{"id":"49c85d57-e29d-45f2-9618-1f8a921ab259","arxiv_id":"2411.16890","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"A U-Net enhanced wavelet neural operator for fetal head ultrasound segmentation is reported, with a maximum Dice score of 0.65 and no baseline comparison.","lead":"The authors combine a wavelet neural operator with a U-Net style decoder and apply the result to fetal head segmentation in 2D ultrasound images, reporting a maximum Dice score of about 0.65 on the HC18 test set. The paper is only significant if that score is competitive, but it offers no comparison with a U-Net baseline or published HC18 results.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed strong segmentation performance rests on a single maximum Dice of 0.65 with no baseline, variance, or reproducibility check; this is insufficient to support the central claim.","rationale":"The reader's weakest_assumption identifies the same load-bearing issue: a single maximum Dice of 0.65 without baselines, error bars, or multiple seeds is insufficient evidence for the claim of strong performance. My read of the paper confirms this: the Results section contains only this one quantitative result, and the abstract and conclusion extend it to clinical and educational impact. The known HC18 benchmark provides a clear external reference showing that 0.65 is far below the state of the art, so the reported number actively undercuts the claim rather than supporting it. Additionally, the manuscript's own statement that code access is deferred until after review is a self-acknowledged reproducibility limitation. Because the central claim is unsupported by adequate evidence, the REJECT verdict is appropriate. No further adjustment is needed; the concern is aligned with the reader's assessment.","tokens_in":5556,"tokens_out":1991,"duration_ms":20131,"concrete_test":"Re-run the experiment on the HC18 dataset with an identical training setup: same 128x128 resizing, same train/test split, same batch size, same 500 epochs, and same loss function, but with a standard U-Net as the model. Report the mean and standard deviation of Dice over at least 5 random seeds for both U-WNO and U-Net. If the U-Net baseline achieves Dice above 0.9 while U-WNO remains near 0.65, the claim of strong performance is contradicted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4 reports a maximum Dice score of approximately 0.65 after 500 epochs on a fixed train/test split, with no baseline comparison, no error bars, and no multiple-seed statistics. The central claim of 'strong performance' depends entirely on this number. On the HC18 challenge, which is the source dataset, published methods exceed 0.95 Dice; even a plain U-Net typically achieves near 0.9 or higher on this task. Without any comparison to a U-Net baseline (or any baseline), a score of 0.65 cannot be interpreted as strong. The paper's own description of the architecture is also too vague to reproduce: the fusion of wavelet output, ConvNet output, and U-Net output is described only as 'x1 + x2 + x3', with no precise tensor shapes, feature dimensions, or operation order. The repository link is provided but access is deferred until after review, so no independent verification is currently possible. The load-bearing assumption is that a single maximum Dice of 0.65, taken in isolation, supports the claim of strong segmentation performance and the broader clinical and educational conclusions. This assumption is not justified by the evidence presented.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes U-WNO, a hybrid architecture combining a wavelet neural operator, a ConvNet, and a U-Net branch, applied to fetal head segmentation in 2D ultrasound images from the HC18 challenge. The branches are fused as x1+x2+x3, followed by fully convolutional layers. The model is trained for 500 epochs on images resized to 128x128, and the paper reports a maximum Dice score of approximately 0.65. From this single number, it concludes that the method achieves strong performance and has potential clinical and educational applications. The code repository is provided, but access is deferred until after review.","tokens_in":5925,"tokens_out":5793,"duration_ms":53208,"significance":"If validated by a proper comparison, a neural-operator-based segmentation model could be an interesting research direction. The paper has strengths: it uses a public benchmark dataset, defines the Dice score explicitly, and proposes a simple fusion idea. There is no circularity issue, since the Dice score is a measured quantity; the problem is empirical support. The current evidence, however, is a single maximum Dice value with no baseline, no error bars, no multiple seeds, and no ablation, and on the HC18 benchmark the reported value is far below published results. The manuscript's central claim of strong performance and its clinical and educational conclusions are therefore not supported by the data as presented.","major_comments":[{"comment":"The only quantitative result is a single maximum Dice of approximately 0.65, obtained after 500 epochs on a fixed train/test split. There is no baseline comparison, no standard deviation, no multiple-seed experiment, and no validation-based model selection. On the HC18 challenge, which is the source dataset, established methods report Dice scores well above 0.9; a value of 0.65, taken in isolation, cannot support the sentence 'yields strong performance in the segmentation task.' This is the load-bearing evidence for the paper's central claim, so the claim is currently unsupported.","section":"Section 4, Results"},{"comment":"The fusion operation is described only as 'x1 + x2 + x3', with no information about tensor shapes, channel counts, what x1 and x2 represent after the inverse wavelet transform and the ConvNet branch, or how the U-Net output is aligned before summation. The repository link is given, but access is deferred until after review (Section 4), so the architecture cannot currently be verified. This makes the method irreproducible as described.","section":"Section 3, final paragraph"},{"comment":"The abstract states that the operator is demonstrated for different pregnancy trimesters, but the results section reports only an aggregate Dice score without any trimester-specific breakdown. If the claim is to remain, per-trimester results must be reported; otherwise the statement overstates the evidence.","section":"Abstract and Section 4"},{"comment":"No ablation is provided for the three components. Because the contribution is the integration of WNO, ConvNet, and U-Net, the paper should at least compare U-WNO with U-Net alone and WNO alone. Without such comparisons, the reported architecture's added value is untestable.","section":"Section 4, Results"}],"minor_comments":[{"comment":"The dataset sizes are given as 'approximately 976' images and 'around 325' test images, which do not match the official HC18 split (999 training, 335 test); please clarify the exact preprocessing and split.","section":"Section 2, Dataset"},{"comment":"The manuscript contains an unedited drafting note: 'Here's a refined version with enhanced grammar, structure, and readability:' followed by the results paragraph. This editorial artifact should be removed so that the final text reads as the refined paragraph only.","section":"Section 4, Results"},{"comment":"The text says 'fetal presentation' but spells it as 'fatal presentation'; this typo should be corrected.","section":"Section 1, Introduction"},{"comment":"Several reference URLs are malformed, beginning with 'urlhttps://' (e.g., Li et al., Sobhaninia et al., Tripura and Chakraborty, Wen et al.); these should be corrected.","section":"References"},{"comment":"The claim that WNO 'exhibits superior performance compared to FNO in the spatial domain' is made without a specific citation or quantitative comparison; it should be either supported by the cited work or softened.","section":"Section 3, paragraph 2"}],"recommendation":"reject","confidential_remarks":"The manuscript appears to be an early draft whose central quantitative claim is contradicted by its own reported Dice score. The absence of any baseline or error analysis means the result cannot be interpreted, and the architectural description is incomplete. If the authors can add a U-Net baseline, ablations, and sufficiently improved and reproducible results, a new submission could be appropriate; however, these changes go beyond a routine revision of the current manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this paper's central claim—that U-WNO gives strong fetal head segmentation—is not supported by the evidence it presents. A single maximum Dice of 0.65 on HC18, with no baseline, no error bars, and no multiple-seed statistics, is a red flag; on that challenge, standard U-Nets are typically in the 0.9x range. The paper itself calls 0.65 \"strong performance,\" which is the load-bearing overinterpretation.\n\nWhat's actually new: very little. The authors cite U-FNO, and U-WNO is exactly the same recipe with the Fourier block swapped for a wavelet block. That is a routine engineering combination, not a new algorithmic idea. The paper does give a reasonable summary of WNO and explains the choice of db4 wavelets, and it is honest in one respect: it reports the actual Dice number rather than hiding it. But the report is then framed in a way that misreads a low score as a success.\n\nThe soft spots are the core, not the edges. The architecture description is too vague to reproduce: the fusion of wavelet output, ConvNet output, and U-Net output is just \"x1 + x2 + x3\" with no tensor shapes or operation order. The code is promised but withheld until after review, which means no independent check is possible now. The evaluation has no validation-based model selection, no comparison to a plain U-Net or any other baseline, and no variance estimate. The clinical and educational conclusions in the abstract and conclusion are speculative and go well beyond what a 0.65 Dice score can support.\n\nI don't think there is a hidden gem here. The idea is a direct analogue of U-FNO, and the evidence is far too thin to establish the claimed benefit. The paper could be improved by running a proper comparison with a U-Net baseline, reporting mean and standard deviation over seeds, and making the code available. As submitted, it does not meet the bar for serious peer review; a desk reject is appropriate. I would not bring this to reading group, and I would not cite it in its current form. If the authors add the missing baseline and reproducibility, the comparison might be worth a second look, but right now the load-bearing assumption fails.","headline":"A single maximum Dice of 0.65 with no baseline or error bars cannot support the paper's claim of strong segmentation performance; the architecture is a routine wavelet swap of U-FNO and the evidence does not justify peer review.","tokens_in":6261,"tokens_out":2087,"would_cite":false,"duration_ms":18732,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that U-WNO, a U-Net-enhanced Wavelet Neural Operator, achieves strong fetal head segmentation with a maximum Dice score of about 0.65 on the HC18 test set.","keywords":["Operator Learning","Wavelet Neural Operator","Image Segmentation","Fetal ultrasound","U-Net","Computer Vision","Neural Networks","Dice score"],"falsifier":"Train a standard U-Net on the same HC18 split, at the same 128x128 resolution and 500-epoch schedule, and compare test Dice; if the plain U-Net matches or exceeds the reported 0.65, the contribution of the wavelet operator branch is not supported. Repeating U-WNO training with five random seeds and checking whether the mean minus one standard deviation stays above 0.65 would also test whether the single maximum is a reliable estimate.","tokens_in":5367,"feed_emoji":"🩺","tokens_out":7806,"duration_ms":63374,"temperature":0.7,"pith_summary":"The paper proposes U-WNO, a segmentation architecture that combines a Wavelet Neural Operator, a ConvNet, and a U-Net encoder-decoder, and applies it to fetal head segmentation in two-dimensional ultrasound images. It reports a maximum Dice score of approximately 0.65 on the HC18 test set after 500 epochs of training and interprets this as strong performance under noisy, low-resolution conditions. The authors present the model as a bridge from operator learning, originally developed for PDE problems, to practical medical image segmentation, with downstream benefits for clinical monitoring and trainee education. A sympathetic reading is that the paper is trying to establish that wavelet-based operator learning can be productively fused with U-Net to segment anatomical structures in challenging imaging data.","feed_headline":"Fusing wavelets with U-Net scores 0.65 Dice on fetal head segmentation","feed_subtitle":"The paper tests whether operator learning, borrowed from PDE solvers, transfers to ultrasound image segmentation.","key_machinery":"The load-bearing object is the U-WNO architecture itself, a three-branch fusion of a Wavelet Neural Operator, a ConvNet, and a U-Net. The WNO branch applies a Daubechies 4 wavelet transform, parameterizes the kernel in the wavelet domain as $K_\\psi(u)(x) = W^{-1}(K \\cdot W(u))(x)$, and returns an inverse-transformed approximation; the ConvNet branch transforms features without changing their shape; the U-Net branch produces segmentation maps through its contracting and expanding paths. The three outputs are concatenated as $x_1 + x_2 + x_3$ and processed by fully convolutional layers trained with the Adam optimizer. The Db4 wavelet is chosen for its balance of time and frequency localization, which the paper argues isolates critical features under noise and low contrast.","core_discovery":"On its own terms, the paper's central claim is that integrating the Wavelet Neural Operator (WNO) with U-Net as a unified operator yields strong fetal head segmentation performance. The evidence offered is a maximum Dice score of approximately 0.65, achieved on a roughly 325-image test set drawn from the HC18 challenge, using images resized to 128 by 128 pixels and a model trained for 500 epochs. The architecture concatenates wavelet-decomposed features (x1), ConvNet-transformed features (x2), and U-Net segmentation maps (x3) into a final output, after which fully convolutional layers with ReLU activation produce the segmentation map. The paper interprets this result as showing strong performance despite noise and reduced resolution, and extends the claim to potential clinical decision support and ultrasound training tools.","pith_inferences":["A direct comparison against a plain U-Net on the same 128x128 split would isolate the contribution of the wavelet neural operator branch from the U-Net backbone.","Repeating training over several random seeds would convert the single reported maximum Dice into a mean and confidence interval, giving a more reliable performance estimate.","Applying the same three-branch fusion to other anatomical segmentation tasks, such as placenta or fetal organs, would test whether the wavelet operator generalizes beyond the fetal head.","Injecting controlled noise into test images would test the paper's claim that wavelet decomposition specifically provides noise tolerance in ultrasound data."],"forward_implications":["If U-WNO performs as claimed, operator learning originally developed for PDEs can be transferred to medical image segmentation tasks.","The model is presented as able to segment fetal heads across different pregnancy trimesters from 2D ultrasound images.","The reported tolerance to noise and small image size would support deployment as a clinical decision aid.","The same architecture could serve as a simulated practice tool for training sonographers when real patients are unavailable."],"supporting_citations":[{"why":"Supplies the Wavelet Neural Operator that U-WNO builds on, including the wavelet-domain kernel parameterization and admissibility condition.","marker":"Tripura and Chakraborty, 2022"},{"why":"Defines the Fourier Neural Operator, the operator-learning baseline that WNO is claimed to surpass in spatial locality.","marker":"Li et al., 2021"},{"why":"Introduces U-FNO, the Fourier-based analogue of U-WNO, providing the template of enhancing a neural operator with U-Net.","marker":"Wen et al., 2022"},{"why":"Prior fetal ultrasound segmentation work that frames the task and the use of multi-task deep learning for biometric parameters.","marker":"Sobhaninia et al., 2019"},{"why":"Review of deep-learning fetal ultrasound analysis that grounds the claim that automated segmentation can support clinical decision-making.","marker":"Fiorentino et al., 2022"}],"fun_headline_variants":["Fusing wavelets and U-Net yields 0.65 Dice on fetal heads","U-WNO: A wavelet-U-Net hybrid for fetal head segmentation","From PDEs to ultrasound: U-WNO hits 0.65 Dice","New operator learning model scores 0.65 Dice on fetal heads"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on taking one maximum Dice score of about 0.65 from a single training run on one fixed split, with no error bars, no multiple seeds, and no comparison to a U-Net baseline, as sufficient evidence of strong segmentation performance.","fun_headline_variants_meta":{"raw":{"variants":["Fusing wavelets and U-Net yields 0.65 Dice on fetal heads","U-WNO: A wavelet-U-Net hybrid for fetal head segmentation","From PDEs to ultrasound: U-WNO hits 0.65 Dice","New operator learning model scores 0.65 Dice on fetal heads"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000229,"raw_usage":{"total_tokens":1419,"prompt_tokens":829,"completion_tokens":590,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":445,"completion_tokens_details":{"reasoning_tokens":508}},"tokens_in":445,"tokens_out":590,"duration_ms":6065,"temperature":1.0,"reasoning_tokens":508,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:45:30.045848+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a standard U-Net on the same HC18 split, at the same 128x128 resolution and 500-epoch schedule, and compare test Dice; if the plain U-Net matches or exceeds the reported 0.65, the contribution of the wavelet operator branch is not supported. Repeating U-WNO training with five random seeds and checking whether the mean minus one standard deviation stays above 0.65 would also test whether the single maximum is a reliable estimate.","supporting_citations":[{"cited_title":"Fetal Ultrasound Image Segmentation for Measuring Biometric Parameters Using Multi-Task Deep Learning","cited_arxiv_id":"1909.00273","evidence_quote":"Prior fetal ultrasound segmentation work that frames the task and the use of multi-task deep learning for biometric parameters."}],"review_version":1}