Pith. sign in

REVIEW 3 major objections 7 minor 1 cited by

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

T0 review · 3 major / 7 minor · reviewed 2026-08-14 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A multi-task network lifts fetal-head ultrasound segmentation to 96.8 Dice.

desk verdict A solid multi-task segmentation idea, but the augmentation-then-split protocol likely leaks training data into the test set and inflates every headline number. read the letter →

arxiv 1909.00273 v1 pith:R5GUY4OF submitted 2019-08-31 eess.IV cs.CVcs.LG

Zahra Sobhaninia , Shima Rafiei , Ali Emami , Nader Karimi , Kayvan Najarian , Shadrokh Samavi , S.M.Reza Soroushmehr This is my paper · ORCID
classification eess.IVcs.CVcs.LG
keywords fetalultrasoundheadcircumferenceimagesegmentationmulti-tasklearningLink-Netellipsefittingdeepbiometricparameters
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 tries to show that a single deep network can both segment the fetal head in 2D ultrasound images and fit an ellipse whose perimeter gives the head circumference, and that doing the two tasks together is better than segmentation alone. The proposed multi-task network reaches a Dice score of 96.84 percent, up from 92.67 percent for the same network without the ellipse branch, and it reports boundary errors below those of a leading non-deep method. If this holds, automatic head-circumference measurement could be made more reliable across all pregnancy trimesters, saving manual steps in prenatal screening. The cost is modest: five extra ellipse parameters and a compound loss that joins a Dice-plus-cross-entropy segmentation term with a mean-squared-error term on those parameters.

What carries the argument

The load-bearing object is the multi-task Link-Net (MTLN): an encoder-decoder segmentation network with residual encoder blocks, multi-scale inputs, and skip connections, augmented by an Ellipse Tuner made of three fully connected layers that read feature maps between encoder and decoder and output five ellipse parameters (center coordinates, the two radii, and an angle). The training objective is $L_T = \alpha_1 L_{Seg} + \alpha_2 L_{ET}$, where $L_{Seg}$ is a boundary-weighted sum of cross-entropy and Dice loss and $L_{ET}$ is the mean squared error between predicted and ground-truth ellipse parameters. The ellipse branch is not just an output head; its gradients flow back into the segmentation features, so the network is explicitly pushed toward masks whose boundaries trace an ellipse. The parametric ellipse equation ties the five outputs to the head-circumference measurement.

What would settle it

Re-run the competitor's method on the exact same 25% test split, compute DSC, DF, ADF and HD with the paper's formulas, and compare with the reported 96.84±2.89, 1.13±2.69, 2.12±1.87 and 1.72±1.39; if the competitor's HD and ADF are not worse, or if the single-task network reproduces the multi-task Dice when retrained, the central claims fail.

Watch

Extended reading notes

Core claim

The central claim is that adding an Ellipse Tuner to a Link-Net segmentation architecture and training with a compound loss simultaneously improves segmentation and produces a direct head-circumference estimate. The paper reports a DSC of 92.67±2.70 for the single-task network and 96.84±2.89 for the multi-task network, with the multi-task outputs smoother, cleaner, and more elliptical. Against the prior automated method, the multi-task network achieves a Hausdorff distance of 1.72±1.39 mm and an absolute head-circumference difference of 2.12±1.87 mm, both better than the competitor's 2.0±1.6 mm and 2.8±3.3 mm, while its DSC of 96.84±2.89 is comparable to the competitor's 97.0±2.8.

Load-bearing premise

The claim that the method is comparable to or better than the existing automated system rests on the assumption that the competitor's scores were produced on the same test images with the same metric definitions, which the paper does not explicitly confirm.

Editorial extensions

If this is right

  • Adding a small ellipse-parameter branch to an existing segmentation network can raise Dice by more than four points without changing the segmentation backbone.
  • The network outputs both a pixel mask and an explicit ellipse, so head circumference can be read directly from the ellipse perimeter instead of from boundary-tracing post-processing.
  • On the test set, the boundary errors measured by Hausdorff distance and absolute circumference difference are lower than those of the prior non-deep pipeline, implying the multi-task masks sit closer to the radiologist's border.
  • Because both losses back-propagate through shared features, the segmentation and ellipse tasks are trained jointly with no additional annotation beyond the fitted ground-truth ellipse.

Reading between the lines

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

  • One extension the paper leaves implicit is that the same ellipse-supervision trick could transfer to other roughly elliptical fetal structures, such as the abdominal circumference, where boundary dropout and speckle noise are similar; the five-parameter regression head would need no architectural change.
  • The paper attributes the Dice jump to multi-task learning, but it does not test an equally sized auxiliary branch trained with a different auxiliary loss; a control experiment would tell whether the improvement comes from the ellipse shape prior specifically or from any extra gradient signal.
  • A per-trimester breakdown would be a natural next check, since first-trimester images have poorly visible skulls; if the gain concentrates there, the method is most useful exactly where manual measurement is hardest.
  • Because the competitor's scores are reported as published rather than re-run on the same split (the paper does not state either way), a direct head-to-head reimplementation on the identical test images would settle whether the reported HD and ADF advantages are real or an artifact of different test sets.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 7 minor

Summary. The paper proposes a multi-task deep learning system for fetal head segmentation and head circumference (HC) estimation from 2D ultrasound images. The architecture is based on Link-Net with multi-scale inputs, augmented by a fully connected 'Ellipse Tuner' branch that predicts ellipse parameters from bottleneck features. Training minimizes a compound loss combining segmentation Dice/cross-entropy weighted near boundaries and MSE of ellipse parameters. The authors report experiments on 999 images from the Radboud University Medical Center, with data augmentation to 8823 images, and compare a single-task version without the Ellipse Tuner against the full multi-task network (MTLN). They report DSC improvement from 92.67±2.70 to 96.84±2.89, and compare MTLN with the method of Heuvel et al. in terms of DSC, DF, ADF, and HD. The central claims are that multi-task learning improves segmentation and that MTLN is comparable to or better than the state of the art for fetal HC measurement.

Significance. If the reported results are valid, the proposed MTLN would be a practically useful, relatively simple extension of Link-Net for automated fetal HC measurement, with a clean controlled comparison between single-task and multi-task variants of the same architecture. The paper's strengths include the explicit controlled experiment on the architecture (single-task vs. multi-task), the use of a clinically sourced dataset spanning trimesters, and reporting of multiple clinically relevant metrics (DSC, DF, ADF, HD). However, the manuscript as written leaves two load-bearing issues unresolved: (1) the order of augmentation and train/test splitting is ambiguous and may imply leakage of near-duplicate images into the test set, and (2) the comparison with Heuvel et al. is not demonstrated to be on the same test split and protocol. These issues directly affect the validity of the central improvement and state-of-the-art claims. The paper does not provide code, trained models, or a precise description of the evaluation protocol (e.g., how standard deviations are computed), which limits reproducibility.

major comments (3)
  1. [III-B] The data augmentation and splitting order is ambiguous and load-bearing. The text states, 'For data augmentation, we transform each image ... We generated 8823 augmented US images in total. The data set was randomly split into (75%, 25%) for training and test.' As written, the split is applied after augmentation, which would put flipped/rotated versions of the same fetal head in both training and test sets. Because in-plane rotations and flips are nearly label-preserving for fetal head contours, such near-duplicate test images could substantially inflate the reported DSC (96.84±2.89), HD (1.72±1.39), and the measured improvement over the single-task network (92.67->96.84). Please clarify explicitly whether the random split was performed on the original 999 images before augmentation. If it was not, the experiments must be rerun with an independent test set; the sentence in the Conclusion claiming evaluation on 'an independent test set' would also need to be reconciled with the Methods text.
  2. [Table II] The comparison with Heuvel et al. [10] in Table II is not adequately controlled. The paper reports that MTLN 'outperforms the competitor in terms of HD and ADF' and has comparable DSC, but it does not state whether the scores for [10] were produced by re-running that method on the same 75/25 test split used for MTLN, or whether they were quoted from the original publication, which had a different dataset and evaluation protocol. Differences in test image composition, pixel spacing, and metric implementations (e.g., whether HD is computed on contour pixels or mask boundaries) can change these numbers by more than the reported differences. Please specify the exact evaluation setup, and ideally re-implement or re-run [10] on the identical test split; otherwise the comparative claims are not substantiated.
  3. [III-C] The central claim that multi-task learning improves segmentation is supported only by mean±std values (92.67±2.70 vs. 96.84±2.89) with no statistical significance testing. Given the overlapping standard deviations, a paired or per-image statistical test (e.g., Wilcoxon signed-rank or a paired t-test on per-image DSC) or confidence intervals from repeated splits is needed to establish that the improvement is not due to random variation. Please also state explicitly whether the reported values are per-image mean±std, over how many images, and how the standard deviation is computed.
minor comments (7)
  1. [III-C, Eq. (9)] Equation (9) defines DSC using Area_S and Area_R, but the text immediately below says 'Area_S is the ground truth area and Area_R is the area extracted from the segmentation network.' Earlier in Eq. (4), G is ground truth and S is the network segmentation. Please make the notation consistent to avoid confusion about which area is the reference.
  2. [IV] The Conclusion states the network 'was evaluated on an independent test set,' which is inconsistent with the augmentation-then-split description in Section III-B unless the split was performed on the original images. Please align the wording with the actual protocol.
  3. [III-C] The definition of DF and ADF in Eqs. (10)-(11) uses HC_P and HC_GT, but the method for computing the perimeter from the segmentation result is not described; please specify how the circumference is derived from the segmentation mask or the fitted ellipse.
  4. [II-B] The Ellipse Tuner is trained on ground-truth ellipse parameters that are derived from the same radiologist contours used for the segmentation target. A brief comment on whether this makes the auxiliary task redundant, or what specific benefit the shared features provide, would strengthen the methodological discussion.
  5. [Table I and Section III-A] The paper provides hyperparameter values but no sensitivity analysis; e.g., the loss weights α1 and α2 are fixed at 1 and 2 without justification. A small ablation on these weights would increase confidence in the robustness of the multi-task improvement.
  6. [References] Reference [16] is cited for the boundary-weight map w(x), but that reference is about liver segmentation in CT; please also cite the original U-Net paper where this weighting scheme was introduced, or clarify the provenance.
  7. [General] There are several minor typographical and stylistic issues: 'dice scores' in the Abstract should be 'Dice scores,' the caption of Figure 5 mentions 'red borders in both columns' while the figure appears to have three panels, and the phrase 'DSC score' is redundant. A careful proofreading pass is recommended.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper is an empirical supervised-learning study whose segmentation and ellipse-tuning results are measured against external ground-truth labels, not derived from the method's own assumptions.

full rationale

This paper does not claim a first-principles derivation; it proposes a multi-task Link-Net with an Ellipse Tuner and evaluates it on fetal ultrasound images against radiologist ground truth. The compound loss in Eq. (1) combines a segmentation loss and an ellipse-parameter MSE, and both are trained on the same annotated labels. That is ordinary supervised multi-task training, not circular reasoning: the ellipse parameters are a second output derived from the same ground-truth contours, and the reported DSC, HD, and ADF improvements are measured empirically on a held-out test portion. The only self-citation is reference [16] for the boundary weighting map w(x), which is a standard U-Net-style edge-emphasis term and is not load-bearing for the paper's central claim of improved segmentation. The comparison with Heuvel et al. [10] is an external benchmark, so the headline result does not reduce to the paper's own definitions or fitted parameters. A possible train/test leakage concern from augmentation-before-split is a data-protocol issue, not a circularity issue, and the paper's text leaves the ordering ambiguous rather than demonstrating equivalence of inputs and outputs. Therefore no circular step can be quoted and exhibited, and the appropriate score is 0.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new physical entities. Its free parameters are manually chosen hyperparameters. The main implicit assumptions concern the ellipticity of ground-truth contours, the data split, and the comparability of the external baseline.

free parameters (4)
  • alpha1 = 1
    Weight for the segmentation loss in the compound loss function, chosen by hand.
  • alpha2 = 2
    Weight for the ellipse tuner loss, chosen by hand.
  • w0 = 30
    Predefined constant in the boundary weighting map, set manually.
  • sigma = 10
    Variance of the Gaussian kernel in the boundary weighting map, set manually.
assumptions (3)
  • domain assumption All radiologist ground-truth head contours can be represented as ellipses with five parameters (center, diameters, angle).
    The Ellipse Tuner regresses these parameters from ground-truth masks, and the segmentation loss emphasizes elliptical shapes; Section II-B.
  • domain assumption The 75/25 random split yields an independent test set across patients.
    Section III-B describes a random split but does not specify patient-level stratification, so images from the same patient may appear in both training and test sets.
  • domain assumption The comparison with Heuvel et al. [10] uses the same test data and evaluation protocol.
    Section III-C and Table II report scores for [10] without stating whether they were re-run on the authors' test split or quoted from the original paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fetal Ultrasound Image Segmentation for Measuring Biometric Parameters Using Multi-Task Deep Learning." pith.science (2026). https://pith.science/paper/R5GUY4OF

@misc{pith2026190900273,
  author       = {Pith},
  title        = {Pith review of: Fetal Ultrasound Image Segmentation for Measuring Biometric Parameters Using Multi-Task Deep Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R5GUY4OF}},
  note         = {Machine review of arXiv:1909.00273}
}
read the original abstract

Ultrasound imaging is a standard examination during pregnancy that can be used for measuring specific biometric parameters towards prenatal diagnosis and estimating gestational age. Fetal head circumference (HC) is one of the significant factors to determine the fetus growth and health. In this paper, a multi-task deep convolutional neural network is proposed for automatic segmentation and estimation of HC ellipse by minimizing a compound cost function composed of segmentation dice score and MSE of ellipse parameters. Experimental results on fetus ultrasound dataset in different trimesters of pregnancy show that the segmentation results and the extracted HC match well with the radiologist annotations. The obtained dice scores of the fetal head segmentation and the accuracy of HC evaluations are comparable to the state-of-the-art.

Figures

Figures reproduced from arXiv: 1909.00273 by the authors.

Figure 5
Figure 5. , the segmentation results of our system are comparable to [10], but in some cases provides more reliable results. IV. CONCLUSION In this paper, we presented a Multi-Task deep network based on Link-Net structure with multi-scale inputs, for segmentation and estimation of fetal head circumference in 2D ultrasound images. Our proposed network was trained on 999 images and was evaluated on an independent test set which… view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

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

    eess.IV 2024-11 reject novelty 3.0 of 10

    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.

Reference graph

Works this paper leans on

18 extracted references · 18 canonical work pages · cited by 1 Pith paper

  1. [10]

    Automated measurement of fetal head circumference using 2D ultrasound images,

    T. L. A. van den Heuvel, D. de Bruijn, C. L. de Korte, and B. van Ginneken, “Automated measurement of fetal head circumference using 2D ultrasound images,” PLoS One, vol. 13, no. 8, p. e0200412, 2018

    work page 2018

  2. [1]

    Evaluation and comparison of current fetal ultrasound image segmentation methods for biometric measurements: a grand challenge,

    S. Rueda et al., “Evaluation and comparison of current fetal ultrasound image segmentation methods for biometric measurements: a grand challenge,” IEEE Trans. Med. Imaging , vol. 33, no. 4, pp. 797 –813, 2014

    work page 2014

  3. [2]

    Intelligent estimation of noise and blur variances using ANN for the restoration of ultrasound images,

    M. S. Uddin et al., “Intelligent estimation of noise and blur variances using ANN for the restoration of ultrasound images, ” Appl. Opt., vol. 55, no. 31, pp. 8905–8915, 2016

    work page 2016

  4. [3]

    Fetal size and dating: charts recommended for clinical obstetric practice,

    P. Loughna, L. Chitty, T. Evans, and T. Chudleigh, “Fetal size and dating: charts recommended for clinical obstetric practice, ” Ultrasound, vol. 17, no. 3, pp. 160–166, 2009

    work page 2009

  5. [4]

    Does the use of automated fetal biometry improve clinical workflow efficiency?,

    J. Espinoza, S. Good, E. Russell, and W. Lee, “Does the use of automated fetal biometry improve clinical workflow efficiency?,” J. Ultrasound Med., vol. 32, no. 5, pp. 847–850, 2013

    work page 2013

  6. [5]

    A semi-supervised patch- based approach for segmentation of fetal ultrasound imagin g,

    A. Ciurte, X. Bresson, and M. B. Cuadra, “A semi-supervised patch- based approach for segmentation of fetal ultrasound imagin g,” Proc. Chall. US Biometric Meas. from Fetal Ultrasound Images, ISBI, pp. 5– 7, 2012

    work page 2012

  7. [6]

    A multilevel thresholding combined with edge detection and shape-based recognition for segmentation of fetal ultrasound images,

    G. V Ponomarev, M. S. Gelfand, and M. D. Kazanov, “A multilevel thresholding combined with edge detection and shape-based recognition for segmentation of fetal ultrasound images,” Proc. Chall. US biometric Meas. from fetal ultrasound images, ISBI , pp. 17 –19, 2012

    work page 2012

  8. [7]

    A boundary fragment model for head segmentation in fetal ultrasound,

    R. V Stebbing and J. E. McManigle, “A boundary fragment model for head segmentation in fetal ultrasound, ” Proc. Chall. U S Biometric Meas. from Fetal Ultrasound Images, ISBI, pp. 9–11, 2012

    work page 2012

Show all 18 references
  1. [8]

    Automated Telehealth System for Fetal Growth Detection and Approximation of Ultrasound Images,

    Wisnu Jatmiko, Ikhsanul Habibie, M. Anwar Ma’sum, Robeth Rahmatullah, I Putu Satwika, “Automated Telehealth System for Fetal Growth Detection and Approximation of Ultrasound Images, ” International Journal on Smart Sensing and Intelligent Systems, vol. 8, no. 1, 2015

    2015

  2. [9]

    Automatic Fetal Head Measurements from Ultrasound Images Using Optimal Ellipse Detection and Texture Maps,

    J. L. Perez -Gonzalez, J. C. Bello Muńoz , M. C. Rolon Porras , Fernando Arámbula -Cosío, V. Medina -Bańuelos, “Automatic Fetal Head Measurements from Ultrasound Images Using Optimal Ellipse Detection and Texture Maps,” Springer, vol. 49, p. 329–332, 2015

    2015

  3. [11]

    A survey on deep learning in medical image analysis,

    G. Litjens et al. , “A survey on deep learning in medical image analysis,” Med. Image Anal., vol. 42, pp. 60–88, 2017

    2017

  4. [12]

    Deep learning with ultrasound physics for fetal skull segmentation,

    J. J. Cerrolaza et al., “Deep learning with ultrasound physics for fetal skull segmentation,” in 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018), 2018, pp. 564–567

    2018

  5. [13]

    Cascaded Transforming Multi -task Networks For Abdominal Biometric Estimation from Ultrasound,

    M. D. Sinclair et al., “Cascaded Transforming Multi -task Networks For Abdominal Biometric Estimation from Ultrasound,” 2018

    2018

  6. [14]

    Linknet: Exploiting encoder representations for efficient semantic segmentation,

    A. Chaurasia and E. C ulurciello, “Linknet: Exploiting encoder representations for efficient semantic segmentation, ” in 2017 IEEE Visual Communications and Image Processing (VCIP), 2017, pp. 1–4

    2017

  7. [15]

    D-linknet: Linknet with pretrained encoder and dilated convolution for high resolution satellite imagery road extraction,

    L. Zhou, C. Zhang, and M. Wu, “D-linknet: Linknet with pretrained encoder and dilated convolution for high resolution satellite imagery road extraction,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, 2018, pp. 182–186

    2018

  8. [16]

    Liver Segmentation in CT Images Using Three Dimensional to Two Dimensional Fully Convolutional Network,

    S. Rafiei, E. Nasr-Esfahani, K. Najarian, N. Karimi, S. Samavi, and S. M. R. Soroushmehr, “Liver Segmentation in CT Images Using Three Dimensional to Two Dimensional Fully Convolutional Network, ” in 2018 25th IEEE International Conference on Image Processing (ICIP), 2018, pp....

    2018

  9. [17]

    Vorlesungen ub¨ er Darstellenden Geometrie. Vandenhoeck & Ruprecht,

    K. Strubecker, “Vorlesungen ub¨ er Darstellenden Geometrie. Vandenhoeck & Ruprecht,” in Vandenhoeck & Ruprecht, 1967. Figure 5. Segmentation results of fetal head (ground truth is red) From left to right: 1) Original image, 2) Green: results of [10], 3) Blue: MTLN results. Fig...

    1967

  10. [18]

    Original image, 2) Result without FC, 3) MTLN Segmentation result

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.