Pith. sign in

REVIEW 4 major objections 6 minor 20 references

USEANet: Ultrasound-Specific Edge-Aware Multi-Branch Network for Lightweight Medical Image Segmentation

T0 review · 4 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A 3.64M-parameter ultrasound-specific edge-aware network beats nine baselines on five ultrasound datasets, reporting 67.01 IoU on BUSI at 0.79G FLOPs.

desk verdict The architecture has merit, but the paper's central claim of SOTA performance is contradicted by its own Table 1. read the letter →

arxiv 2509.08860 v1 pith:5L34WHHG submitted 2025-09-09 eess.IV

Jingyi Gao , Di Wu , Baha lhnaini This is my paper · ORCID
classification eess.IV
keywords ultrasoundimagesegmentationedge-awareattentionmulti-branchnetworklightweightneuralmedicalanalysisPVT-B0backboneBUSIdataset
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 USEANet, a compact encoder-decoder network for ultrasound image segmentation, and argues that a network designed around ultrasound's own failure modes—speckle noise, low contrast, ambiguous boundaries—can beat both heavy transformers and extremely small generic models. If the claims hold, accurate segmentation becomes feasible on portable ultrasound machines without sacrificing accuracy. The evidence is a five-dataset comparison against nine baselines plus ablations on BUSI attributing each component's contribution. The headline result is 67.01 IoU (Intersection over Union) on BUSI at 3.64M parameters and 0.79G FLOPs.

What carries the argument

The load-bearing object is the Multi-Branch Feature Processor (MBFP), which runs three parallel streams on each encoder stage: a noise-reduction branch built from residual depthwise-separable convolutions, an edge-enhancement branch that adds a scaled Laplacian-filtered response, and a contrast-enhancement branch that applies channel attention to a global-pooled feature. Around it sit Edge-Aware Attention (EAA), which gates features with gradient information; Hierarchical Feature Aggregation (HFA), which multiplies multi-scale features from coarser levels into finer ones; and Ultrasound-Aware Decoder Enhancement (UADE), which uses the foreground-minus-background prediction as a multiplicativ

What would settle it

Re-run all five datasets with a patient-level 70:15:15 split, ensuring every image of a patient stays in one partition, and compare BUSI IoU against the same nine baselines. If the reported margin collapses or reverses, the performance claim is an artifact of leakage; if the margin persists, the central claim survives.

Watch

Extended reading notes

Core claim

The central claim is that a lightweight ultrasound-specific design can close the gap to much larger models while beating them on ultrasound-specific metrics. USEANet reports the best mIoU/F1 on all five datasets—BUSI, BUS-BRA, Breast-Lesions-USG, DDTI, TN3K—with 3.64M parameters and 0.79G FLOPs. Its BUSI IoU of 67.01 sits 8.69 points above U-Net and above transformer baselines such as SwinUnet and XboundFormer at a fraction of their compute. Ablations on BUSI attribute 2.79 IoU points to the attention mechanisms, 2.42 to the multi-branch architecture, 2.85 to ultrasound-specific modules, and 2.99 to multi-scale aggregation.

Load-bearing premise

The evaluation assumes the random 70:15:15 split is patient-level—no image from the same patient appears in more than one partition—but the paper never states this; if the split is image-level, the reported IoU gains could be inflated.

Editorial extensions

If this is right

  • Ultrasound segmentation no longer requires a heavyweight model: transformer-scale accuracy is reported at 7.5x fewer parameters than SwinUnet and 8.1x fewer than XboundFormer.
  • Extremely small skin-lesion models (0.05M-0.33M parameters) fall far behind on ultrasound, with EGEUNet, MALUNet, and PMFSNet trailing USEANet by roughly 21-25 IoU points on BUSI.
  • Each of the four modules contributes independently, with multi-scale aggregation providing the largest single gain (2.99 IoU on BUSI).
  • The 0.79G FLOPs footprint puts real-time deployment on resource-constrained hardware within reach.

Reading between the lines

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

  • The split is described only as random 70:15:15; if it is image-level rather than patient-level, the reported margins could shrink under patient-stratified evaluation. A re-run with patient-level splits would test the generalisation claim.
  • The four modules are presented as portable: MBFP, EAA, HFA, and UADE could be grafted onto another lightweight encoder, though the paper only validates them in combination.
  • The observed frontier implies an optimal operating point around 3-4M parameters for ultrasound: 0.05M models lack capacity, while 25-35M transformers buy little extra accuracy.
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

4 major / 6 minor

Summary. The manuscript proposes USEANet, a lightweight encoder-decoder architecture for ultrasound image segmentation built on a PVT-B0 backbone (3.64M parameters, 0.79G FLOPs). The design includes a multi-branch feature processor for noise reduction, edge enhancement, and contrast enhancement; edge-aware attention mechanisms; hierarchical feature aggregation; and an ultrasound-aware decoder enhancement module. The method is evaluated on five public ultrasound datasets (BUSI, BUS-BRA, Breast-Lesions-USG, DDTI, TN3K) and compared with nine baselines. The abstract and Section 3.3 claim that USEANet 'significantly outperforms existing methods across five ultrasound datasets' and 'achieves superior performance across all five datasets.'

Significance. If the central performance claim were correct, the paper would offer a useful contribution to efficient ultrasound segmentation: the parameter and FLOP counts are genuinely low, the code is promised to be public, and the modular architecture is clearly described. The ablation study, although limited, does suggest each module contributes positively on BUSI. However, the primary empirical claim is contradicted by the paper's own Table 1, and no statistical evidence is provided for any 'significant' improvement. The efficiency story is plausible, but the accuracy-superiority claim is not supported as presented.

major comments (4)
  1. [Abstract and §3.3, Table 1] The central claim that USEANet 'achieves superior performance across all five datasets' is directly contradicted by Table 1. XboundFormer reports higher IoU on BUS-BRA (81.56 vs 80.63), DDTI (83.29 vs 81.72), and TN3K (75.83 vs 74.81). USEANet is the best only on BUSI and Breast-Lesions-USG. Even the average IoU across the five datasets is slightly in favor of XboundFormer (~74.94 vs ~74.90). The abstract's 'significantly outperforms existing methods across five ultrasound datasets' is therefore not supported by the presented numbers.
  2. [§3.2, Evaluation Setup] All results are reported as point estimates from single runs, with no error bars, confidence intervals, or statistical significance tests. The word 'significantly' appears in the abstract and throughout §3.3, but no test (e.g., paired t-test or Wilcoxon) is reported. Without variance estimates or repeated runs, the improvements over UNet/AttUNet on BUSI and the small margins in other datasets cannot be evaluated. This is a load-bearing deficiency for an empirical paper that makes a superiority claim.
  3. [§3.1, Datasets] The 70:15:15 random split is not specified as patient-level. Public ultrasound datasets such as BUSI, BUS-BRA, DDTI, and TN3K commonly contain multiple images per patient. If the split is at image level, same-patient images can appear in both training and test sets, inflating the reported IoU values and making comparisons with literature results unreliable. The authors should clarify the split level and, if necessary, re-run experiments with a patient-level split.
  4. [§3.4, Table 2] The ablation study is limited to a single dataset (BUSI) and a single run, with no variance reporting. In addition, 'w/o Ultrasound-Specific' removes the multi-branch module that contains three branches at once, so the individual contributions of noise reduction, edge enhancement, and contrast enhancement are not isolated. The stated gains (2.85 IoU for 'ultrasound-specific modules') may be dominated by parameter count or training dynamics. This is secondary to the main claim, but it weakens the paper's component-wise validation.
minor comments (6)
  1. [Title and author affiliations] The title renders 'Edge-Aware' as 'Edge-A W are' and the author name 'Baha lhnaini' appears to contain a typo. These should be corrected.
  2. [§3.2 and Table 1] The metric is called 'mIoU' in the evaluation setup but 'IoU' in Table 1 and elsewhere. The naming should be consistent.
  3. [§2.2, Eq. (2)] The dimensions of W_i in the adaptive fusion are not specified. Since W = Softmax(FC(GAP(Fcat))), it is unclear how the scalar/vector weights are broadcast over the feature maps F^i_branch. Please clarify the tensor shapes.
  4. [§2.3 and §2.4] The hyperparameters alpha and beta are reused with different meanings (edge enhancement α=0.1 in Eq. (2), EAA β=0.1 in Eq. (4), and HFA α=0.5 in Eq. (5)). This is confusing; use distinct symbols or define each locally.
  5. [References] Reference [12] lists the author as 'Pawłwska', which appears to be a typo for 'Pawłowska'. Please verify all author names and dataset citations.
  6. [Conclusion] The conclusion states 'Experimental results demonstrate superior segmentation performance,' but given the Table 1 results, a more measured statement such as 'competitive performance with strong efficiency' would be accurate. The paper would benefit from a limitations paragraph.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: predictions are trained test-set measurements; no self-citation chain.

full rationale

This is an empirical architecture-comparison paper. The claimed results (IoU/F1/Acc in Table 1) are produced by training on 70% splits and measuring on 15% test splits; they are not derived from the method's definitions. Each component (MBFP, ECA/EAA, HFA, UADE) is defined by explicit equations (1)–(7) using trainable parameters and standard convolution/attention operations, and the final loss is a weighted BCE+IoU supervision on ground-truth masks (Sec. 2.6). There is no fitted input that is later renamed a prediction, no self-citation chain (the authors do not cite any prior work of their own), and no uniqueness theorem imported from the authors. The ablation study on BUSI (Table 2) is a standard contribution analysis; it does not make the method circular because removing components changes the trained model and the resulting differences are empirical measurements, not consequences of how the components were defined. The main weaknesses are evidentiary rather than circular: Table 1 does not support 'superior across all five datasets' since XboundFormer has higher IoU on BUS-BRA, DDTI, and TN3K, and the random split is not stated to be patient-level. Those are correctness/validity concerns, not circularity. Score 0.

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

The network's claimed performance depends on a handful of hand-set scaling constants (alpha, beta, loss weights) and on the choice of lightweight building blocks. These are common hyperparameters in empirical network design, but their values are not swept or justified, so the central result is conditional on them.

free parameters (4)
  • MBFP edge scale alpha = 0.1
    Fixed constant in Eq. (2) scaling Laplacian edge branch; chosen by hand, not tuned. Any value could alter the balance.
  • EAA edge scale beta = 0.1
    Constant in Eq. (4) controlling gradient enhancement; chosen by hand.
  • HFA fusion strength alpha = 0.5
    Constant in Eq. (5) controlling multi-scale fusion; chosen by hand.
  • Loss weights lambda_s = not specified
    Multi-scale loss weights in Section 2.6 are not reported; they affect supervision balance.
assumptions (4)
  • domain assumption Depthwise separable convolution is sufficient to replace standard convolution without loss of representational capacity.
    Used throughout via LConv (Eq. 1); if false, efficiency claims would not hold at stated accuracy.
  • domain assumption Adding Laplacian-based gradient features improves boundary segmentation rather than introducing high-frequency noise.
    Central to MBFP and EAA; ablation on BUSI supports it, but only on one dataset with no statistical test.
  • domain assumption The PVT-B0 features are a suitable lightweight encoder for ultrasound images.
    Encoder choice is not varied in ablation.
  • domain assumption Standard U-Net style encoder-decoder is an appropriate base for ultrasound segmentation.
    Taken as baseline design, not justified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of USEANet: Ultrasound-Specific Edge-Aware Multi-Branch Network for Lightweight Medical Image Segmentation." pith.science (2026). https://pith.science/paper/5L34WHHG

@misc{pith2026250908860,
  author       = {Pith},
  title        = {Pith review of: USEANet: Ultrasound-Specific Edge-Aware Multi-Branch Network for Lightweight Medical Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5L34WHHG}},
  note         = {Machine review of arXiv:2509.08860}
}
read the original abstract

Ultrasound image segmentation faces unique challenges including speckle noise, low contrast, and ambiguous boundaries, while clinical deployment demands computationally efficient models. We propose USEANet, an ultrasound-specific edge-aware multi-branch network that achieves optimal performance-efficiency balance through four key innovations: (1) ultrasound-specific multi-branch processing with specialized modules for noise reduction, edge enhancement, and contrast improvement; (2) edge-aware attention mechanisms that focus on boundary information with minimal computational overhead; (3) hierarchical feature aggregation with adaptive weight learning; and (4) ultrasound-aware decoder enhancement for optimal segmentation refinement. Built on an ultra-lightweight PVT-B0 backbone, USEANet significantly outperforms existing methods across five ultrasound datasets while using only 3.64M parameters and 0.79G FLOPs. Experimental results demonstrate superior segmentation accuracy with 67.01 IoU on BUSI dataset, representing substantial improvements over traditional approaches while maintaining exceptional computational efficiency suitable for real-time clinical applications. Code is available at https://github.com/chouheiwa/USEANet.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 19 canonical work pages

  1. [1]

    Deep learn- ing approaches, particularly U-Net [1] and its variants includ- ing AttUNet [2], have established strong baselines for medical image analysis

    INTRODUCTION Automated medical image segmentation has become crucial for diagnostic accuracy and treatment planning. Deep learn- ing approaches, particularly U-Net [1] and its variants includ- ing AttUNet [2], have established strong baselines for medical image analysis. Recent advances include Transformer-based methods like SwinUnet [3] and XboundFormer ...

    work page

  2. [2]

    METHOD This section presents USEANet, an ultrasound-specific edge- aware multi-branch network that addresses the inherent chal- lenges of ultrasound image segmentation through specialized architectural innovations, as illustrated in Fig. 1. 2.1. Overall Network Architecture USEANet employs an encoder-decoder U-Net architecture with four key innovations: (...

    work page Pith review arXiv 2025

  3. [3]

    Datasets We evaluate our proposed USEANet on five publicly available ultrasound datasets: three breast (BUSI, BUS-BRA, Breast- Lesions-USG) and two thyroid (DDTI, TN3K)

    EXPERIMENTS 3.1. Datasets We evaluate our proposed USEANet on five publicly available ultrasound datasets: three breast (BUSI, BUS-BRA, Breast- Lesions-USG) and two thyroid (DDTI, TN3K). BUSI[10] contains 780 breast ultrasound images,BUS- BRA[11] consists of 1,875 breast images,Breast-Lesions- USG[12] comprises 256 breast images,DDTI[13] contains 637 thyr...

    work page 1969

  4. [4]

    Our approach addresses ultrasound imaging challenges through specialized multi-branch feature process- ing and edge-aware attention mechanisms

    CONCLUSION We presented USEANet, a novel ultrasound-specific edge- aware multi-branch network for lightweight medical image segmentation. Our approach addresses ultrasound imaging challenges through specialized multi-branch feature process- ing and edge-aware attention mechanisms. Experimental results demonstrate superior segmentation performance while ma...

    work page

  5. [5]

    U-net: Convolutional networks for biomedical image segmentation,

    Olaf Ronneberger, Philipp Fischer, and Thomas Brox, “U-net: Convolutional networks for biomedical image segmentation,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2015, pp. 234–241

    work page 2015

  6. [6]

    Attu-net: at- tention u-net for brain tumor segmentation,

    Sihan Wang, Lei Li, and Xiahai Zhuang, “Attu-net: at- tention u-net for brain tumor segmentation,” inInterna- tional MICCAI brainlesion workshop. Springer, 2021, pp. 302–311

    work page 2021

  7. [7]

    Swin- unet: Unet-like pure transformer for medical image seg- mentation,

    Hu Cao, Yueyue Wang, Joy Chen, Dongsheng Jiang, Xi- aopeng Zhang, Qi Tian, and Manning Wang, “Swin- unet: Unet-like pure transformer for medical image seg- mentation,” inEuropean conference on computer vision. Springer, 2022, pp. 205–218

    work page 2022

  8. [8]

    Xbound-former: Toward cross- scale boundary modeling in transformers,

    Jiacheng Wang, Fei Chen, Yuxi Ma, Liansheng Wang, Zhaodong Fei, Jianwei Shuai, Xiangdong Tang, Qichao Zhou, and Jing Qin, “Xbound-former: Toward cross- scale boundary modeling in transformers,”IEEE Trans- actions on Medical Imaging, vol. 42, no. 6, pp. 1735– 1745, 2023

    work page 2023

Show all 20 references
  1. [9]

    Unext: Mlp-based rapid medical image segmentation network,

    Jeya Maria Jose Valanarasu and Vishal M Patel, “Unext: Mlp-based rapid medical image segmentation network,” inInternational conference on medical image comput- ing and computer-assisted intervention. Springer, 2022, pp. 23–33

    2022

  2. [10]

    Malunet: A multi-attention and light- weight unet for skin lesion segmentation,

    Jiacheng Ruan, Suncheng Xiang, Mingye Xie, Ting Liu, and Yuzhuo Fu, “Malunet: A multi-attention and light- weight unet for skin lesion segmentation,” in2022 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). IEEE, 2022, pp. 1150–1156

    2022

  3. [11]

    Ege-unet: an efficient group enhanced unet for skin lesion segmentation,

    Jiacheng Ruan, Mingye Xie, Jingsheng Gao, Ting Liu, and Yuzhuo Fu, “Ege-unet: an efficient group enhanced unet for skin lesion segmentation,” inInternational conference on medical image computing and computer- assisted intervention. Springer, 2023, pp. 481–490

    2023

  4. [12]

    Pyramid vision transformer: A versatile back- bone for dense prediction without convolutions,

    Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao, “Pyramid vision transformer: A versatile back- bone for dense prediction without convolutions,” inPro- ceedings of the IEEE/CVF international conference on computer visio...

    2021

  5. [13]

    Squeeze-and- excitation networks,

    Jie Hu, Li Shen, and Gang Sun, “Squeeze-and- excitation networks,” inProceedings of the IEEE con- ference on computer vision and pattern recognition, 2018, pp. 7132–7141

    2018

  6. [14]

    Dataset of breast ultrasound images,

    Walid Al-Dhabyani, Mohammed Gomaa, Hussien Khaled, and Aly Fahmy, “Dataset of breast ultrasound images,”Data in Brief, vol. 28, pp. 104863, 2020

    2020

  7. [15]

    Bus- bra: A breast ultrasound dataset for assessing computer- aided diagnosis systems,

    Hugo Figueiras and Jo ˜ao Manuel R. S. Tavares, “Bus- bra: A breast ultrasound dataset for assessing computer- aided diagnosis systems,”Medical Physics, vol. 50, no. 11, pp. 7025–7037, 2023

    2023

  8. [16]

    A curated benchmark dataset for ultrasound based breast lesion analysis (breast-lesions- usg)(version 1),

    A Pawłwska et al., “A curated benchmark dataset for ultrasound based breast lesion analysis (breast-lesions- usg)(version 1),”The Cancer Imaging Archive, 2024

    2024

  9. [17]

    Digital database for thyroid imaging,

    Lina Pedraza, Carlos Vargas, Fabi ´an Narv ´aez, Oscar Dur´an, Emma Mu ˜noz, and Eduardo Romero, “Digital database for thyroid imaging,” Available online, 2015, Accessed: 2024

    2015

  10. [18]

    Thyroid region prior guided attention for ultrasound segmentation of thyroid nodules,

    Hao Gong, Junyu Chen, Shiliang Chen, Yu Li, Gang Xiao, et al., “Thyroid region prior guided attention for ultrasound segmentation of thyroid nodules,”Com- puters in Biology and Medicine, vol. 155, pp. 106389, 2023

    2023

  11. [19]

    Trans- fuse: Fusing transformers and cnns for medical image segmentation,

    Yundong Zhang, Huiye Liu, and Qiang Hu, “Trans- fuse: Fusing transformers and cnns for medical image segmentation,” inInternational conference on medi- cal image computing and computer-assisted interven- tion. Springer, 2021, pp. 14–24

    2021

  12. [20]

    Pmfsnet: Po- larized multi-scale feature self-attention network for lightweight medical image segmentation,

    Jiahui Zhong, Wenhong Tian, Yuanlun Xie, Zhijia Liu, Jie Ou, Taoran Tian, and Lei Zhang, “Pmfsnet: Po- larized multi-scale feature self-attention network for lightweight medical image segmentation,”Computer Methods and Programs in Biomedicine, vol. 261, pp. 108611, 2025

    2025

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.