Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

A Self-Supervised Framework for Improved Generalisability in Ultrasound B-mode Image Segmentation

T0 review · 3 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read Self-supervised pretraining with ultrasound-specific augmentations, scored by a learned Relation Contrastive Loss, beats fully supervised training for breast lesion segmentation, especially when labels are scarce and test data come from a…

desk verdict The paper's broad empirical claim — SSL pretraining with domain-style augmentations helps low-data breast ultrasound segmentation — is credible, but the specific claim that the new Relation Contrastive Loss drives the gains is not supported by the experiments as presented. read the letter →

arxiv 2502.02489 v1 pith:JNP6BTFO submitted 2025-02-04 cs.CV cs.AIcs.LG

Edward Ellis , Andrew Bulpitt , Nasim Parsa , Michael F Byrne , Sharib Ali This is my paper · ORCID
classification cs.CVcs.AIcs.LG
keywords self-supervisedlearningcontrastiveultrasoundB-modeimagingbreastlesionsegmentationgeneralisabilityRelationLossfrequencyaugmentationmedicalimage
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 claims that a contrastive self-supervised pretraining scheme designed for B-mode ultrasound—standard grayscale ultrasound—can learn features that transfer to breast-lesion segmentation better than training the same network fully supervised. The scheme adds two ingredients: a novel Relation Contrastive Loss (RCL), which uses a small learned network to score how similar image and patch representations are, and data augmentations in both spatial and frequency domains that mimic real ultrasound noise and tissue structure. Across three public breast ultrasound datasets, the best self-supervised configurations consistently raise Dice scores relative to the supervised Res-UNet baseline, and the gains are largest when only 20% or 50% of the training labels are available. In the cross-dataset test, where a model trained on BUSI and BrEaST is evaluated on the held-out UDIAT set, the supervised baseline drops to 0.690 Dice with 20% training data while the best self-supervised method reaches 0.896. The paper concludes that domain-inspired self-supervised pretraining with these augmentations is a practical route to improved ultrasound segmentation, particularly when annotated data are limited and deployment data come from different scanners or centres.

What carries the argument

The load-bearing object is the Relation Contrastive Loss (RCL), a contrastive loss whose similarity score is produced by a small relation network rather than by cosine similarity. Image-level and patch-level feature embeddings are projected to 128-d, L2-normalised, combined by element-wise product into positive and negative pairs, and scored by a two-layer network with sigmoid output; the loss is the mean squared error between the scores and the labels 1 and 0. This loss is paired with a perceptual loss between image and patch features at layer 40 of the encoder, and the two are weighted by lambda. The pretext task itself is a Cross-patch Jigsaw operation: a frequency-augmented crop is split into 36 patches, a random patch defines focal row and column sets that are weakly transformed, and the remaining non-focal patches are shuffled, so the model must learn structure while being robust to local frequency distortion that mimics ultrasound noise.

What would settle it

Run the same BUSI+BrEaST pretraining and 20%-label downstream training on UDIAT with several random seeds, re-tuning lambda separately for PIRL+perceptual and RCL+perceptual on each target validation split; if the DSC gap between SSL and supervised training shrinks to within the seed-to-seed spread, or the RCL method no longer beats the PIRL baseline, the central claim of robust data-limited gains is not supported.

Watch

Extended reading notes

Core claim

The central claim is that when a contrastive self-supervised pretraining scheme is tailored to the physics of ultrasound B-mode images—frequency band-stop filtering that mimics speckle and reverberation, plus a spatial rearrangement that preserves partial tissue-layer structure—the learned encoder transfers to downstream breast-lesion segmentation better than a fully supervised Res-UNet, and the advantage grows as labelled data shrink and as the test distribution moves away from training. Technically, the scheme replaces the cosine-similarity scoring of a standard contrastive loss with a relation network that learns to score positive and negative pairs, giving the Relation Contrastive Loss (RCL), and combines it with a perceptual loss at layer 40 of the encoder. The paper reports consistent Dice improvements over the supervised baseline across BUSI, BrEaST, and UDIAT, and in the cross-dataset generalisability study the best SSL configuration reaches 0.914 Dice with 50% of BUSI+BrEaST labels and 0.896 with 20%, where the supervised baseline scores 0.778 and 0.690 respectively.

Load-bearing premise

The strongest assumption is that hyperparameters tuned on one dataset's validation split, including the loss weight lambda, per-method learning rates, and hand-chosen frequency filter ranges, transfer to other datasets, and that results from a single random seed are representative.

Editorial extensions

If this is right

  • On the BUSI dataset, the best SSL configuration reaches Dice 0.882 with 20% of training labels, against 0.847 for the supervised Res-UNet, and roughly 0.900 with 50% of labels.
  • On BrEaST, RCL with perceptual loss and the Cross-patch Jigsaw task reaches Dice 0.883 with 50% of labels, where the supervised baseline scores 0.795.
  • On UDIAT, adding frequency augmentation to the pretext task lifts the best full-data Dice to 0.918, compared with 0.873 for supervised training.
  • In the cross-dataset generalisability test, SSL with 20% of BUSI+BrEaST labels keeps UDIAT Dice near 0.896, while supervised training collapses to 0.690, meaning the distribution-shift penalty is much smaller for the SSL models.
  • Combining perceptual loss with either PIRL or RCL appears in the top-two methods in 11 of 12 dataset and proportion settings, making it a consistent ingredient of the reported gains.

Reading between the lines

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

  • Editorial extension: the same pretraining recipe should be tested on other ultrasound tasks with scarce labels, such as abdominal organ segmentation, which the paper names as future work but does not evaluate.
  • Editorial extension: because lambda and per-method learning rates were tuned only on the BUSI validation set, a natural follow-up is to re-tune these hyperparameters on each target dataset and see whether the reported rankings persist, which would separate the RCL contribution from the choice of hyperparameters.
  • Editorial extension: the frequency augmentation is designed to mimic realistic degradation, so a controlled experiment adding synthetic speckle and reverberation artefacts at test time could quantify how much of the generalisability gain is attributable to band-stop filtering rather than to the contrastive objective.
  • Editorial extension: the reported standard deviations are across test images rather than across training runs, so a multi-seed replication would tell whether the one-to-five percentage-point differences are stable enough to rely on in clinical deployment decisions.
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

3 major / 6 minor

Summary. The paper proposes a self-supervised contrastive pretraining framework for breast ultrasound B-mode segmentation, built on the PIRL jigsaw approach. The contributions are a domain-inspired "Cross-patch Jigsaw" spatial transformation, a frequency-domain band-stop augmentation, a novel Relation Contrastive Loss (RCL) based on relation networks, and a perceptual loss term. The method is evaluated by fine-tuning a Res-UNet on three public breast ultrasound datasets (BUSI, BrEaST, UDIAT) at 100%, 50%, and 20% of the training data, and by a cross-dataset generalisability experiment in which BUSI+BrEaST are used for training and UDIAT is the held-out test set. The authors report that their SSL approaches outperform a fully supervised Res-UNet baseline, particularly in low-data settings, and that frequency augmentation combined with perceptual loss and RCL gives the best results.

Significance. If the claims are upheld, the paper would provide a practical recipe for improving ultrasound segmentation when annotated data are scarce and when the deployment distribution differs from the training distribution. The study has clear strengths: it benchmarks a large number of method variants (two spatial pretext tasks, with and without frequency augmentation, and four loss configurations) on three datasets, reports several clinically relevant metrics (DSC, JC, HD, PPV, recall), and includes a dedicated out-of-distribution generalisability experiment. The frequency augmentation appears repeatedly among the best configurations and its effect is at least partially isolated by ablation. The paper does not ship code or trained models, and the headline novelty attribution to RCL is not supported by the current experiments, as detailed below.

major comments (3)
  1. [V-D, Eq. (13), Tables III–VI] The ablation in Table VII selects λ = 0.1 for RCL+perceptual and λ = 0.75 for PIRL+perceptual, and these settings are then used in all downstream comparisons. With Eq. (13), the deployed RCL+perceptual objective is 90% perceptual loss and only 10% RCL, whereas PIRL+perceptual is 75% NCE and 25% perceptual. The comparison therefore changes two variables at once (loss family and relative weighting), so the observed advantage of RCL+perceptual over PIRL+perceptual cannot be attributed to RCL. This is compounded by the fact that RCL alone is often worse than the PIRL baseline (e.g., BUSI 20% DSC 0.853 vs 0.870 in Table III; BrEaST 20% DSC 0.759 vs 0.811 in Table IV; BrEaST 50% DSC 0.840 vs 0.858 in Table IV). The central claim that the novel RCL drives the improved segmentation is therefore not established by the present experiments. Please add controlled comparisons in which the perceptual-loss weight is matched between methods and report RCL+perceptual at several λ values on the actual deployed pretext tasks (Jig+Freq and CP-Jig+Freq).
  2. [IV-B, V-B, Tables III–VI] All experiments use a single random seed (42), and the standard deviations reported in Tables III–VI are across test images, not across training runs. The two p-values in Section V-B (p = 0.039 and p = 0.090) are given without stating the test, whether it accounts for paired images, or any multiple-comparison correction across the 16 method variants and three data fractions. Since the paper repeatedly chooses the top two of 16 variants, the selection itself inflates the apparent gain. The phrase "significantly outperforms" in the abstract is therefore not supported by the reported statistics. Please provide multi-seed results (at least 3–5 seeds) or a corrected inferential procedure, and report variability across training runs rather than only across test images.
  3. [IV-B, Table VII, Table VI] The generalisability conclusions rest on hyperparameters tuned on the BUSI validation set with the Jigsaw baseline pretext task, including λ (Eq. 13), per-method learning rates, the memory-bank EMA weight, and the hand-chosen frequency filter ranges. These settings are then transferred to BrEaST, UDIAT, and the BUSI+BrEaST→UDIAT generalisability experiment, and to different pretext tasks (Jig+Freq, CP-Jig+Freq). Table VII itself shows that RCL+perceptual achieves DSC 0.890 at λ=0.1 and 0.885 at λ=0.25, so the selected value is not sharply distinguished. The claim that the method generalises across datasets would be considerably stronger with a sensitivity analysis on at least one additional dataset or a validation-based per-dataset selection of λ.
minor comments (6)
  1. [III-A2, Eq. (5)] The definition of Pnf is circular: it reads Pnf = {Pij | i=1..6, j=1..6} \ Pnf, but should presumably be "\ Pf" to denote the complement of the focal patches.
  2. [IV-B, Table II] Table II is captioned as a learning-rate sweep "for Jigsaw pretext task" but it reports DSC for PIRL, PIRL+perceptual, RCL, and RCL+perceptual. Clarify that the sweep covers all four method variants, or rename the table.
  3. [III-B, Fig. 1] The notation for the image-level representation is inconsistent: the text uses I_t1 and v_t1 in Section III-B, while Fig. 1 shows I'_t1 and v'_t1 from the memory bank. Define t0, t1, and t2 explicitly and use them consistently.
  4. [III-A1, III-B] The augmentation description says frequency filtering is applied to a random cropped region, but Section III-B says the patch division applies to "a frequency-augmented random cropped area". Please clarify whether the crop is also the region that is divided into 36 patches, and state the crop size or how it is sampled.
  5. [V-B] The two p-values in the generalisability section should be accompanied by the test name (e.g., paired t-test, Wilcoxon signed-rank test) and a statement about whether any correction for multiple comparisons was applied; as written, the reader cannot interpret them.
  6. [Tables III–VI] The Hausdorff distance values have very large standard deviations (e.g., 79.92 ± 62.64 in Table VI), indicating heavy-tailed outliers. Reporting the median or a percentile-based HD in addition to the mean would make the comparisons more stable and more clinically interpretable.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: reported gains rest on external benchmarks and held-out test sets, not on self-citation or definitional equivalence.

full rationale

The paper's central claim is empirical: contrastive pretraining with frequency/spatial augmentations and an RCL/perceptual objective improves downstream segmentation on BUSI/BrEaST/UDIAT. The pretraining losses (Eqs. 9-13) are defined from image pairs and memory-bank features, not from downstream Dice scores; downstream segmentation uses a separately trained Res-UNet on annotated data. Hyperparameters such as lambda in Eq. (13) are tuned on the BUSI validation set (Table VII) and then applied to held-out test splits and to the UDIAT generalisability set (Table VI), so the reported 'predictions' are not fitted to the evaluated targets. The only close-to-self reference, [36], is cited for implementation details of NCE with normalised embeddings and is not load-bearing; PIRL [16] is an external baseline. The RCL-vs-PIRL comparison uses different lambda values (0.1 vs 0.75), which weakens the attribution of gains to RCL, but this is a confound in the ablation, not a circular derivation.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The framework introduces no new physical entity. Its empirical claims rest on hand-chosen augmentation ranges, validation-tuned loss weights, and the pretrained backbone; these are the main quantities a replication would have to fix. No derivation reduces to fitted constants.

free parameters (6)
  • lambda in combined loss = 0.1 for RCL+perceptual, 0.75 for PIRL+perceptual
    Weight between RCL and perceptual loss in Eq. 13; selected on BUSI validation set (Table VII), so the reported results use a value tuned to one dataset.
  • w in Eq. 10 = 0.5
    Equal weighting between image-level and patch-level RCL; fixed by hand without sensitivity analysis.
  • memory bank EMA weight m_w = 0.5
    Exponential moving average for the memory bank, taken from PIRL [16]; not tuned for ultrasound.
  • per-method learning rate = varies, 0.0001 to 0.05 depending on method
    Learning rates for pretext learning were optimized per method on BUSI validation (Table II), then used across datasets; this is validation-set tuning.
  • frequency filter radii and thickness = inner radius 10 to 100; X-filter thickness 0 to 10
    Hand-chosen ranges for band-stop filtering; no sensitivity analysis is reported, yet the frequency augmentation is one of the two claimed novelty components.
  • random seed = 42
    All experiments use seed 42; no repeated runs with other seeds are reported, so the numbers may not be stable.
assumptions (4)
  • domain assumption ImageNet-pretrained ResNet50 features transfer to ultrasound B-mode images
    The encoder is initialized with ImageNet weights and then pretext-trained; see Sections III-B and IV-B. If this transfer is poor, the method's gains may not replicate in other US domains.
  • domain assumption Public breast ultrasound datasets and their ground-truth masks are reliable
    Section IV-A relies on BUSI, BrEaST and UDIAT annotations; no quality control or label-noise analysis is provided.
  • ad hoc to paper Hyperparameters found on BUSI validation generalize to BrEaST and UDIAT
    Section V-D tunes lambda and learning rates on BUSI; transferring these values to other datasets is assumed, not demonstrated.
  • ad hoc to paper Single-seed training is representative
    Section IV-B fixes seed 42; the paper treats this as sufficient to rank methods across datasets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Self-Supervised Framework for Improved Generalisability in Ultrasound B-mode Image Segmentation." pith.science (2026). https://pith.science/paper/JNP6BTFO

@misc{pith2026250202489,
  author       = {Pith},
  title        = {Pith review of: A Self-Supervised Framework for Improved Generalisability in Ultrasound B-mode Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JNP6BTFO}},
  note         = {Machine review of arXiv:2502.02489}
}
read the original abstract

Ultrasound (US) imaging is clinically invaluable due to its noninvasive and safe nature. However, interpreting US images is challenging, requires significant expertise, and time, and is often prone to errors. Deep learning offers assistive solutions such as segmentation. Supervised methods rely on large, high-quality, and consistently labeled datasets, which are challenging to curate. Moreover, these methods tend to underperform on out-of-distribution data, limiting their clinical utility. Self-supervised learning (SSL) has emerged as a promising alternative, leveraging unlabeled data to enhance model performance and generalisability. We introduce a contrastive SSL approach tailored for B-mode US images, incorporating a novel Relation Contrastive Loss (RCL). RCL encourages learning of distinct features by differentiating positive and negative sample pairs through a learnable metric. Additionally, we propose spatial and frequency-based augmentation strategies for the representation learning on US images. Our approach significantly outperforms traditional supervised segmentation methods across three public breast US datasets, particularly in data-limited scenarios. Notable improvements on the Dice similarity metric include a 4% increase on 20% and 50% of the BUSI dataset, nearly 6% and 9% improvements on 20% and 50% of the BrEaST dataset, and 6.4% and 3.7% improvements on 20% and 50% of the UDIAT dataset, respectively. Furthermore, we demonstrate superior generalisability on the out-of-distribution UDIAT dataset with performance boosts of 20.6% and 13.6% compared to the supervised baseline using 20% and 50% of the BUSI and BrEaST training data, respectively. Our research highlights that domain-inspired SSL can improve US segmentation, especially under data-limited conditions.

Figures

Figures reproduced from arXiv: 2502.02489 by the authors.

Figure 2
Figure 2. Frequency-Based Filtering Augmentation: The first row shows the [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 1
Figure 1. Block Diagram of our proposed SSL framework with domain-inspired data engineered pretext task that integrates perceptual loss ( [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 3
Figure 3. Cross-patch Jigsaw Task: From left to right: Image1: Cropped [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Qualitative evaluation of generalisability study on held out UDIAT dataset. 5 examples were chosen from small regular-shaped lesions to larger [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]

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. Self-Supervised Ultrasound-Video Segmentation with Feature Prediction and 3D Localised Loss

    cs.CV 2025-07 conditional novelty 4.0 of 10

    A 3D relative-localisation auxiliary loss improves V-JEPA pre-training for cardiac ultrasound video segmentation, with larger gains in low-label regimes.

Reference graph

Works this paper leans on

40 extracted references · 40 canonical work pages · cited by 1 Pith paper

  1. [1]

    Application of ultrasound in medicine,

    A. Carovac et al., “Application of ultrasound in medicine,” AIM, vol. 19, pp. 168–171, 2011

    work page 2011

  2. [2]

    Machine learning for medical ultrasound: status, methods, and future opportunities,

    L. J. Brattain et al. , “Machine learning for medical ultrasound: status, methods, and future opportunities,” Abdom. Radiol , vol. 43, pp. 786– 799, 2018

    work page 2018

  3. [3]

    A hybrid enhanced attention transformer network for medical ultrasound image segmentation,

    T. Jiang et al. , “A hybrid enhanced attention transformer network for medical ultrasound image segmentation,” Biomed. Signal Process. Control, vol. 86, p. 105329, 2023

    work page 2023

  4. [4]

    HAU-Net: Hybrid CNN-transformer for breast ultra- sound image segmentation,

    H. Zhang et al. , “HAU-Net: Hybrid CNN-transformer for breast ultra- sound image segmentation,” Biomed. Signal Process. Control , vol. 87, p. 105427, 2024. 11 TABLE VI ASSESSING GENERALISABILITY OF METHODS ON UDIAT DATASET. BUSI AND BREAST DATASETS ARE USED FOR PRETEXT AND DOWNSTREAM TRAINING . ALL DOWNSTREAM MODELS USE RES-UN ET WITH RESNET50 ENCODER . H...

    work page 2024

  5. [5]

    Cross-Image Dependency Modeling for Breast Ultrasound Segmentation,

    H. Wu et al., “Cross-Image Dependency Modeling for Breast Ultrasound Segmentation,” IEEE TMI , vol. 42, pp. 1619–1631, 2023

    work page 2023

  6. [6]

    Unified semantic model for medical image segmenta- tion,

    S. Yuan et al. , “Unified semantic model for medical image segmenta- tion,” Biomed. Signal Process. Control , vol. 98, p. 106711, 2024

    work page 2024

  7. [7]

    Self-supervised learning is more robust to dataset imbalance,

    H. Liu et al. , “Self-supervised learning is more robust to dataset imbalance,” in ICLR, 2022

    work page 2022

  8. [8]

    Toward Generalizability in the Deployment of Artifi- cial Intelligence in Radiology: Role of Computation Stress Testing to Overcome Underspecification,

    T. Eche et al. , “Toward Generalizability in the Deployment of Artifi- cial Intelligence in Radiology: Role of Computation Stress Testing to Overcome Underspecification,” Radiol.: Artif. Intell , vol. 3, p. e210097, 2021

    work page 2021

Show all 40 references
  1. [9]

    Big Self-Supervised Models Advance Medical Image Classification,

    S. Azizi et al. , “Big Self-Supervised Models Advance Medical Image Classification,” in 2021 IEEE/CVF ICCV , 2021, pp. 3458–3468

    2021

  2. [10]

    Self-supervised learning for medical image analysis using image context restoration,

    L. Chen et al. , “Self-supervised learning for medical image analysis using image context restoration,” Med. Image Anal. , vol. 58, p. 101539, 2019

    2019

  3. [11]

    SSL-CPCD: Self-supervised learning with composite pretext-class discrimination for improved generalisability in endoscopic image analysis,

    Z. Xu et al. , “SSL-CPCD: Self-supervised learning with composite pretext-class discrimination for improved generalisability in endoscopic image analysis,” IEEE TMI , vol. 43, pp. 4105–4119, 2024

    2024

  4. [12]

    VanBerlo et al., “A survey of the impact of self-supervised pretraining 12 Fig

    B. VanBerlo et al., “A survey of the impact of self-supervised pretraining 12 Fig. 4. Qualitative evaluation of generalisability study on held out UDIAT dataset. 5 examples were chosen from small regular-shaped lesions to larger irregular-shaped lesions. Segmentation predictio...

    2024

  5. [13]

    Self-Supervised Learning to More Efficiently Generate Segmentation Masks for Wrist Ultrasound,

    Y . Zhou et al. , “Self-Supervised Learning to More Efficiently Generate Segmentation Masks for Wrist Ultrasound,” in Simplifying Medical Ultrasound, 2023, vol. 14337, pp. 79–88

    2023

  6. [14]

    Thyroid ultrasound diagnosis improvement via multi- view self-supervised learning and two-stage pre-training,

    J. Wang et al. , “Thyroid ultrasound diagnosis improvement via multi- view self-supervised learning and two-stage pre-training,” Comput. Biol. Med., vol. 171, p. 108087, 2024

    2024

  7. [15]

    Self-Supervised Learning: Generative or Contrastive,

    X. Liu et al. , “Self-Supervised Learning: Generative or Contrastive,” IEEE Trans. Knowl. Data. Eng. , vol. 35, pp. 857–876, 2023

    2023

  8. [16]

    Self-Supervised Learning of Pretext- Invariant Representations,

    I. Misra and L. van der Maaten, “Self-Supervised Learning of Pretext- Invariant Representations,” in IEEE/CVF CVPR , 2020, pp. 6706–6716

    2020

  9. [17]

    Unsupervised Learning of Visual Represen- tations by Solving Jigsaw Puzzles,

    M. Noroozi and P. Favaro, “Unsupervised Learning of Visual Represen- tations by Solving Jigsaw Puzzles,” in Computer Vision – ECCV , 2016, vol. 9910, pp. 69–84

    2016

  10. [18]

    A simple framework for contrastive learning of visual representations

    T. Chen et al. , “A simple framework for contrastive learning of visual representations.” PMLR, 2020, pp. 1597–1607

    2020

  11. [19]

    U-Net: Convolutional Networks for Biomedical Image Segmentation,

    O. Ronneberger et al. , “U-Net: Convolutional Networks for Biomedical Image Segmentation,” in MICCAI, 2015, vol. 9351, pp. 234–241

    2015

  12. [20]

    ESTAN: Enhanced Small Tumor-Aware Network for Breast Ultrasound Image Segmentation,

    B. Shareef et al. , “ESTAN: Enhanced Small Tumor-Aware Network for Breast Ultrasound Image Segmentation,” Healthcare, vol. 10, p. 2262, 2022

    2022

  13. [21]

    Ultrasound spine image segmentation using multi-scale feature fusion skip-inception U-Net (SIU-Net),

    S. Banerjee et al. , “Ultrasound spine image segmentation using multi-scale feature fusion skip-inception U-Net (SIU-Net),” Biocybern. Biomed. Eng. , vol. 42, pp. 341–361, 2022

    2022

  14. [22]

    Deep Learning for Carotid Plaque Segmentation using a Dilated U-Net Architecture,

    N. H. Meshram et al., “Deep Learning for Carotid Plaque Segmentation using a Dilated U-Net Architecture,” Ultrason. Imaging , vol. 42, pp. 221–230, 2020

    2020

  15. [23]

    Dilated Squeeze-and-Excitation U-Net for Fetal Ultrasound Image Segmentation,

    D. Qiao and F. Zulkernine, “Dilated Squeeze-and-Excitation U-Net for Fetal Ultrasound Image Segmentation,” in IEEE CIBCB, 2020, pp. 1–7

    2020

  16. [24]

    Unsupervised representation learning by predicting image rotations,

    S. Gidaris, P. Singh, and N. Komodakis, “Unsupervised representation learning by predicting image rotations,” in ICLR, 2018

    2018

  17. [25]

    Identification method of thyroid nodule ultrasonography based on self-supervised learning dual-branch attention learning frame- work,

    Y . Xie et al. , “Identification method of thyroid nodule ultrasonography based on self-supervised learning dual-branch attention learning frame- work,” Health Inf. Sci. Syst , vol. 12, p. 7, 2024

    2024

  18. [26]

    Twin self-supervision based semi-supervised learning (TS-SSL): Retinal anomaly classification in SD-OCT images,

    Y . Zhang et al. , “Twin self-supervision based semi-supervised learning (TS-SSL): Retinal anomaly classification in SD-OCT images,” Neuro- computing, vol. 462, pp. 491–505, 2021

    2021

  19. [27]

    SimMIM: a Simple Framework for Masked Image Modeling,

    Z. Xie et al. , “SimMIM: a Simple Framework for Masked Image Modeling,” in IEEE/CVF CVPR , 2022, pp. 9643–9653

    2022

  20. [28]

    Momentum Contrast for Unsupervised Visual Represen- tation Learning,

    K. He et al. , “Momentum Contrast for Unsupervised Visual Represen- tation Learning,” in IEEE/CVF CVPR , 2020, pp. 9726–9735

    2020

  21. [29]

    Bootstrap your own latent-a new approach to self- supervised learning,

    J.-B. Grill et al. , “Bootstrap your own latent-a new approach to self- supervised learning,” Adv. neural inf. process. syst. , vol. 33, pp. 21 271– 21 284, 2020

    2020

  22. [30]

    Siamese neural networks for one-shot image recogni- tion,

    G. Koch et al. , “Siamese neural networks for one-shot image recogni- tion,” in ICML deep learning workshop , vol. 2, 2015, pp. 1–30

    2015

  23. [31]

    Prototypical networks for few-shot learning,

    J. Snell et al., “Prototypical networks for few-shot learning,” Adv. neural inf. process. syst. , vol. 30, 2017

    2017

  24. [32]

    Learning to Compare: Relation Network for Few-Shot Learning,

    F. Sung et al. , “Learning to Compare: Relation Network for Few-Shot Learning,” in 2018 IEEE/CVF CVPR , 2018, pp. 1199–1208

    2018

  25. [33]

    ImageNet: A large-scale hierarchical image database,

    J. Deng et al. , “ImageNet: A large-scale hierarchical image database,” in 2009 IEEE CVPR , 2009, pp. 248–255

    2009

  26. [34]

    Deep Residual Learning for Image Recognition,

    K. He et al., “Deep Residual Learning for Image Recognition,” in 2016 IEEE CVPR , 2016, pp. 770–778

    2016

  27. [35]

    Unsupervised Feature Learning via Non-parametric Instance Discrimination,

    Z. Wu et al. , “Unsupervised Feature Learning via Non-parametric Instance Discrimination,” in 2018 IEEE/CVF CVPR , 2018, pp. 3733– 3742

    2018

  28. [36]

    Self-Supervised Learning for Accurate Liver View Classification in Ultrasound Images with Minimal Labeled Data,

    A.-R. Ali et al. , “Self-Supervised Learning for Accurate Liver View Classification in Ultrasound Images with Minimal Labeled Data,” in 2023 IEEE/CVF CVPRW , 2023, pp. 3087–3093

    2023

  29. [37]

    Dataset of breast ultrasound images,

    W. Al-Dhabyani et al., “Dataset of breast ultrasound images,”Data Brief, vol. 28, p. 104863, 2020

    2020

  30. [38]

    Curated benchmark dataset for ultrasound based breast lesion analysis,

    A. Pawłowska et al. , “Curated benchmark dataset for ultrasound based breast lesion analysis,” Sci. Data , vol. 11, p. 148, 2024

    2024

  31. [39]

    Automated breast ultrasound lesions detection using convolutional neural networks,

    M. H. Yap et al. , “Automated breast ultrasound lesions detection using convolutional neural networks,” IEEE J. Biomed. Health Inform , vol. 22, pp. 1218–1226, 2017

    2017

  32. [40]

    Road extraction by deep Residual U-Net,

    Z. Zhang et al. , “Road extraction by deep Residual U-Net,” IEEE Geoscience and Remote Sensing Letters , vol. 15, pp. 749–753, 2018

    2018

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.