Pith. sign in

REVIEW 3 major objections 5 minor 63 references

Unleashing the Power of Intermediate Domains for Mixed Domain Semi-Supervised Medical Image Segmentation

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

Pith's one-line read A new framework, UST-RUN, claims to solve mixed-domain semi-supervised medical segmentation by constructing intermediate domains between labeled and unlabeled images, with a reported 12.94% Dice improvement on the Prostate dataset.

desk verdict Useful incremental extension of the authors' UST work, with a real reliable/unreliable sample-handling mechanism, but the missing patient-disjoint split statement puts the headline numbers on shaky ground. read the letter →

arxiv 2505.24567 v1 pith:UY7I7DCM submitted 2025-05-30 cs.CV

classification cs.CV
keywords semi-supervisedmedicalimagesegmentationdomainshiftmixedintermediatedomainscopy-pasteaugmentationpseudo-labelqualityamplitudemixupuncertaintyestimation
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

This paper defines a training scenario it calls MiDSS, in which only a handful of labeled images from one medical center are available while large quantities of unlabeled images come from several centers with different scanners, protocols, or patient populations. It argues that existing semi-supervised and unsupervised domain adaptation methods fail here because each tackles only one of the two obstacles, limited annotation and domain shift. The paper claims that UST-RUN, a mean-teacher framework that builds intermediate-domain samples by unified copy-paste between labeled and unlabeled images, adds symmetric pseudo-label guidance, progressive style mixing, and reliable or unreliable sample selection, and that this framework outperforms prior methods on the Fundus, Prostate, M&Ms, and BUSI datasets. The headline quantitative claim is a 12.94% Dice improvement on the Prostate dataset over the best prior method.

What carries the argument

The load-bearing object is the intermediate sample: an image formed by pasting one masked region of a labeled image into an unlabeled image, and vice versa, so that the image lies between the source and target domains in both content and style. UCP generates these samples in both directions; SymGD turns them into a two-way supervisory loop between intermediate samples and the original unlabeled data; TP-RAM modulates the low-frequency amplitude spectrum so that the style mixing ratio grows with training progress; and the reliable-sample queue uses teacher-student Dice disagreement as a hardness score to decide which unlabeled samples deserve high-quality pseudo-labels. The role of this machinery is to keep pseudo-label quality high while the domain gap is being narrowed, so that self-training does not accumulate errors.

What would settle it

A controlled experiment could settle it: construct a MiDSS benchmark where the unlabeled domain differs from the labeled domain by a known semantic flip, such as the same image statistics but with the object moved to a different position or the label inverted, and run UST-RUN. If the framework still improves Dice, then the gain is not coming from accurate pseudo-labels; if performance collapses, the central assumption is confirmed. A cheaper check is to rerun the Prostate experiment with 40 labeled images and replace the pseudo-labels in Eq. (2) with random masks of the same shape, in which case the method's margin over FixMatch should disappear if reliable pseudo-labels are load-bearing.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that domain gaps in semi-supervised medical segmentation can be bridged by explicitly constructing intermediate domains and training the model on them from both directions. UST-RUN creates these intermediate samples by Unified Copy-paste (UCP), applying both embedding directions between a labeled image and an unlabeled image, so each mixed image carries ground truth in one region and a pseudo-label in the other. A Symmetric Guidance (SymGD) loss then merges the pseudo-labels predicted by the teacher on two intermediate samples and uses the merged map to supervise the student on the original unlabeled image, weighting pixels by agreement between the two pseudo-label views. A Training Progress-aware Random Amplitude MixUp (TP-RAM) gradually blends low-frequency amplitude spectra from unlabeled images into labeled images, so stylistic transfer enters the training schedule softly instead of abruptly. The extension UST-RUN adds a queue of reliable unlabeled samples, chosen by teacher-student agreement, to diversify intermediate samples, and a separate strategy for unreliable samples that preserves their background while pasting in labeled foreground. The claim is that together these pieces let knowledge flow from the single labeled domain to multiple unlabeled domains, with consistent gains over comparison methods on four public datasets.

Load-bearing premise

The method assumes that after confidence masking, the teacher's pseudo-labels on the unlabeled parts of intermediate samples are accurate enough to train the student; if domain shift makes those pseudo-labels systematically wrong, the symmetric guidance and reliable-sample selection could amplify mistakes instead of correcting them.

Editorial extensions

If this is right

  • A model trained this way can be deployed across multiple unlabeled centers without knowing which center each image came from, since TP-RAM mixes styles sample-to-sample rather than domain-to-domain.
  • With only 10 labeled Fundus images, UST-RUN still beats methods that use 20 or 40 labeled images, suggesting the framework reduces annotation requirements more sharply than prior semi-supervised methods.
  • The reliable or unreliable sample split turns pseudo-label quality into an explicit training signal: easy samples diversify the intermediate domain, while hard samples receive refined supervision, so performance improves specifically in low-label regimes.
  • Because the upper-bound experiments use UCP with all labels from one domain, the reported gaps between UST-RUN and the upper bound quantify how much annotation quantity still matters even when domain bridging is in place.

Reading between the lines

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

  • Outside this paper, the teacher-student hardness score of Eq. (10) could serve as a general online data-quality filter for any semi-supervised pipeline whose unlabeled data arrive from heterogeneous sources, not only medical images.
  • The principle that style mixing should ramp with training progress is a testable recipe for other domain-bridging methods: one could ablate TP-RAM against a fixed high mixing ratio and against a cosine schedule, and expect the ramp to matter most when labeled data are extremely scarce.
  • The framework's reliance on pseudo-label quality suggests its gains will shrink as the number of target domains grows or the domain gap widens; a natural stress test is to add a third or fourth scanner vendor to the M&Ms setting and watch the performance curve.
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 / 5 minor

Summary. The paper introduces a new problem setting, Mixed Domain Semi-supervised medical image Segmentation (MiDSS), in which labeled data come from a single domain and unlabeled data come from multiple domains. The proposed UST-RUN framework extends the authors' earlier UST method by adding reliable-sample selection into a queue for generating diverse intermediate samples, a hardness-based dynamic threshold, and a targeted training strategy for unreliable samples, on top of the existing Unified Copy-paste (UCP), Symmetric Guidance (SymGD), and Training Process-aware Random Amplitude MixUp (TP-RAM) modules. Extensive experiments are reported on four public datasets (Fundus, Prostate, M&Ms, BUSI) with comparisons to semi-supervised, UDA, and foundation-model baselines, plus ablations and sensitivity analyses. The headline quantitative claim is a 12.94% Dice improvement on the Prostate dataset over prior state-of-the-art methods.

Significance. If the reported results are valid, MiDSS is a practically relevant setting, and UST-RUN provides a coherent, well-engineered solution that combines existing data-augmentation and pseudo-labeling ideas. The paper's strengths include extensive comparisons with standard deviations, module-by-module ablations, hyperparameter sensitivity studies, analyses of the reliable-sample queue, and public code. The main limitation is experimental: the paper does not demonstrate that the train/test split is subject-disjoint or volume-disjoint for the slice-based datasets, which is essential for the cross-domain generalization claim. The hyperparameter selection procedure also appears to use the evaluation datasets themselves, without a separate validation split. These concerns affect the central empirical claim and need to be addressed before the results can be fully trusted.

major comments (3)
  1. [Section IV-A] The split protocol for Prostate and M&Ms is described only as 'randomly split the dataset into training and testing sets based on a ratio of 4:1', with slices resized and randomly cropped. Because both datasets are volumetric and contain multiple adjacent slices per subject, an image-level random split can place slices from the same patient in both the unlabeled training set and the test set. UST-RUN explicitly selects 'reliable' unlabeled samples, stores them in a queue, and mixes them into intermediate samples, so the model can memorize patient-specific appearance rather than learn transferable domain knowledge. The headline 12.94% Prostate Dice improvement and the cross-domain margins on M&Ms could therefore be inflated by leakage. The authors should rerun the experiments with subject-disjoint or volume-disjoint splits, or provide dataset-specific evidence that slice-level independence is valid, and report per-subject evaluation if applicable.
  2. [Section IV-B and Tables IX-X] The hyperparameters beta (low-frequency region size), tau (confidence threshold), delta (hardness threshold increase factor), K (queue capacity), and gamma_0 are tuned on the Fundus and Prostate datasets, and the final results are reported on the same datasets in Tables III and IV. No separate validation split or nested tuning procedure is described. This means the reported numbers may reflect selection on the test distribution. The authors should either specify a held-out validation protocol used to choose these values or show that the final conclusions are stable across a range of hyperparameters without selecting the best values on the test set.
  3. [Section III-C and Fig. 12] The hardness score in Eq. (10) measures agreement between teacher and student pseudo-labels on the same weakly augmented image. High teacher-student agreement is a standard self-training signal, but it does not by itself guarantee correctness under domain shift, and the paper does not provide a formal argument that the dynamic threshold controls error accumulation. This is not fatal because Fig. 13 does show qualitative ground-truth comparisons and Fig. 12a reports that selected reliable samples have higher Dice than unreliable ones, which is supportive evidence. Still, a quantitative analysis of pseudo-label precision against ground truth for reliable versus unreliable samples, with respect to the domain of origin, would strengthen the central claim that the selection mechanism improves knowledge transfer rather than simply selecting easy samples.
minor comments (5)
  1. [Section IV-D, first paragraph] The text says 'As shown in Fig. IX' but the referenced content is Table IX, and the list of beta values omits 0.05 while duplicating 0.005; in the following paragraph 'increasing coefficients beta' should refer to the coefficient delta from Eq. (12).
  2. [Section III-B.2] The sentence 'In addition to L_in and L_in' should read 'L_in and L_out', since the losses for the two UCP directions are defined in Eq. (3).
  3. [Table I and Eq. (10)] The notation q_w used for the student pseudo-label in Eq. (10) is not defined in Table I; the table lists p_s and q_hat_w but not q_w, which makes the hardness-score definition harder to follow.
  4. [Section IV-D.4] The phrase 'randomly relected' should be 'randomly selected', and 'the full RUN-UST' should be 'the full UST-RUN'.
  5. [Section IV-A and IV-B] There are several minor typos: 'Postate' should be 'Prostate', and 'M&Ms datasetM&Ms [58]' needs a space before the citation.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: UST-RUN's components are defined by explicit equations and validated on external benchmarks; the only self-citation is the authors' conference version and is not load-bearing.

full rationale

The derivation chain is self-contained. UCP (Eq. 2), SymGD (Eqs. 5-7), TP-RAM (Eqs. 8-9), and the reliable/unreliable sample mechanisms (Eqs. 10-13) are all defined by explicit formulas and are not fitted to the headline Dice improvement. The 12.94% Prostate gain is a comparison against FixMatch (74.57 to 87.51 avg DC in Table IV), not against an input of the method. The hardness score h = 1 - Dice(q̂_w, q_w) is a teacher-student consistency heuristic; selecting low-h samples and using them as pseudo-labeled data is a standard self-training loop, and the paper validates the selection against ground truth in Fig. 13 rather than merely restating the selection criterion. The paper does cite the authors' conference version [20] and states 'This work is built upon our conference version [20]', but the UST equations are re-derived in Section III-B and the extensions are supported by fresh ablations (Tables VIII, XI, XII) and four external datasets, so the self-citation is not load-bearing. The test-split description in Section IV-A ('randomly split the dataset into training and testing sets based on a ratio of 4:1') does not state a patient-disjoint split; if slices from the same subject appear in both the unlabeled training set and the test set, the reported gains could be inflated. That is a data-hygiene/validity concern, not a circularity of the derivation chain.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The framework introduces no new physical or algorithmic entities beyond the modules themselves. Its parameters are fitted on the evaluation datasets, and its domain assumptions (style in amplitude, hardness as reliability) are carried from prior work or standard practice.

free parameters (5)
  • beta (low-frequency region size) = 0.01
    Chosen on Fundus dataset (Table IX) to maximize Dice; controls the mixing area in TP-RAM.
  • tau (confidence threshold) = 0.95
    Chosen on Fundus dataset (Table IX); pixels below this threshold are excluded from pseudo-label supervision.
  • delta (hardness threshold increase factor) = 1.0005
    Chosen on Prostate dataset (Table X) to adapt the reliable-sample selection criterion.
  • K (reliable queue capacity) = 20
    Chosen on Prostate dataset (Table X); affects the diversity and quality of reliable samples.
  • gamma_0 (initial hardness threshold) = not stated
    Initial threshold for reliable sample selection; value not reported in the paper.
assumptions (4)
  • domain assumption Low-frequency amplitude of the Fourier transform encodes style; mixing it transfers style while preserving content.
    Invoked in TP-RAM (Sec III-B3, Eq. 9). Cited from [19], not independently verified; if false, style transition could corrupt semantics.
  • domain assumption Teacher-student disagreement (hardness) correlates with segmentation accuracy.
    Invoked in Eq. (10), Sec III-C1. Supported by Fig. 12a, but the assumption is not proven; selection of reliable samples depends on it.
  • domain assumption Copy-paste between images from different domains yields valid intermediate-domain training samples.
    Invoked in UCP (Sec III-B1, Eq. 2). Assumes pasted patches retain plausible anatomy and do not introduce unrealistic artifacts.
  • standard math EMA teacher provides stable and accurate pseudo-labels.
    Standard mean-teacher assumption from [50], used throughout the method.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Unleashing the Power of Intermediate Domains for Mixed Domain Semi-Supervised Medical Image Segmentation." pith.science (2026). https://pith.science/paper/UY7I7DCM

@misc{pith2026250524567,
  author       = {Pith},
  title        = {Pith review of: Unleashing the Power of Intermediate Domains for Mixed Domain Semi-Supervised Medical Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UY7I7DCM}},
  note         = {Machine review of arXiv:2505.24567}
}
read the original abstract

Both limited annotation and domain shift are prevalent challenges in medical image segmentation. Traditional semi-supervised segmentation and unsupervised domain adaptation methods address one of these issues separately. However, the coexistence of limited annotation and domain shift is quite common, which motivates us to introduce a novel and challenging scenario: Mixed Domain Semi-supervised medical image Segmentation (MiDSS), where limited labeled data from a single domain and a large amount of unlabeled data from multiple domains. To tackle this issue, we propose the UST-RUN framework, which fully leverages intermediate domain information to facilitate knowledge transfer. We employ Unified Copy-paste (UCP) to construct intermediate domains, and propose a Symmetric GuiDance training strategy (SymGD) to supervise unlabeled data by merging pseudo-labels from intermediate samples. Subsequently, we introduce a Training Process aware Random Amplitude MixUp (TP-RAM) to progressively incorporate style-transition components into intermediate samples. To generate more diverse intermediate samples, we further select reliable samples with high-quality pseudo-labels, which are then mixed with other unlabeled data. Additionally, we generate sophisticated intermediate samples with high-quality pseudo-labels for unreliable samples, ensuring effective knowledge transfer for them. Extensive experiments on four public datasets demonstrate the superiority of UST-RUN. Notably, UST-RUN achieves a 12.94% improvement in Dice score on the Prostate dataset. Our code is available at https://github.com/MQinghe/UST-RUN

Figures

Figures reproduced from arXiv: 2505.24567 by the authors.

Figure 1
Figure 1. The upper illustrates SSIMS, UDA, and MiDSS. The [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The overall framework of our UST-RUN emphasizes domain knowledge transfer through data augmentation and training [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The illustration of UCP between images. Cut refers to splitting the image into two parts according to Mα, while Mix implies merging two parts of images back together. A. Notation and Problem Setting In MiDSS scenario, the domain gap exists among images originating from K data centers {Di} K i=1. The training set comprises N labeled images {(xi , yi)} N i=1 from a single domain Dj ∈ {Di} K i=1 and M unlabeled images … view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: The illustration of TP-RAM. Fast Fourier transform [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: In UST, intermediate samples are solely generated [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Examples from the four datasets. Each pair of images [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Visual results from the Fundus dataset. The top test [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 10
Figure 10. Figure 10: Visual results of BUSI dataset. The first row shows [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]
Figure 9
Figure 9. Figure 9: Visual results of M&Ms dataset. The first row shows [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 11
Figure 11. Figure 11: Ablation experiments on the Prostate dataset directly [PITH_FULL_IMAGE:figures/full_fig_p010_11.png]
Figure 12
Figure 12. Figure 12: Further analysis of reliable samples. The results are obtained from an experiment using 40 labeled data from the [PITH_FULL_IMAGE:figures/full_fig_p011_12.png]
Figure 13
Figure 13. Figure 13: Examples of reliable samples and unreliable samples. [PITH_FULL_IMAGE:figures/full_fig_p012_13.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

63 extracted references · 56 canonical work pages

  1. [1]

    Shape-aware meta-learning for generalizing prostate mri segmentation to unseen domains,

    Q. Liu, Q. Dou, and P.-A. Heng, “Shape-aware meta-learning for generalizing prostate mri segmentation to unseen domains,” inMICCAI. Springer, 2020, pp. 475–485

  2. [2]

    Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation,

    L. Yu, S. Wang, X. Li, C.-W. Fu, and P.-A. Heng, “Uncertainty-aware self-ensembling model for semi-supervised 3d left atrium segmentation,” inMICCAI. Springer, 2019, pp. 605–613

  3. [3]

    Inf-net: Automatic covid-19 lung infection segmentation from ct images,

    D.-P. Fan, T. Zhou, G.-P. Ji, Y . Zhou, G. Chen, H. Fu, J. Shen, and L. Shao, “Inf-net: Automatic covid-19 lung infection segmentation from ct images,”TMI, 2020

  4. [4]

    Transformation-consistent self-ensembling model for semisupervised medical image segmentation,

    X. Li, L. Yu, H. Chen, C.-W. Fu, L. Xing, and P.-A. Heng, “Transformation-consistent self-ensembling model for semisupervised medical image segmentation,”TNNLS, 2020

  5. [5]

    Ss-tbn: A semi-supervised tri-branch network for covid-19 screening and lesion segmentation,

    L.-L. Zeng, K. Gao, D. Hu, Z. Feng, C. Hou, P. Rong, and W. Wang, “Ss-tbn: A semi-supervised tri-branch network for covid-19 screening and lesion segmentation,”TPAMI, 2023

  6. [6]

    Momentum contrast for unsupervised visual representation learning,

    K. He, H. Fan, Y . Wu, S. Xie, and R. Girshick, “Momentum contrast for unsupervised visual representation learning,” inCVPR, 2020

  7. [7]

    Big data in healthcare: management, analysis and future prospects,

    S. Dash, S. K. Shakyawar, M. Sharma, and S. Kaushik, “Big data in healthcare: management, analysis and future prospects,”Journal of big data, vol. 6, no. 1, pp. 1–25, 2019

  8. [8]

    Opportunities and challenges in using real-world data for health care,

    V . A. Rudrapatna, A. J. Butteet al., “Opportunities and challenges in using real-world data for health care,”The Journal of Clinical Investigation, 2020

Show all 63 references
  1. [9]

    Semi-supervised learning by aug- mented distribution alignment,

    Q. Wang, W. Li, and L. V . Gool, “Semi-supervised learning by aug- mented distribution alignment,” inICCV, 2019

  2. [10]

    Domain adaptation for medical image analysis: a survey,

    H. Guan and M. Liu, “Domain adaptation for medical image analysis: a survey,”T-BME, 2021

  3. [11]

    Unsupervised bidirectional cross-modality adaptation via deeply synergistic image and feature alignment for medical image segmentation,

    C. Chen, Q. Dou, H. Chen, J. Qin, and P. A. Heng, “Unsupervised bidirectional cross-modality adaptation via deeply synergistic image and feature alignment for medical image segmentation,”TMI, 2020

  4. [12]

    From source to target and back: symmetric bi-directional adaptive gan,

    P. Russo, F. M. Carlucci, T. Tommasi, and B. Caputo, “From source to target and back: symmetric bi-directional adaptive gan,” inCVPR, 2018

  5. [13]

    Unsupervised cross-modality domain adaptation of convnets for biomedical image segmentations with adversarial loss,

    Q. Dou, C. Ouyang, C. Chen, H. Chen, and P.-A. Heng, “Unsupervised cross-modality domain adaptation of convnets for biomedical image segmentations with adversarial loss,”arXiv, 2018. 13

  6. [14]

    Ecacl: A holistic framework for semi-supervised domain adaptation,

    K. Li, C. Liu, H. Zhao, Y . Zhang, and Y . Fu, “Ecacl: A holistic framework for semi-supervised domain adaptation,” inProceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 8578–8587

  7. [15]

    Attract, perturb, and explore: Learning a feature alignment network for semi-supervised domain adaptation,

    T. Kim and C. Kim, “Attract, perturb, and explore: Learning a feature alignment network for semi-supervised domain adaptation,” inEuropean conference on computer vision

  8. [16]

    Multi-level consistency learning for semi-supervised domain adaptation,

    Z. Yan, Y . Wu, G. Li, Y . Qin, X. Han, and S. Cui, “Multi-level consistency learning for semi-supervised domain adaptation,”arXiv preprint arXiv:2205.04066, 2022

  9. [17]

    Cutmix: Reg- ularization strategy to train strong classifiers with localizable features,

    S. Yun, D. Han, S. J. Oh, S. Chun, J. Choe, and Y . Yoo, “Cutmix: Reg- ularization strategy to train strong classifiers with localizable features,” inICCV, 2019

  10. [18]

    Separated con- trastive learning for organ-at-risk and gross-tumor-volume segmentation with limited annotation,

    J. Wang, X. Li, Y . Han, J. Qin, L. Wang, and Z. Qichao, “Separated con- trastive learning for organ-at-risk and gross-tumor-volume segmentation with limited annotation,” inAAAI, 2022

  11. [19]

    Feddg: Federated domain generalization on medical image segmentation via episodic learning in continuous frequency space,

    Q. Liu, C. Chen, J. Qin, Q. Dou, and P.-A. Heng, “Feddg: Federated domain generalization on medical image segmentation via episodic learning in continuous frequency space,” inCVPR, 2021

  12. [20]

    Constructing and exploring intermediate domains in mixed domain semi-supervised medical image segmentation,

    Q. Ma, J. Zhang, L. Qi, Q. Yu, Y . Shi, and Y . Gao, “Constructing and exploring intermediate domains in mixed domain semi-supervised medical image segmentation,” inCVPR, 2024

  13. [21]

    Challenges and methodologies of fully automatic whole heart segmentation: a review,

    X. Zhuang, “Challenges and methodologies of fully automatic whole heart segmentation: a review,”Journal of Healthcare Engineering, 2013

  14. [22]

    Shape-aware semi-supervised 3d semantic segmentation for medical images,

    S. Li, C. Zhang, and X. He, “Shape-aware semi-supervised 3d semantic segmentation for medical images,” inMICCAI, 2020

  15. [23]

    Semi-supervised medical image segmentation through dual-task consistency,

    X. Luo, J. Chen, T. Song, and G. Wang, “Semi-supervised medical image segmentation through dual-task consistency,” inAAAI, 2021

  16. [24]

    Exploring smoothness and class-separation for semi-supervised medical image segmentation,

    Y . Wu, Z. Wu, Q. Wu, Z. Ge, and J. Cai, “Exploring smoothness and class-separation for semi-supervised medical image segmentation,” in MICCAI, 2022

  17. [25]

    Caussl: Causality- inspired semi-supervised learning for medical image segmentation,

    J. Miao, C. Chen, F. Liu, H. Wei, and P.-A. Heng, “Caussl: Causality- inspired semi-supervised learning for medical image segmentation,” in ICCV, 2023

  18. [26]

    Adaptive bidirectional displace- ment for semi-supervised medical image segmentation,

    H. Chi, J. Pang, B. Zhang, and W. Liu, “Adaptive bidirectional displace- ment for semi-supervised medical image segmentation,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2024, pp. 4070–4080

  19. [27]

    Segment anything,

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Loet al., “Segment anything,” inICCV, 2023, pp. 4015–4026

  20. [28]

    Customized segment anything model for medical image segmentation,

    K. Zhang and D. Liu, “Customized segment anything model for medical image segmentation,”arXiv, 2023

  21. [29]

    Sam fails to segment anything?–sam-adapter: Adapting sam in underperformed scenes: Camouflage, shadow, medical image segmentation, and more,

    T. Chen, L. Zhu, C. Ding, R. Cao, Y . Wang, Z. Li, L. Sun, P. Mao, and Y . Zang, “Sam fails to segment anything?–sam-adapter: Adapting sam in underperformed scenes: Camouflage, shadow, medical image segmentation, and more,”arXiv preprint arXiv:2304.09148, 2023

  22. [30]

    Segment anything in medical images,

    J. Ma, Y . He, F. Li, L. Han, C. You, and B. Wang, “Segment anything in medical images,”Nature Communications, vol. 15, no. 1, p. 654, 2024

  23. [31]

    Unleashing the potential of sam for medical adaptation via hierarchical decoding,

    Z. Cheng, Q. Wei, H. Zhu, Y . Wang, L. Qu, W. Shao, and Y . Zhou, “Unleashing the potential of sam for medical adaptation via hierarchical decoding,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 3511–3522

  24. [32]

    Adversar- ial image synthesis for unpaired multi-modal cardiac data,

    A. Chartsias, T. Joyce, R. Dharmakumar, and S. A. Tsaftaris, “Adversar- ial image synthesis for unpaired multi-modal cardiac data,” inSASHIMI, 2017

  25. [33]

    Domain-adversarial training of neural networks,

    Y . Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Lavio- lette, M. Marchand, and V . Lempitsky, “Domain-adversarial training of neural networks,”JMLR, 2016

  26. [34]

    Le- uda: Label-efficient unsupervised domain adaptation for medical image segmentation,

    Z. Zhao, F. Zhou, K. Xu, Z. Zeng, C. Guan, and S. K. Zhou, “Le- uda: Label-efficient unsupervised domain adaptation for medical image segmentation,”TMI, 2022

  27. [35]

    Learning to adapt structured output space for semantic segmentation,

    Y .-H. Tsai, W.-C. Hung, S. Schulter, K. Sohn, M.-H. Yang, and M. Chandraker, “Learning to adapt structured output space for semantic segmentation,” inCVPR, 2018

  28. [36]

    Rectifying pseudo label learning via uncertainty estimation for domain adaptive semantic segmentation,

    Z. Zheng and Y . Yang, “Rectifying pseudo label learning via uncertainty estimation for domain adaptive semantic segmentation,”IJCV, 2021

  29. [37]

    Col- laborative unsupervised domain adaptation for medical image diagnosis,

    Y . Zhang, Y . Wei, Q. Wu, P. Zhao, S. Niu, J. Huang, and M. Tan, “Col- laborative unsupervised domain adaptation for medical image diagnosis,” TIP, 2020

  30. [38]

    De- liberated domain bridging for domain adaptive semantic segmentation,

    L. Chen, Z. Wei, X. Jin, H. Chen, M. Zheng, K. Chen, and Y . Jin, “De- liberated domain bridging for domain adaptive semantic segmentation,” NeurIPS, vol. 35, pp. 15 105–15 118, 2022

  31. [39]

    Fda: Fourier domain adaptation for semantic segmentation,

    Y . Yang and S. Soatto, “Fda: Fourier domain adaptation for semantic segmentation,” inCVPR, 2020

  32. [40]

    Tent: Fully test-time adaptation by entropy minimization,

    D. Wang, E. Shelhamer, S. Liu, B. Olshausen, and T. Darrell, “Tent: Fully test-time adaptation by entropy minimization,”arXiv, 2020

  33. [41]

    Each test image deserves a specific prompt: Continual test-time adaptation for 2d medical image segmentation,

    Z. Chen, Y . Pan, Y . Ye, M. Lu, and Y . Xia, “Each test image deserves a specific prompt: Continual test-time adaptation for 2d medical image segmentation,” inCVPR, 2024

  34. [42]

    When source-free domain adaptation meets learning with noisy labels,

    L. Yi, G. Xu, P. Xu, J. Li, R. Pu, C. Ling, A. I. McLeod, and B. Wang, “When source-free domain adaptation meets learning with noisy labels,” arXiv preprint arXiv:2301.13381, 2023

  35. [43]

    Improving semi-supervised domain adaptation using effective target selection and semantics,

    A. Singh, N. Doraiswamy, S. Takamuku, M. Bhalerao, T. Dutta, S. Biswas, A. Chepuri, B. Vengatesan, and N. Natori, “Improving semi-supervised domain adaptation using effective target selection and semantics,” inProceedings of the IEEE/CVF conference on computer vision and patte...

  36. [44]

    Contradictory structure learning for semi-supervised domain adaptation,

    C. Qin, L. Wang, Q. Ma, Y . Yin, H. Wang, and Y . Fu, “Contradictory structure learning for semi-supervised domain adaptation,” inProceed- ings of the 2021 SIAM International Conference on Data Mining (SDM). SIAM, 2021, pp. 576–584

  37. [45]

    Bidirectional adversarial training for semi-supervised domain adaptation

    P. Jiang, A. Wu, Y . Han, Y . Shao, M. Qi, and B. Li, “Bidirectional adversarial training for semi-supervised domain adaptation.” inIJCAI, 2020, pp. 934–940

  38. [46]

    Semi-supervised domain adap- tation via prototype-based multi-level learning,

    X. Huang, C. Zhu, and W. Chen, “Semi-supervised domain adap- tation via prototype-based multi-level learning,”arXiv preprint arXiv:2305.02693, 2023

  39. [47]

    mixup: Beyond empirical risk minimization,

    H. Zhang, M. Cisse, Y . N. Dauphin, and D. Lopez-Paz, “mixup: Beyond empirical risk minimization,”arXiv, 2017

  40. [48]

    Bidirectional copy-paste for semi-supervised medical image segmentation,

    Y . Bai, D. Chen, Q. Li, W. Shen, and Y . Wang, “Bidirectional copy-paste for semi-supervised medical image segmentation,” inCVPR, 2023

  41. [49]

    Ucc: Uncertainty guided cross- head co-training for semi-supervised semantic segmentation,

    J. Fan, B. Gao, H. Jin, and L. Jiang, “Ucc: Uncertainty guided cross- head co-training for semi-supervised semantic segmentation,” inCVPR, 2022

  42. [50]

    Mean teachers are better role mod- els: Weight-averaged consistency targets improve semi-supervised deep learning results,

    A. Tarvainen and H. Valpola, “Mean teachers are better role mod- els: Weight-averaged consistency targets improve semi-supervised deep learning results,”NeurIPS, vol. 30, pp. 1195–1204, 2017

  43. [51]

    Fixmatch: Simplifying semi- supervised learning with consistency and confidence,

    K. Sohn, D. Berthelot, N. Carlini, Z. Zhang, H. Zhang, C. A. Raffel, E. D. Cubuk, A. Kurakin, and C.-L. Li, “Fixmatch: Simplifying semi- supervised learning with consistency and confidence,”NeurIPS, vol. 33, pp. 596–608, 2020

  44. [52]

    Fixmatchseg: Fixing fixmatch for semi- supervised semantic segmentation,

    P. Upretee and B. Khanal, “Fixmatchseg: Fixing fixmatch for semi- supervised semantic segmentation,”arXiv, 2022

  45. [53]

    Revisiting weak-to- strong consistency in semi-supervised semantic segmentation,

    L. Yang, L. Qi, L. Feng, W. Zhang, and Y . Shi, “Revisiting weak-to- strong consistency in semi-supervised semantic segmentation,” inCVPR, 2023

  46. [54]

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

    J.-B. Grill, F. Strub, F. Altch ´e, C. Tallec, P. Richemond, E. Buchatskaya, C. Doersch, B. Avila Pires, Z. Guo, M. Gheshlaghi Azaret al., “Bootstrap your own latent-a new approach to self-supervised learning,” NeurIPS, vol. 33, pp. 21 271–21 284, 2020

  47. [55]

    Instance credibility inference for few-shot learning,

    Y . Wang, C. Xu, C. Liu, L. Zhang, and Y . Fu, “Instance credibility inference for few-shot learning,” inCVPR, 2020

  48. [56]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inMICCAI, 2015

  49. [57]

    Dofe: Domain-oriented feature embedding for generalizable fundus image segmentation on unseen datasets,

    S. Wang, L. Yu, K. Li, X. Yang, C.-W. Fu, and P.-A. Heng, “Dofe: Domain-oriented feature embedding for generalizable fundus image segmentation on unseen datasets,”TMI, 2020

  50. [58]

    Multi- centre, multi-vendor and multi-disease cardiac segmentation: the m&ms challenge,

    V . M. Campello, P. Gkontra, C. Izquierdo, C. Martin-Isla, A. Sojoudi, P. M. Full, K. Maier-Hein, Y . Zhang, Z. He, J. Maet al., “Multi- centre, multi-vendor and multi-disease cardiac segmentation: the m&ms challenge,”TMI, 2021

  51. [59]

    Dataset of breast ultrasound images,

    W. Al-Dhabyani, M. Gomaa, H. Khaled, and A. Fahmy, “Dataset of breast ultrasound images,”Data in brief, 2020

  52. [60]

    Unsupervised domain adaptation for cardiac segmentation: Towards structure mutual information maximiza- tion,

    C. Lu, S. Zheng, and G. Gupta, “Unsupervised domain adaptation for cardiac segmentation: Towards structure mutual information maximiza- tion,” inCVPR, 2022

  53. [61]

    Classmix: Segmentation-based data augmentation for semi-supervised learning,

    V . Olsson, W. Tranheden, J. Pinto, and L. Svensson, “Classmix: Segmentation-based data augmentation for semi-supervised learning,” inWACV, 2021

  54. [62]

    Milking cowmask for semi- supervised image classification,

    G. French, A. Oliver, and T. Salimans, “Milking cowmask for semi- supervised image classification,”arXiv, 2020

  55. [63]

    Fmix: Enhancing mixed sample data augmentation,

    E. Harris, A. Marcu, M. Painter, M. Niranjan, A. Pr ¨ugel-Bennett, and J. Hare, “Fmix: Enhancing mixed sample data augmentation,”arXiv, 2020

Pith tools

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