DSSAU-Net:U-Shaped Hybrid Network for Pubic Symphysis and Fetal Head Segmentation

Hongxing Li; Libin Lan; Zunhui Xia

REVIEW 3 major objections 5 minor 18 references

DSSAU-Net:U-Shaped Hybrid Network for Pubic Symphysis and Fetal Head Segmentation

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Dual sparse selection attention segments fetal head and pubic symphysis at 86.34 DSC in the IUGC 2024 challenge.

desk verdict A clearly described sparse-attention U-Net variant with a plausible competition result, but the evidence is thin: 40-image validation, no error bars, and a suspicious table entry. read the letter →

arxiv 2506.03684 v1 pith:SW4ITSHK submitted 2025-06-04 cs.CV

Zunhui Xia , Hongxing Li , Libin Lan This is my paper

classification cs.CV

keywords fetalheadsegmentationpubicsymphysissparseself-attentionDSSAU-shapedencoder-decoderintrapartumultrasoundmultiscalefeaturefusionAngleofProgression

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 sets out to show that a U-shaped hybrid network named DSSAU-Net can segment the fetal head and pubic symphysis in intrapartum ultrasound images accurately and efficiently. Its attention module, Dual Sparse Selection Attention (DSSA), prunes tokens twice—first at the region level and then at the pixel level—so the network keeps the most relevant features while spending far less compute than dense self-attention. On the 40-image IUGC 2024 validation set the authors report a Dice Similarity Coefficient of 86.34%, a Hausdorff Distance of 31.08 mm, and an Average Surface Distance of 8.39 mm, with 29.25 million parameters and 7.15 GFLOPs. On the challenge test set the method placed fourth overall and second on HD and ASD for the segmentation task. The paper argues this makes the network a practical tool for deriving the labor-progress parameters AoP and HSD from ultrasound without invasive exams.

What carries the argument

DSSA is a sparse self-attention module that, for each query, first selects the $k_1$ most relevant image regions via a top-$k$ operation on a region-level attention matrix, then selects the $k_2$ most relevant pixels within those regions via a second top-$k$ on a pixel-level attention matrix, with $k_2 = \lambda \cdot k_1 \cdot \frac{HW}{S^2}$. The selected pixel values are weighted by the normalized attention scores and summed with a 5$\times$5 depth-wise convolution term for local context. This module is stacked in varying numbers (2, 2, 8, 2 in the encoder and again in the decoder) inside a symmetric U-shaped network whose decoder features are fused at multiple scales, with convolutional skip connections at $1/4$, $1/8$, and $1/16$ resolutions carrying encoder detail into the decoder.

What would settle it

An independent evaluation on a larger held-out cohort—say, several hundred intrapartum ultrasound images—that fails to reproduce DSSAU-Net's reported DSC near 86.34, HD near 31.08, and ASD near 8.39, or that reverses the ablation ordering (no-skip, or other $k_1$/ $\lambda$ settings, beating the chosen ones), would refute the central claim.

Watch

Extended reading notes

Core claim

The central claim is that performing sparse token selection twice, at both region and pixel granularity, is enough to make a U-shaped transformer-style segmentation network both accurate and computationally light for fetal-head and pubic-symphysis segmentation. DSSA first scores the relevance of $S \times S$ image regions and keeps the top $k_1$ for each query; it then scores individual pixels inside those kept regions and keeps the top $k_2$, where $k_2$ scales as $\lambda$ times the number of pixels in a region. This two-stage filtering removes both irrelevant regions and speckle-like noise tokens, and the remaining tokens are used to compute the attention output with a local-context term added. The paper reports that this design, combined with convolutional skip connections at three resolutions, a pyramid pooling module, and multiscale decoder fusion, reaches DSC of 86.34%, HD of 31.08 mm, and ASD of 8.39 mm on the validation set, and places fourth overall (second on HD and ASD, third on DSC) in the IUGC 2024 segmentation ranking.

Load-bearing premise

The 40-image validation set provided by the challenge organizers is representative enough of the test and clinical distributions that the reported metric differences and the hyperparameter choices ($k_1$, $\lambda$) made on it will transfer.

Editorial extensions

If this is right

A clinician-facing ultrasound tool could compute AoP and HSD automatically from DSSAU-Net's segmentations, giving objective labor-progress measures without invasive vaginal examination.
DSSAU-Net's compute footprint (7.15 GFLOPs at 256×256 input) is low enough for real-time or near-real-time deployment in delivery-room ultrasound systems.
Each additional skip connection improves all three segmentation metrics on the validation set, so the full three-connection design is the recommended configuration.
The multiscale feature fusion module raises DSC by 0.37, lowers HD by 1.2 mm, and lowers ASD by 0.81 mm relative to the no-MFF variant.
The dual sparse selection mechanism keeps segmentation accurate at low FLOPs, making the architecture a candidate for settings where dense transformers are too heavy.

Reading between the lines

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

The two-stage sparse selection recipe is not specific to obstetrics; the same region-then-pixel top-k attention could be transplanted into other U-shaped segmentation networks for small structures in noisy backgrounds, though the paper does not test this.
Because the k1 and lambda hyperparameters were tuned on the same 40-image validation set used to report the headline metrics, an independent larger cohort is needed to confirm the gains are not an artifact of that small sample.
The measured FLOPs vary only slightly with lambda (7.17 to 7.14 GFLOPs in the reported settings), suggesting the architecture could be swept into an accuracy-latency trade-off family; the paper does not explore that frontier.

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.

Desk Editor's Note

read the letter

You should know this paper is a solid engineering write-up of a U-Net variant that swaps in a dual sparse selection attention (DSSA) derived from BiFormer's bi-level routing attention. The authors are transparent about that derivation and about using the mechanism in their own unpublished work. The contribution is modest but real: explicit top-k selection at both region and pixel levels, packaged cleanly into a symmetric encoder-decoder with skip connections and a pyramid pooling module for multiscale fusion. The equations are consistent, the architecture is implementable from the text, and the efficiency numbers (29.25M params, 7.15 GFLOPs) are genuinely good. What the paper does well: it is written clearly, the method section is complete enough to reimplement, and the challenge result (4th overall, 2nd on HD and ASD in segmentation) is credible on its face. The authors also do not oversell the classification performance, admitting it is suboptimal. I credit that honesty. The soft spots are mostly empirical, and they are not minor. The headline results rest on a 40-image validation set. No standard deviations, confidence intervals, or significance tests are reported. The ablation hyperparameters (k1, lambda) are tuned on this same small set, so the reported gains, like MFF's +0.37 DSC, could easily be noise. Table 5 contains an apparent error: the last row reports ASD=29.25, which is exactly the Params(M) value from Table 1. And the abstract's DSC (86.43) disagrees with Table 1 (86.34) by 0.09. These are concrete, fixable problems, not fatal ones. The architecture itself is not circular; the attention is defined independently of the results, and the external test set ranking provides a sanity check. If I were the editor, I would send this to review rather than desk reject. The architecture is plausible, the writing is competent, and the clinical task is relevant. But the reviewers should ask for public code, per-fold or bootstrapped confidence intervals, a fix to Table 5, and a careful reconciliation of the DSC numbers. The central claim—accurate and efficient segmentation—may well hold up, but right now the evidence is only suggestive.

Referee Report

3 major / 5 minor

Summary. The paper proposes DSSAU-Net, a U-shaped hybrid CNN-transformer for segmenting the pubic symphysis and fetal head in intrapartum ultrasound images. The core attention module, DSSA, performs two rounds of sparse token selection—a region-level top-k selection followed by a pixel-level top-k selection—with a local context enhancement term (LCE) and an MLP block. The encoder-decoder uses skip connections with convolutions and a pyramid pooling module for multiscale feature fusion, trained with a hybrid dice and cross-entropy loss. Experiments on the IUGC 2024 challenge validation set (40 images) report DSC 86.34, HD 31.08 mm, and ASD 8.39 mm with 29.25 M parameters and 7.15 G FLOPs. On the challenge test set, the team ranked fourth overall and second on ASD and HD among the listed groups. Ablation studies examine the number of skip connections, the multiscale feature fusion module, and the top-k hyperparameters k1 and λ. The code is stated to be available, but no link is given in the manuscript.

Significance. If the quantitative claims hold, the paper provides an efficient and reasonably accurate architecture for an important clinical task, and the challenge test-set ranking is a genuine external benchmark. The DSSA equations are explicit enough that the module can be reimplemented, and the design choices (region-level then pixel-level sparse selection) are clearly motivated. However, the central empirical claim is supported by point estimates on a single 40-image validation set without uncertainty quantification, and the key hyperparameters are selected on that same set; the load-bearing evidence for the architecture's superiority therefore needs strengthening before the result can be considered established.

major comments (3)

[§3.1, §3.4, §3.5 (Tables 1, 3–5)] The paper's central empirical claim is supported only by point estimates on the 40-image validation set described in §3.1. No standard deviations, confidence intervals, or significance tests are reported for any metric in Tables 1, 3, 4, or 5, and the hyperparameters k1 and λ (Table 5) as well as the architecture components (Tables 3–4) are selected on this same small set. For an n=40 ultrasound dataset with substantial patient-level variability, the MFF ablation gain of +0.37 DSC (85.97→86.34) and the corresponding HD/ASD differences are within plausible resampling noise, so the reported superiority of the full model is not yet established. The independent evidence in Table 2 is a self-reported leaderboard entry without error bars or released code; it is encouraging but does not replace a statistical comparison. Please report per-image distributions or bootstrap confidence intervals and, if possible, multi-seed training results.
[§3.5, Table 5] The last row of Table 5 (k1=[2,8,32,64], λ=1/16) lists ASD=29.25, which is exactly the Params(M) value from Table 1. This appears to be a data-entry error that invalidates the printed row. Since this table is used to justify the chosen λ and k1 values, the correct ASD value should be recomputed and reported.
[§2.1, Eq. (9)] The stated shape of Vgg, R^{S^2 × k1 HW/S^2 × k2 × C}, is inconsistent with the preceding gather operation: the indices IP have shape S^2 × HW/S^2 × k2, so gathering k2 value tokens for each of the HW/S^2 queries should produce a tensor of shape S^2 × HW/S^2 × k2 × C, not S^2 × k1 HW/S^2 × k2 × C. Please correct this shape so that the algorithm can be implemented as written.

minor comments (5)

[§1] The sentence stating that DSSA 'has been used in our other unpublished work' is a missing reference; please either cite the prior work or explicitly state that this manuscript is the first public disclosure of DSSA.
[Abstract and §3.4] The abstract reports a DSC of 86.43, while Table 1 and §3.4 report 86.34; these values should be reconciled.
[§2.2, Eq. (11)] Equation (11) has unbalanced parentheses and duplicates the term 'DWConv(z_{l-1})' inside and outside the layer norm; the expression should be rewritten to match the described three-residual-block structure.
[Fig. 3] The figure panel label 'DSSA-UNet' should be 'DSSAU-Net' to match the model name used throughout the text.
[§3.5] The hyperparameter ablation explicitly does not use pre-trained weights, so Table 5 is not directly comparable with Tables 1, 3, and 4; this should be stated in the table caption or in the main text.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DSSAU-Net is validated against the external IUGC 2024 test set, and its DSSA mechanism is defined independently of the reported metrics.

full rationale

The paper's central claim is that the proposed DSSA-based U-Net achieves accurate and efficient fetal-head and pubic-symphysis segmentation. The architecture is defined by explicit equations (Eqs. 1-22) for region-level and pixel-level sparse token selection, and these definitions do not contain the target metrics (DSC, HD, ASD) or the test-set ranking, so there is no self-definitional reduction. The headline numbers in Table 1 are point estimates on the 40-image organizer validation set, but they are measurements, not quantities derived by construction from the architecture's definition. Hyperparameters k1 and lambda are selected on that same validation set (Section 3.5, Table 5), which is a statistical overfitting risk, not circularity: the validation metrics are not fitted parameters renamed as predictions, and the test-set result (Table 2, fourth place) comes from the external organizer leaderboard. Reference [3] (BRAU-Net) is prior work by co-author Libin Lan, but it is cited as a comparative baseline, not as the justification for DSSA's design or performance; the DSSA mechanism is instead attributed to external works [16,18] and to the paper's own equations. No uniqueness theorem is imported from the authors' prior work, and no known empirical result is merely renamed. The apparent data anomaly in Table 5 (ASD=29.25 duplicating Params(M)=29.25) is a correctness and data-integrity concern, not evidence of circularity.

Assumptions & free parameters 7 free parameters · 3 assumptions · 0 invented entities

The central claim is empirical, not a derivation. The architecture relies on standard attention math and standard deep learning training practices. The main freely chosen quantities are the architectural hyperparameters (number of regions, channels, block counts) and the attention sparsity parameters (k1, lambda), all set by hand or by validation-set ablation. The dataset annotations are taken as ground truth. No new physical or mathematical entities are introduced.

free parameters (7)

k1 (region-level top-k) = [1, 4, 16, 64]
Number of regions selected in each of the four encoder stages; chosen via ablation on the validation set (Table 5).
lambda (pixel-level scaling factor) = 1/8
Determines k2, the number of selected pixel-level tokens; chosen via ablation on the validation set (Table 5).
Number of regions S = 8
The image is divided into 8x8 non-overlapped regions in each DSSA stage (Section 3.3, Eq. 1).
Encoder channel widths and decoder dimension = [96, 192, 384, 768], Cd=64
Channel configuration for the four encoder stages and the fixed decoder feature dimension (Section 3.3).
Initial learning rate = 1e-4
Set in Section 3.3; no schedule details provided.
Loss weights = 0.5 dice, 0.5 CE
Equal weighting of dice and cross-entropy losses (Eq. 25).
ImageNet pretraining = ImageNet weights
Backbone initialized with ImageNet-pretrained weights (Section 3.3).

assumptions (3)

standard math Standard linear algebra and softmax attention definitions
Matrix multiplication, top-k selection, and softmax normalization as defined in Eqs. 1-10.
domain assumption IUGC dataset ground-truth annotations are accurate and sufficient for evaluating AoP/HSD
The paper relies on the challenge dataset's annotations as ground truth for training and evaluation (Section 3.1).
domain assumption The 40-image validation set is representative of the test distribution
All ablations and hyperparameter choices are based on a single 40-image validation split (Sections 3.1, 3.4, 3.5).

how reviews work

0 comments

Cite this review

Pith. "Pith review of DSSAU-Net:U-Shaped Hybrid Network for Pubic Symphysis and Fetal Head Segmentation." pith.science (2026). https://pith.science/paper/SW4ITSHK

@misc{pith2026250603684,
  author       = {Pith},
  title        = {Pith review of: DSSAU-Net:U-Shaped Hybrid Network for Pubic Symphysis and Fetal Head Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SW4ITSHK}},
  note         = {Machine review of arXiv:2506.03684}
}

read the original abstract

In the childbirth process, traditional methods involve invasive vaginal examinations, but research has shown that these methods are both subjective and inaccurate. Ultrasound-assisted diagnosis offers an objective yet effective way to assess fetal head position via two key parameters: Angle of Progression (AoP) and Head-Symphysis Distance (HSD), calculated by segmenting the fetal head (FH) and pubic symphysis (PS), which aids clinicians in ensuring a smooth delivery process. Therefore, accurate segmentation of FH and PS is crucial. In this work, we propose a sparse self-attention network architecture with good performance and high computational efficiency, named DSSAU-Net, for the segmentation of FH and PS. Specifically, we stack varying numbers of Dual Sparse Selection Attention (DSSA) blocks at each stage to form a symmetric U-shaped encoder-decoder network architecture. For a given query, DSSA is designed to explicitly perform one sparse token selection at both the region and pixel levels, respectively, which is beneficial for further reducing computational complexity while extracting the most relevant features. To compensate for the information loss during the upsampling process, skip connections with convolutions are designed. Additionally, multiscale feature fusion is employed to enrich the model's global and local information. The performance of DSSAU-Net has been validated using the Intrapartum Ultrasound Grand Challenge (IUGC) 2024 \textit{test set} provided by the organizer in the MICCAI IUGC 2024 competition\footnote{\href{https://codalab.lisn.upsaclay.fr/competitions/18413\#learn\_the\_details}{https://codalab.lisn.upsaclay.fr/competitions/18413\#learn\_the\_details}}, where we win the fourth place on the tasks of classification and segmentation, demonstrating its effectiveness. The codes will be available at https://github.com/XiaZunhui/DSSAU-Net.

Figures

Figures reproduced from arXiv: 2506.03684 by the authors.

**Figure 1.** Illustration of dual sparse selection attention (DSSA). DSSA performs twice sparse token selections at both the region and pixel levels, which can reduce computational complexity while extracting the most relevant features. 2.1 Dual Sparse Selection Attention (DSSA) The core idea of DSSA is to explicitly perform sparse token selections at both region and pixel levels. It involves three main steps. First, for a give… view at source ↗

**Figure 2.** (a): The overall architecture of DSSAU-Net, which is a u-shaped hybrid network and uses a sparse attention mechanism: dual sparse selection attention (DSSA) as a core building idea to hierarchically design the encoder-decoder structure. Furthermore, a pyramid pooling module (PPM) is used to fuse multiscale features, which is beneficial to improve the segmentation performance. (b): The details of the DSSA block. afte… view at source ↗

**Figure 3.** The visualization results of DSSAU-Net on the Intrapartum Ultrasound Grand Challenge 2024 validation set. The red and green indicate the segmentation results of PS and FH, respectively. 3.5 Ablation Study In this section, to deeply explore the specific impact of each component of the DSSAU-Net on overall performance, we follow the above experimental settings and conduct a series of ablation studies on the provided t… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

18 extracted references · 14 canonical work pages

[1]

Fetal Diagnosis and Therapy42(4), 249–256 (2017)

Bellussi, F., Ghi, T., Youssef, A., Cataneo, I., Salsi, G., Simonazzi, G., Pilu, G.: Intrapartum ultrasound to differentiate flexion and deflexion in occipitoposterior rotation. Fetal Diagnosis and Therapy42(4), 249–256 (2017)

work page 2017
[2]

Obstetrics & Gynecology122(1), 33–40 (2013)

Boyle, A., Reddy, U.M., Landy, H.J., Huang, C.C., Driggers, R.W., Laughon, S.K.: Primary cesarean delivery in the united states. Obstetrics & Gynecology122(1), 33–40 (2013)

work page 2013
[3]

Cai, P., Lu, J., Li, Y., Lan, L.: Pubic symphysis-fetal head segmentation using pure transformer with bi-level routing attention (2023),https://arxiv.org/abs/ 2310.00289

work page Pith review arXiv 2023
[4]

In: European conference on computer vision

Cao, H., Wang, Y., Chen, J., Jiang, D., Zhang, X., Tian, Q., Wang, M.: Swin- unet: Unet-like pure transformer for medical image segmentation. In: European conference on computer vision. pp. 205–218. Springer (2022)

2022
[5]

IEEE Journal of Biomedical and Health Infor- matics (2024)

Chen, Z., Lu, Y., Long, S., Campello, V.M., Bai, J., Lekadir, K.: Fetal head and pubic symphysis segmentation in intrapartum ultrasound image using a dual-path boundary-guided residual network. IEEE Journal of Biomedical and Health Infor- matics (2024)

work page 2024
[6]

Expert Systems with Applications245, 123096 (2024)

Chen, Z., Ou, Z., Lu, Y., Bai, J.: Direction-guided and multi-scale feature screening for fetal head–pubic symphysis segmentation and angle of progression calculation. Expert Systems with Applications245, 123096 (2024)

work page 2024
[7]

Cohen, S., Lipschuetz, M., Yagel, S.: Is a prolonged second stage of labor too long? Ultrasound in Obstetrics & Gynecology50(4), 423–426 (2017)

work page 2017
[8]

European Journal of Obstetrics & Gynecology and Reproductive Biology123(2), 193–197 (2005)

Dupuis, O., Ruimark, S., Corinne, D., Simone, T., André, D., René-Charles, R.: Fetal head position during the second stage of labor: comparison of digital vaginal examination and transabdominal ultrasonographic examination. European Journal of Obstetrics & Gynecology and Reproductive Biology123(2), 193–197 (2005)

work page 2005

Show all 18 references

[9]

Medical image analysis83, 102629 (2023)

Fiorentino, M.C., Villani, F.P., Di Cosmo, M., Frontoni, E., Moccia, S.: A review on deep-learning algorithms for fetal ultrasound-image analysis. Medical image analysis83, 102629 (2023)

2023
[10]

Ultrasound in Obstetrics & Gynecology52(1), 128–139 (2018)

Ghi, T., Eggebø, T., Lees, C., Kalache, K., Rozenberg, P., Youssef, A., Salomon, L., Tutschek, B.: Isuog practice guidelines: intrapartum ultrasound. Ultrasound in Obstetrics & Gynecology52(1), 128–139 (2018)

2018
[11]

The Journal of Maternal-Fetal & Neonatal Medicine27(5), 520–526 (2014)

Malvasi, A., Tinelli, A., Barbera, A., Eggebø, T., Mynbaev, O., Bochicchio, M., Pacella,E.,DiRenzo,G.:Occiputposteriorpositiondiagnosis:vaginalexamination or intrapartum sonography? a clinical review. The Journal of Maternal-Fetal & Neonatal Medicine27(5), 520–526 (2014)

2014
[12]

BJOG: An International Journal of Obstetrics & Gynaecology121(8), 1029–1038 (2014) 14 Z

Ramphul, M., Ooi, P.V., Burke, G., Kennelly, M.M., Said, S.A., Montgomery, A.A., Murphy, D.J.: Instrumental delivery and ultrasound: a multicentre randomised controlled trial of ultrasound assessment of the fetal head position versus standard care as an approach to prevent mor...

2014
[13]

In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, Oc- tober 5-9, 2015, proceedings, part III 18

Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, Oc- tober 5-9, 2015, proceedings, part III 18....

2015
[14]

IEEE Transactions on Medical Imaging33(4), 797–813 (2014)

Rueda, S., Fathima, S., Knight, C.L., et al.: Evaluation and comparison of current fetalultrasoundimagesegmentationmethodsforbiometricmeasurements:Agrand challenge. IEEE Transactions on Medical Imaging33(4), 797–813 (2014)

2014
[15]

In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y

Xiao, T., Liu, Y., Zhou, B., Jiang, Y., Sun, J.: Unified perceptual parsing for scene understanding. In: Ferrari, V., Hebert, M., Sminchisescu, C., Weiss, Y. (eds.) Computer Vision – ECCV 2018. pp. 432–448. Springer International Publishing, Cham (2018)

2018
[16]

org/abs/1912.11637

Zhao, G., Lin, J., Zhang, Z., Ren, X., Su, Q., Sun, X.: Explicit sparse trans- former: Concentrated attention through explicit selection (2019),https://arxiv. org/abs/1912.11637

2019 arXiv
[17]

In: Proceedings of the IEEE conference on computer vision and pattern recognition

Zhao, H., Shi, J., Qi, X., Wang, X., Jia, J.: Pyramid scene parsing network. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 2881–2890 (2017)

2017
[18]

In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

Zhu, L., Wang, X., Ke, Z., Zhang, W., Lau, R.: Biformer: Vision transformer with bi-level routing attention. In: 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 10323–10333 (2023)

2023

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Fetal Diagnosis and Therapy42(4), 249–256 (2017)

Bellussi, F., Ghi, T., Youssef, A., Cataneo, I., Salsi, G., Simonazzi, G., Pilu, G.: Intrapartum ultrasound to differentiate flexion and deflexion in occipitoposterior rotation. Fetal Diagnosis and Therapy42(4), 249–256 (2017)

work page 2017

[2] [2]

Obstetrics & Gynecology122(1), 33–40 (2013)

Boyle, A., Reddy, U.M., Landy, H.J., Huang, C.C., Driggers, R.W., Laughon, S.K.: Primary cesarean delivery in the united states. Obstetrics & Gynecology122(1), 33–40 (2013)

work page 2013

[3] [3]

Cai, P., Lu, J., Li, Y., Lan, L.: Pubic symphysis-fetal head segmentation using pure transformer with bi-level routing attention (2023),https://arxiv.org/abs/ 2310.00289

work page Pith review arXiv 2023

[4] [4]

In: European conference on computer vision

Cao, H., Wang, Y., Chen, J., Jiang, D., Zhang, X., Tian, Q., Wang, M.: Swin- unet: Unet-like pure transformer for medical image segmentation. In: European conference on computer vision. pp. 205–218. Springer (2022)

2022

[5] [5]

IEEE Journal of Biomedical and Health Infor- matics (2024)

Chen, Z., Lu, Y., Long, S., Campello, V.M., Bai, J., Lekadir, K.: Fetal head and pubic symphysis segmentation in intrapartum ultrasound image using a dual-path boundary-guided residual network. IEEE Journal of Biomedical and Health Infor- matics (2024)

work page 2024

[6] [6]

Expert Systems with Applications245, 123096 (2024)

Chen, Z., Ou, Z., Lu, Y., Bai, J.: Direction-guided and multi-scale feature screening for fetal head–pubic symphysis segmentation and angle of progression calculation. Expert Systems with Applications245, 123096 (2024)

work page 2024

[7] [7]

Cohen, S., Lipschuetz, M., Yagel, S.: Is a prolonged second stage of labor too long? Ultrasound in Obstetrics & Gynecology50(4), 423–426 (2017)

work page 2017

[8] [8]

European Journal of Obstetrics & Gynecology and Reproductive Biology123(2), 193–197 (2005)

Dupuis, O., Ruimark, S., Corinne, D., Simone, T., André, D., René-Charles, R.: Fetal head position during the second stage of labor: comparison of digital vaginal examination and transabdominal ultrasonographic examination. European Journal of Obstetrics & Gynecology and Reproductive Biology123(2), 193–197 (2005)

work page 2005