Pith. sign in

REVIEW 4 major objections 5 minor 19 references

U-WNO:U-Net-enhanced Wavelet Neural Operator for fetal head segmentation

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

Pith's one-line read 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.

desk verdict 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. read the letter →

arxiv 2411.16890 v1 pith:SM5YYS6C submitted 2024-11-25 eess.IV cs.CV

classification eess.IVcs.CV
keywords OperatorLearningWaveletNeuralImageSegmentationFetalultrasoundU-NetComputerVisionNetworksDicescore
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

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.

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 (4)
  1. [Section 4, Results] 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.
  2. [Section 3, final paragraph] 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.
  3. [Abstract and Section 4] 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.
  4. [Section 4, Results] 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.
minor comments (5)
  1. [Section 2, Dataset] 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.
  2. [Section 4, Results] 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.
  3. [Section 1, Introduction] The text says 'fetal presentation' but spells it as 'fatal presentation'; this typo should be corrected.
  4. [References] 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.
  5. [Section 3, paragraph 2] 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.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found; the paper reports a measured Dice score and does not derive its central claim from its own assumptions by construction.

full rationale

The paper contains no derivation chain that reduces to its own inputs. U-WNO is a proposed architecture assembled from published components (WNO, ConvNet, U-Net), and its reported result is an empirical Dice score measured on a held-out test set. The central claim of 'strong performance' is an interpretive judgment based on that measured value, not a quantity fitted from or defined by the architecture itself. No parameter is fitted to a subset of data and then renamed as a prediction. The reliance on the WNO paper by Tripura and Chakraborty is a normal citation of prior work, not a load-bearing self-citation, and the authors do not invoke any uniqueness theorem from their own prior work. The weakness of the paper lies in insufficient evaluation (single run, no baseline, no error bars), which is a correctness and evidentiary concern, not circularity. Accordingly, the appropriate circularity score is 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new physical entities. Its central claim relies on unspecified training details, a post-hoc selected metric, and an untested architectural fusion rule.

free parameters (3)
  • maximum Dice score selection = 0.65
    The reported result is the maximum over the training run (500 epochs), not a test-set score selected by validation; this post-hoc selection inflates the apparent performance.
  • training hyperparameters = not reported
    Learning rate, loss function, optimizer settings, layer widths, and number of filters are not stated; the reported Dice depends on these choices, which are chosen by hand.
  • image resize to 128x128 and db4 wavelet basis = 128x128, db4
    Downsampling from native ultrasound resolution and the choice of Daubechies 4 wavelet are modeling choices that affect the measured Dice and are not justified empirically.
assumptions (3)
  • standard math Wavelet admissibility condition Cψ ensures lossless representation
    Invoked in Section 3; standard wavelet theory.
  • domain assumption The HC18 dataset masks and the train/test split used by the authors give an unbiased evaluation
    Section 2 describes the split only vaguely; the paper does not describe the split procedure or compare with the official challenge protocol.
  • ad hoc to paper The simple additive fusion x1+x2+x3 of wavelet, ConvNet, and U-Net branches improves segmentation
    Section 3 states the fusion as x1+x2+x3 with no ablation or theoretical justification; this is load-bearing for the 'unified operator' claim.

how reviews work

0 comments
Cite this review

Pith. "Pith review of U-WNO:U-Net-enhanced Wavelet Neural Operator for fetal head segmentation." pith.science (2026). https://pith.science/paper/SM5YYS6C

@misc{pith2026241116890,
  author       = {Pith},
  title        = {Pith review of: U-WNO:U-Net-enhanced Wavelet Neural Operator for fetal head segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SM5YYS6C}},
  note         = {Machine review of arXiv:2411.16890}
}
read the original abstract

This article describes the development of a novel U-Net-enhanced Wavelet Neural Operator (U-WNO),which combines wavelet decomposition, operator learning, and an encoder-decoder mechanism. This approach harnesses the superiority of the wavelets in time frequency localization of the functions, and the combine down-sampling and up-sampling operations to generate the segmentation map to enable accurate tracking of patterns in spatial domain and effective learning of the functional mappings to perform regional segmentation. By bridging the gap between theoretical advancements and practical applications, the U-WNO holds potential for significant impact in multiple science and industrial fields, facilitating more accurate decision-making and improved operational efficiencies. The operator is demonstrated for different pregnancy trimesters, utilizing two-dimensional ultrasound images.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

19 extracted references · 3 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address archive author booktitle chapter doi edition editor eid eprint howpublished institution journal key keywords month note number organization pages publisher school series title type url volume year archivePrefix primaryClass adsurl adsnote version label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sent...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION add.period duplicate empty 'skip "." * add.blank if FUNCTION if.digit duplicate "0" = swap duplicate "1" = swap duplicate "2" = swap duplicate "3" = swap duplicate "4" = swap duplicate "5" = swap duplicate "6" = swap duplicate "7" = swap duplicate "8" = swap "9" = or or or or or or or or or FUNCTION ...

  3. [3]

    Best Pract Res Clin Obstet Gynaecol 28

    Dias T, Sairam S, Kumarasiri S (2014) Ultrasound diagnosis of fetal renal abnormalities. Best Pract Res Clin Obstet Gynaecol 28. doi:10.1016/j.bpobgyn.2014.01.009

  4. [4]

    Med Image Anal 83:102629

    Fiorentino MC, Villani F, di Cosmo M, et al (2022) A review on deep-learning algorithms for fetal ultrasound-image analysis. Med Image Anal 83:102629. doi:10.1016/j.media.2022.102629

  5. [5]

    Available at: ://arxiv.org/abs/2010.08895

    Li Z, Kovachki N, Azizzadenesheli K, et al (2021) Fourier neural operator for parametric partial differential equations. Available at: ://arxiv.org/abs/2010.08895

  6. [6]

    Fetal Ultrasound Image Segmentation for Measuring Biometric Parameters Using Multi-Task Deep Learning

    Sobhaninia Z, Rafiei S, Emami A, et al (2019) Fetal ultrasound image segmentation for measuring biometric parameters using multi-task deep learning. Available at: ://arxiv.org/abs/1909.00273

  7. [7]

    Available at: ://arxiv.org/abs/2205.02191

    Tripura T, Chakraborty S (2022) Wavelet neural operator: A neural operator for parametric partial differential equations. Available at: ://arxiv.org/abs/2205.02191

  8. [8]

    Phys Med Biol 51(13):R83

    Wells PNT (2006) Ultrasound imaging. Phys Med Biol 51(13):R83. doi:10.1088/0031-9155/51/13/R06, ://dx.doi.org/10.1088/0031-9155/51/13/R06

Show all 19 references
  1. [9]

    Available at: ://arxiv.org/abs/2109.03697

    Wen G, Li Z, Azizzadenesheli K, et al (2022) U-FNO -- an enhanced fourier neural operator-based deep-learning model for multiphase flow. Available at: ://arxiv.org/abs/2109.03697

  2. [10]

    write newline

    " write newline " cite write " FUNCTION editor.postfix editor num.names #1 > "( )" "( )" if FUNCTION editor.trans.postfix editor num.names #1 > "( )" "( )" if FUNCTION trans.postfix translator num.names #1 > "( )" "( )" if FUNCTION authors.editors.reflist.apa5 'field := 'dot :...

  3. [11]

    , " * write output.state after.block = add.period write newline

    ENTRY address author booktitle chapter doi edition editor eid howpublished institution journal key keywords month note number organization pages publisher school series title type url volume year eprint archive archivePrefix primaryClass adsurl adsnote version label INTEGERS o...

  4. [12]

    write newline

    " write newline "" before.all 'output.state := FUNCTION if.digit duplicate "0" = swap duplicate "1" = swap duplicate "2" = swap duplicate "3" = swap duplicate "4" = swap duplicate "5" = swap duplicate "6" = swap duplicate "7" = swap duplicate "8" = swap "9" = or or or or or or...

  5. [13]

    write newline

    " write newline "" before.all 'output.state := FUNCTION output.doi doi empty skip "doi:" doi * "" * output if FUNCTION format.archive archivePrefix empty "" archivePrefix ":" * if FUNCTION format.primaryClass primaryClass empty "" " [" primaryClass * "] " * if FUNCTION format....

  6. [14]

    write newline

    " write newline "" before.all 'output.state := FUNCTION string.to.integer 't := t text.length 'k := #1 'char.num := t char.num #1 substring 's := s is.num s "." = or char.num k = not and char.num #1 + 'char.num := while char.num #1 - 'char.num := t #1 char.num substring FUNCTI...

  7. [15]

    write newline

    " write newline "" before.all 'output.state := FUNCTION string.to.integer 't := t text.length 'k := #1 'char.num := t char.num #1 substring 's := s is.num s "." = or char.num k = not and char.num #1 + 'char.num := while char.num #1 - 'char.num := t #1 char.num substring FUNCTI...

  8. [16]

    , " * write output.state after.block = add.period write newline

    ENTRY address archive author booktitle chapter edition editor eprint howpublished institution journal key keywords month note number organization pages publisher school series title type url doi volume year archivePrefix primaryClass eid adsurl adsnote version label INTEGERS o...

  9. [17]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  10. [18]

    Available from:

    ENTRY address assignee author booktitle chapter cartographer day edition editor howpublished institution inventor journal key keywords month note number organization pages part publisher school series title type volume word year eprint doi url lastchecked updated archive archi...

  11. [19]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.