across all five.' A reproduction test is the cleanest way to confirm the contradiction or expose a typographical error. The reader's formal weakest_assumption (split level) is different, but their rationale also flags the table overclaim, so my agreement is partial. The verdict remains REJECT, unchanged by this stress-test pass.","tokens_in":6213,"tokens_out":7259,"duration_ms":70787,"concrete_test":"Reproduce the full benchmark with the authors' released code (https://github.com/chouheiwa/USEANet) on all five datasets using the stated 70:15:15 split and evaluation protocol, recomputing per-dataset and mean IoU/F1 for USEANet and all nine baselines. If the reproduced values match Table 1, the claim of 'superior performance across all five datasets' is false as printed (XboundFormer wins three datasets and the mean). If the reproduced values differ, the paper's reported numbers need correction before the claim can be assessed. This single check settles whether the contradiction is real or a table typo.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim (abstract and §3.3) is that USEANet 'significantly outperforms existing methods across five ultrasound datasets' and 'achieves superior performance across all five datasets.' Table 1 is the sole quantitative support and contradicts this: on BUS-BRA, XboundFormer reports IoU 81.56 vs USEANet 80.63; on DDTI, 83.29 vs 81.72; on TN3K, 75.83 vs 74.81. USEANet is best on only two of five datasets. Even the average IoU across the five datasets slightly favors XboundFormer (≈74.94 vs ≈74.90). No error bars, confidence intervals, or significance tests are reported, so 'significantly outperforms' is unsupported by the presented evidence. The split-level concern in the Reader's report is plausible, but it is secondary: the numbers as printed already fail to establish the claim. The architecture and efficiency figures are not at issue; the falsifiable empirical claim is.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes USEANet, a lightweight encoder-decoder architecture for ultrasound image segmentation built on a PVT-B0 backbone (3.64M parameters, 0.79G FLOPs). The design includes a multi-branch feature processor for noise reduction, edge enhancement, and contrast enhancement; edge-aware attention mechanisms; hierarchical feature aggregation; and an ultrasound-aware decoder enhancement module. The method is evaluated on five public ultrasound datasets (BUSI, BUS-BRA, Breast-Lesions-USG, DDTI, TN3K) and compared with nine baselines. The abstract and Section 3.3 claim that USEANet 'significantly outperforms existing methods across five ultrasound datasets' and 'achieves superior performance across all five datasets.'","tokens_in":6526,"tokens_out":3505,"duration_ms":42299,"significance":"If the central performance claim were correct, the paper would offer a useful contribution to efficient ultrasound segmentation: the parameter and FLOP counts are genuinely low, the code is promised to be public, and the modular architecture is clearly described. The ablation study, although limited, does suggest each module contributes positively on BUSI. However, the primary empirical claim is contradicted by the paper's own Table 1, and no statistical evidence is provided for any 'significant' improvement. The efficiency story is plausible, but the accuracy-superiority claim is not supported as presented.","major_comments":[{"comment":"The central claim that USEANet 'achieves superior performance across all five datasets' is directly contradicted by Table 1. XboundFormer reports higher IoU on BUS-BRA (81.56 vs 80.63), DDTI (83.29 vs 81.72), and TN3K (75.83 vs 74.81). USEANet is the best only on BUSI and Breast-Lesions-USG. Even the average IoU across the five datasets is slightly in favor of XboundFormer (~74.94 vs ~74.90). The abstract's 'significantly outperforms existing methods across five ultrasound datasets' is therefore not supported by the presented numbers.","section":"Abstract and §3.3, Table 1"},{"comment":"All results are reported as point estimates from single runs, with no error bars, confidence intervals, or statistical significance tests. The word 'significantly' appears in the abstract and throughout §3.3, but no test (e.g., paired t-test or Wilcoxon) is reported. Without variance estimates or repeated runs, the improvements over UNet/AttUNet on BUSI and the small margins in other datasets cannot be evaluated. This is a load-bearing deficiency for an empirical paper that makes a superiority claim.","section":"§3.2, Evaluation Setup"},{"comment":"The 70:15:15 random split is not specified as patient-level. Public ultrasound datasets such as BUSI, BUS-BRA, DDTI, and TN3K commonly contain multiple images per patient. If the split is at image level, same-patient images can appear in both training and test sets, inflating the reported IoU values and making comparisons with literature results unreliable. The authors should clarify the split level and, if necessary, re-run experiments with a patient-level split.","section":"§3.1, Datasets"},{"comment":"The ablation study is limited to a single dataset (BUSI) and a single run, with no variance reporting. In addition, 'w/o Ultrasound-Specific' removes the multi-branch module that contains three branches at once, so the individual contributions of noise reduction, edge enhancement, and contrast enhancement are not isolated. The stated gains (2.85 IoU for 'ultrasound-specific modules') may be dominated by parameter count or training dynamics. This is secondary to the main claim, but it weakens the paper's component-wise validation.","section":"§3.4, Table 2"}],"minor_comments":[{"comment":"The title renders 'Edge-Aware' as 'Edge-A W are' and the author name 'Baha lhnaini' appears to contain a typo. These should be corrected.","section":"Title and author affiliations"},{"comment":"The metric is called 'mIoU' in the evaluation setup but 'IoU' in Table 1 and elsewhere. The naming should be consistent.","section":"§3.2 and Table 1"},{"comment":"The dimensions of W_i in the adaptive fusion are not specified. Since W = Softmax(FC(GAP(Fcat))), it is unclear how the scalar/vector weights are broadcast over the feature maps F^i_branch. Please clarify the tensor shapes.","section":"§2.2, Eq. (2)"},{"comment":"The hyperparameters alpha and beta are reused with different meanings (edge enhancement α=0.1 in Eq. (2), EAA β=0.1 in Eq. (4), and HFA α=0.5 in Eq. (5)). This is confusing; use distinct symbols or define each locally.","section":"§2.3 and §2.4"},{"comment":"Reference [12] lists the author as 'Pawłwska', which appears to be a typo for 'Pawłowska'. Please verify all author names and dataset citations.","section":"References"},{"comment":"The conclusion states 'Experimental results demonstrate superior segmentation performance,' but given the Table 1 results, a more measured statement such as 'competitive performance with strong efficiency' would be accurate. The paper would benefit from a limitations paragraph.","section":"Conclusion"}],"recommendation":"reject","confidential_remarks":"The main reason for rejection is the mismatch between the paper's central claim and its own results in Table 1. This is not a stylistic issue: the paper says 'superior across all five datasets' while the data show otherwise. The architecture and efficiency numbers are interesting, and a revised manuscript framed as a lightweight, competitive alternative to heavier transformers—with patient-level splits and variance estimates—could be reconsidered. As it stands, the core selling point is not supported."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read USEANet. The headline is simple: the paper claims to outperform all baselines on five ultrasound datasets, but its own Table 1 shows XboundFormer beating it on three of the five, and the average IoU is a virtual tie. That's not a rounding issue. On BUS-BRA, XboundFormer gets 81.56 vs 80.63; on DDTI 83.29 vs 81.72; on TN3K 75.83 vs 74.81. The absolute differences are small, but so are the ones in USEANet's favor on the other two. There's no way to call that 'significantly outperforms existing methods.'\n\nNow the upside. The architecture is a sensible engineering effort: PVT-B0 backbone with multi-branch noise/edge/contrast processors, edge-aware attention, hierarchical fusion, and a foreground-background gating decoder. Each component is a known trick, but the assembly is new and the resource budget (3.64M params, 0.79G FLOPs) is genuinely lightweight. Code is available, which gives me some confidence the experiments actually ran. The ablation on BUSI shows each module contributes a couple of IoU points, which at least suggests the components aren't inert.\n\nThe soft spots beyond the overclaim: no error bars, no significance tests, and the evaluation protocol only says 'random 70:15:15 split' without saying whether the split is per-image or per-patient. If it's per-image, the numbers are inflated. That concern is real but secondary: the printed numbers already fail to support the abstract.\n\nThe paper is not incoherent or unserious; the engineering is fine and the task is practical. But the evaluation does not support the stated conclusion. I'd send it to review if I were an editor, because a good referee can point out exactly what needs fixing, and the underlying architecture might be worth salvaging. But I wouldn't trust the numbers as-is, and I wouldn't cite the performance claims.\n\nFor your reading group, it's a useful case study in how to spot a contradicted claim, but not because of the network design. Maybe include it if you want a discussion of overclaiming in medical imaging papers.","headline":"The architecture has merit, but the paper's central claim of SOTA performance is contradicted by its own Table 1.","tokens_in":6947,"tokens_out":2840,"would_cite":false,"duration_ms":31515,"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 3.64M-parameter ultrasound-specific edge-aware network beats nine baselines on five ultrasound datasets, reporting 67.01 IoU on BUSI at 0.79G FLOPs.","keywords":["ultrasound image segmentation","edge-aware attention","multi-branch network","lightweight neural network","medical image analysis","PVT-B0 backbone","BUSI dataset"],"falsifier":"Re-run all five datasets with a patient-level 70:15:15 split, ensuring every image of a patient stays in one partition, and compare BUSI IoU against the same nine baselines. If the reported margin collapses or reverses, the performance claim is an artifact of leakage; if the margin persists, the central claim survives.","tokens_in":6168,"feed_emoji":"🩺","tokens_out":8383,"duration_ms":78371,"temperature":0.7,"pith_summary":"The paper proposes USEANet, a compact encoder-decoder network for ultrasound image segmentation, and argues that a network designed around ultrasound's own failure modes—speckle noise, low contrast, ambiguous boundaries—can beat both heavy transformers and extremely small generic models. If the claims hold, accurate segmentation becomes feasible on portable ultrasound machines without sacrificing accuracy. The evidence is a five-dataset comparison against nine baselines plus ablations on BUSI attributing each component's contribution. The headline result is 67.01 IoU (Intersection over Union) on BUSI at 3.64M parameters and 0.79G FLOPs.","feed_headline":"Small ultrasound network beats nine baselines on five datasets","feed_subtitle":"Edge-aware design hits 67.01 IoU on BUSI at 0.79 GFLOPs, making real-time portable segmentation plausible.","key_machinery":"The load-bearing object is the Multi-Branch Feature Processor (MBFP), which runs three parallel streams on each encoder stage: a noise-reduction branch built from residual depthwise-separable convolutions, an edge-enhancement branch that adds a scaled Laplacian-filtered response, and a contrast-enhancement branch that applies channel attention to a global-pooled feature. Around it sit Edge-Aware Attention (EAA), which gates features with gradient information; Hierarchical Feature Aggregation (HFA), which multiplies multi-scale features from coarser levels into finer ones; and Ultrasound-Aware Decoder Enhancement (UADE), which uses the foreground-minus-background prediction as a multiplicativ","core_discovery":"The central claim is that a lightweight ultrasound-specific design can close the gap to much larger models while beating them on ultrasound-specific metrics. USEANet reports the best mIoU/F1 on all five datasets—BUSI, BUS-BRA, Breast-Lesions-USG, DDTI, TN3K—with 3.64M parameters and 0.79G FLOPs. Its BUSI IoU of 67.01 sits 8.69 points above U-Net and above transformer baselines such as SwinUnet and XboundFormer at a fraction of their compute. Ablations on BUSI attribute 2.79 IoU points to the attention mechanisms, 2.42 to the multi-branch architecture, 2.85 to ultrasound-specific modules, and 2.99 to multi-scale aggregation.","pith_inferences":["The split is described only as random 70:15:15; if it is image-level rather than patient-level, the reported margins could shrink under patient-stratified evaluation. A re-run with patient-level splits would test the generalisation claim.","The four modules are presented as portable: MBFP, EAA, HFA, and UADE could be grafted onto another lightweight encoder, though the paper only validates them in combination.","The observed frontier implies an optimal operating point around 3-4M parameters for ultrasound: 0.05M models lack capacity, while 25-35M transformers buy little extra accuracy."],"forward_implications":["Ultrasound segmentation no longer requires a heavyweight model: transformer-scale accuracy is reported at 7.5x fewer parameters than SwinUnet and 8.1x fewer than XboundFormer.","Extremely small skin-lesion models (0.05M-0.33M parameters) fall far behind on ultrasound, with EGEUNet, MALUNet, and PMFSNet trailing USEANet by roughly 21-25 IoU points on BUSI.","Each of the four modules contributes independently, with multi-scale aggregation providing the largest single gain (2.99 IoU on BUSI).","The 0.79G FLOPs footprint puts real-time deployment on resource-constrained hardware within reach."],"supporting_citations":[{"why":"Provides the ultra-lightweight PVT-B0 encoder backbone that produces the four multi-scale feature maps USEANet processes.","marker":"[8]"},{"why":"Standard U-Net baseline whose BUSI IoU (58.32) is the main lower bar USEANet claims to beat by 8.69 points.","marker":"[1]"},{"why":"Attention U-Net baseline included in all five dataset comparisons.","marker":"[2]"},{"why":"Transformer-based SwinUnet baseline; USEANet claims 7.5x fewer parameters while matching or beating accuracy.","marker":"[3]"},{"why":"Cross-scale boundary-modeling transformer baseline; its boundary focus is the closest competing idea to USEANet's edge-aware modules.","marker":"[4]"},{"why":"UNext, an MLP-based lightweight baseline that represents the lightweight approaches USEANet argues degrade on ultrasound.","marker":"[5]"},{"why":"MALUNet, a multi-attention lightweight U-Net baseline designed for skin lesions; USEANet claims it underperforms on ultrasound.","marker":"[6]"},{"why":"EGEUNet, the extremely lightweight 0.05M-parameter baseline; its poor ultrasound results motivate the domain-specific design.","marker":"[7]"},{"why":"BUSI dataset is the main benchmark for the headline IoU number and for the ablation study.","marker":"[10]"},{"why":"TN3K, the largest of the five thyroid datasets, used in the multi-dataset generalisation claim.","marker":"[14]"}],"fun_headline_variants":["Ultrasound net: 3.64M params, top scores on 5 sets","Edge-aware US segmentation, 8.7 IoU over U-Net at 0.79 GFLOPs","Lightweight U-Net rival: beats 9 baselines on 5 ultrasound datasets","Tiny ultrasound model, big gains: 67.01 IoU on BUSI at 0.79 GFLOPs","USEANet: ultrasound-specific edge focus, wins on all five datasets"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The evaluation assumes the random 70:15:15 split is patient-level—no image from the same patient appears in more than one partition—but the paper never states this; if the split is image-level, the reported IoU gains could be inflated.","fun_headline_variants_meta":{"raw":{"variants":["Ultrasound net: 3.64M params, top scores on 5 sets","Edge-aware US segmentation, 8.7 IoU over U-Net at 0.79 GFLOPs","Lightweight U-Net rival: beats 9 baselines on 5 ultrasound datasets","Tiny ultrasound model, big gains: 67.01 IoU on BUSI at 0.79 GFLOPs","USEANet: ultrasound-specific edge focus, wins on all five datasets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000153,"raw_usage":{"total_tokens":1050,"prompt_tokens":757,"completion_tokens":293,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":501,"completion_tokens_details":{"reasoning_tokens":168}},"tokens_in":501,"tokens_out":293,"duration_ms":3833,"temperature":1.0,"reasoning_tokens":168,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T21:05:34.548822+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run all five datasets with a patient-level 70:15:15 split, ensuring every image of a patient stays in one partition, and compare BUSI IoU against the same nine baselines. If the reported margin collapses or reverses, the performance claim is an artifact of leakage; if the margin persists, the central claim survives.","supporting_citations":[],"review_version":1}