Pith. sign in

REVIEW 5 major objections 5 minor 40 references

Semi-supervised medical image segmentation can be made reliable without a teacher by disentangling foreground and background features and weighting pseudo-labels by structural consistency.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 20:09 UTC pith:QOLJSUDG

load-bearing objection A legitimate within-subfield advance: teacher-free SSL segmentation with a structural reliability map that is plausible but not yet proven; deserves peer review, not unconditional acceptance. the 5 major comments →

arxiv 2607.16705 v1 pith:QOLJSUDG submitted 2026-07-18 cs.CV cs.AI

OFD-Net: Teacher-Free Reliable Semi-supervised Medical Image Segmentation with Orthogonal Feature Disentanglement Net of Foreground-Background

Shao-feng Jiang , Zhe-yang Jing , Qin Lu , Huan-huan Shi , Zhen Chen , Cong-xuan zhang , Chen Yi This is my paper
classification cs.CV cs.AI
keywords semi-supervised learningmedical image segmentationfeature disentanglementteacher-freepseudo-label reliabilityorthogonality constraintreliability-aware learningforeground-background separation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Semi-supervised segmentation relies on pseudo-labels whose quality is normally judged by confidence, but confidence can be high while the structure is wrong. This paper argues that within a single network, without a teacher model or cross-network consistency, one can explicitly split the deepest features into foreground and background branches, force them to be orthogonal, and use them as an internal structural reference. That reference both guides the decoder and produces a pixel-wise reliability weight for pseudo-label supervision. The authors report consistent gains over teacher-based alternatives on four public benchmarks under 10–20% labels, with the largest improvements on the hardest unlabeled samples. The claim to a fair reader is that structure-based reliability, not confidence, is the missing ingredient for reliable unlabeled supervision.

Core claim

OFD-Net replaces the teacher-student loop with an internal foreground-background structural reference. An Orthogonal Feature Disentanglement Module projects the deepest encoder feature into two mutually exclusive representations, minimized in cosine similarity per spatial location; a Disentanglement Guidance Module feeds these as prototypes into the decoder; and a reliability map Mrel = (1−|Pred−Pfg|)·|Pfg−Pbg| weights pseudo-label loss. The authors show that each component contributes in ablations, that the orthogonal constraint is needed for clean separation, and that Mrel's gains grow monotonically with sample difficulty, which they read as evidence the mechanism suppresses exactly the un

What carries the argument

The load-bearing object is the reliability map Mrel, defined as (1−|Pred−Pfg|)·|Pfg−Pbg|. It encodes two criteria: main prediction agrees with the disentangled foreground branch, and foreground and background responses are far apart. OFDM supplies the two branches with an orthogonality loss on cosine similarity; DGM injects those branches into the decoder. Mrel then becomes the pixel-wise weight on pseudo-label supervision in Stage II, gating both main and auxiliary heads.

Load-bearing premise

The whole mechanism rests on the assumption that the reliability map Mrel, built from the network's own predictions and disentangled branches, is a faithful measure of which pseudo-labels are actually wrong; the paper never demonstrates this against true pseudo-label errors.

What would settle it

Train OFD-Net as described, freeze it on a held-out unlabeled set, and compute the pixel-wise correlation between Mrel and the agreement of the hard pseudo-label with the ground-truth mask. If the correlation is near zero or negative on hard samples, or if replacing Mrel with a uniform weight changes nothing, the claimed structural reliability criterion collapses into a confidence heuristic.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Semi-supervised segmentation can avoid the memory and complexity cost of a teacher network while keeping pseudo-label quality high.
  • Pseudo-label reliability can be judged by structure rather than confidence, which may resist confirmation bias on datasets with large appearance variation.
  • The same disentanglement and reliability weighting could transfer to any dense prediction task where foreground/background structure is well-defined, such as lesion, organ, or object segmentation.
  • The quartile result implies that reliability weighting mainly rescues hard unlabeled samples, so it should complement rather than replace confidence filtering.
  • The ablation evidence suggests the orthogonality loss is a necessary precondition for the reliability map to be informative, not a minor regularizer.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Mrel's validity is asserted but never checked against oracle pseudo-label error; a direct test would be to freeze the model and compute the correlation between Mrel and pixel-wise agreement with ground truth on a validation set.
  • A testable extension: if Mrel truly encodes structure, then replacing the trained foreground/background heads with frozen, independently trained heads should degrade Mrel's usefulness; if it does not, the mechanism reduces to another confidence heuristic.
  • The framework leaves 3D volumetric segmentation open; a natural extension is to apply the same reliability logic with 3D convolutions or across slice sequences.
  • Since Mrel multiplies the pseudo-label loss, it changes the effective label balance spatially; a simpler variant might achieve similar gains by scheduling Mrel's temperature rather than using it as a hard gate.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes OFD-Net, a teacher-free single-network semi-supervised learning framework for medical image segmentation. It introduces an Orthogonal Feature Disentanglement Module (OFDM) to split bottleneck features into foreground/background branches, a Disentanglement Guidance Module (DGM) to inject these structural priors into the decoder, and a reliability map Mrel (Eq. 10) that weights pseudo-labels by their estimated structural consistency. The method is evaluated on ISIC-2016, Kvasir-SEG, Synapse, and ACDC under 5%/10%/20% labeled-data settings, reporting improvements over several published baselines. The paper also includes ablations and a quartile analysis suggesting that Mrel helps hardest samples.

Significance. If the reliability mechanism truly provides a structure-aware alternative to confidence-based pseudo-label weighting, the framework would be a meaningful advance for medical-image SSL. The two-stage training with semantic-frequency dual perturbations and the exploration of a teacher-free architecture are interesting. The paper also evaluates on four public datasets and includes a backbone ablation. However, the central claim is not currently supported: Mrel is constructed from the same network outputs that it is used to gate, and Eq. (7) makes the consistency term in Eq. (10) a self-comparison rather than an independent structural check. The reported gains are plausible but not statistically substantiated, and the DGM module is not specified to a reproducible level. The ideas are worth pursuing, but the evidence and presentation need substantial revision.

major comments (5)
  1. [§III-C, §III-D, Eq. (7) and Eq. (10)] Mrel = (1-|Pred-Pfg|)·|Pfg-Pbg| is used to gate pseudo-labels, but Pred is defined in Eq. (7) as Conv(P_mlp ⊕ P_fg), so P_fg is an input to Pred. The first factor therefore compares Pred to one of its own inputs, and the fusion conv can make this factor close to 1 by construction. Since Pfg and Pbg are supervised by y and 1-y in Stage I, after training Pbg≈1-Pfg, so the second factor becomes approximately |2Pfg-1|, a confidence measure of the foreground branch. No experiment validates Mrel against true pseudo-label error, nor is it compared with a confidence-only reweighting. This is the load-bearing claim of the paper; please add a direct validation on labeled data (e.g., correlation of Mrel with pseudo-label correctness) and an ablation that replaces Mrel with a simple confidence map.
  2. [§III-D, Eq. (14)] Lrel = Σ W ⊙ BCE(Mrel, 1) / Σ W, with W = 1(Pred>τ or Pred<1-τ) ⊙ Mrel. Since W includes Mrel, setting Mrel ≡ 1 makes BCE(Mrel,1) = 0 and reduces W to the confidence mask, giving a trivial minimum. This could collapse the reliability mechanism and make Mrel vacuous. Please detach Mrel in the weighting term or add a non-triviality constraint, and report the distribution of Mrel values during training to show that collapse does not occur.
  3. [§III-C, DGM] DGM is described only in prose: it 'maps' F_fg and F_bg to each scale and uses 'deformable guidance' to produce F_i_out, but no equations specify the offset generation, the scale-mapping operation, or how the structural priors modulate the decoder features. As DGM is a core contribution and is claimed to improve discrimination, the lack of formal specification prevents reproducibility. Please provide a precise module definition with tensor operations and dimensions.
  4. [§IV, Tables I–IV] All reported results are single-run numbers with no standard deviations or significance tests, and baselines are quoted from prior papers rather than reproduced under identical settings. The ablation is limited to one dataset (ISIC-2016 at 20% labels), and Table IV is formatted so poorly that individual rows cannot be reliably attributed to specific model variants. This makes it impossible to verify the claimed contributions of OFDM, Mrel, DHF, and DP. Please provide multiple seeds with error bars, statistical comparisons, and a clearly labeled ablation table with explicit row descriptions.
  5. [§III-D, multi-class extension] For multi-class segmentation, only M^k_rel is defined in Eq. (18); pseudo-label generation, confidence filtering, and the exact weighting procedure for the multi-class case are described as 'analogous' but not specified. Since Synapse and ACDC are central experimental results, please provide the full multi-class pseudo-label and weighting equations, or implement the reliability mechanism in a way that is fully specified for the multi-class setting.
minor comments (5)
  1. [Abstract and Table III] The abstract states that OFD-Net exceeds KnowSAM by 0.86 Dice and 4.59 Jaccard on ACDC, but Table III shows KnowSAM has better HD95 (1.28 vs 1.36) and ASD (0.36 vs 0.78). Please qualify the claim or discuss boundary trade-offs.
  2. [§IV-B and Eq. (24)] The text in §IV-B refers to a 'sigmoid ramp-up schedule,' but Eq. (24) is a Gaussian-style ramp-up λ(t)=λmax·e^{-5(1-t/tmax)^2}. Please correct the inconsistency.
  3. [Table IV] The table header appears as 'Baseline OFDM DHF DPM rel' without proper spacing, and the row/column alignment is broken. This makes even the component labels hard to read.
  4. [§III-E, Eq. (20)] The frequency perturbation uses a 'learnable transformation G(·)' but its architecture is not described. Please specify G or cite a reference for the transformation.
  5. [Declarations] No code or trained models are provided. Given the reproducibility concerns, releasing code would be helpful.

Circularity Check

0 steps flagged

No significant circularity: the reported gains are benchmark-measured; the reliability map is self-referential by design but not a fitted parameter renamed as a prediction.

full rationale

The paper's claimed derivation chain is not one that reduces to its inputs. The reliability map Mrel = (1 - |Pred - Pfg|) * |Pfg - Pbg| (Eq. 10) is a heuristic constructed from the network's own outputs, and in Stage II the auxiliary foreground/background branches are themselves supervised with the same reliability-weighted pseudo-labels (Eq. 13). This is intrinsic self-training feedback and a genuine validation concern, but it is not a case of fitting a parameter to a subset and then 'predicting' a closely related quantity. The paper's central evidence consists of external held-out metrics on ISIC-2016, Kvasir-SEG, Synapse, and ACDC, plus ablations; no parameter is fitted to those test sets and then reported as a prediction. There are no load-bearing self-citations, no imported uniqueness theorem, and no ansatz smuggled in via prior work by the same authors. The Discussion explicitly concedes the key limitation: 'the effectiveness of the proposed reliability mechanism depends on the quality of the learned foreground-background disentanglement; if the bottleneck features are not sufficiently separated, the reliability map may become less discriminative' (Section V). That is an acknowledged empirical conditionality, not an admission that a result equals its input by construction. Under the requirement to exhibit an equation-level reduction or a fitted-parameter-renamed-as-prediction, no circular step is present.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 3 invented entities

The central claim depends on hand-set hyperparameters (lambda_orth, lambda_cons, lambda_max, tau), an assumed link between per-pixel orthogonality and reliable separation, a meaningful OVR decomposition for multi-class organs, and unstated DGM machinery. The invented computational modules have only internal ablation support.

free parameters (4)
  • lambda_orth = 0.1
    Weight on the orthogonality loss; chosen by hand, with sensitivity tested only on ISIC-2016 over 0.08-0.20 in Fig. 8.
  • lambda_cons = 0.1
    Weight on the labeled consistency loss in Eqs. (21) and (25); set by hand, no sensitivity study reported.
  • lambda_max = 0.1
    Maximum coefficient for reliability-weighted unlabeled loss in Eq. (24); chosen by hand.
  • tau = 0.95
    Confidence threshold for bidirectional pseudo-label filtering in Eq. (11); tested across 0.80-0.99 on only one dataset.
axioms (4)
  • domain assumption Per-pixel orthogonality between Ffg and Fbg at the bottleneck implies clean foreground/background separation at the prediction level.
    Used in Eq. (3)/(6); the ablation replacing L_orth with FDM shows small drops, but there is no independent evidence that orthogonality is necessary or sufficient for reliability.
  • domain assumption Foreground/background one-vs-rest decomposition is meaningful for multi-class organs with a shared encoder.
    Section III-B and III-D; class-wise branches are described but no per-class validation or class-wise reliability analysis is provided for Synapse or ACDC.
  • domain assumption Semantic-frequency dual perturbations expose the model to the variation present in unlabeled data without destroying target structure.
    Section III-E; standard SSL augmentation assumption, asserted but not independently verified here.
  • ad hoc to paper DGM's deformable guidance can be implemented as described.
    Section III-C names deformable guidance but provides no equation for offsets, scale mapping, or modulation; reproducibility depends on an unstated implementation.
invented entities (3)
  • OFDM (Orthogonal Feature Disentanglement Module) no independent evidence
    purpose: Decomposes deepest encoder features into foreground and background branches under an orthogonality loss.
    Support is internal ablation only; the intermediate disentangled features are not externally validated.
  • DGM (Disentanglement Guidance Module) no independent evidence
    purpose: Injects disentangled foreground/background structure into the decoder via deformable guidance.
    No explicit formulation is given, so its claimed effect cannot be independently reproduced from the text.
  • Mrel reliability map no independent evidence
    purpose: Spatially weights pseudo-label supervision by structural consistency and separability.
    The map is derived from the model's own outputs and is not validated against an external measure of pseudo-label noise.

pith-pipeline@v1.3.0-alltime-deepseek · 20673 in / 14146 out tokens · 129862 ms · 2026-08-01T20:09:57.598165+00:00 · methodology

0 comments
read the original abstract

Semi-supervised learning (SSL) is an effective solution for medical image segmentation with limited annotations. Existing SSL methods mainly rely on pseudo-labels generated by teacher-student supervision or cross-network consistency. However, these methods lack an explicit structural reference for judging pseudo-label quality. Low-quality pseudo-labels may lead to unreliable training, error accumulation and confirmation bias when processing unlabeled data with substantial appearance variations. To address this issue, we proposed OFD-Net, a teacher-free single-network framework for reliable semi-supervised medical image segmentation. OFD-Net employs an Orthogonal Feature Disentanglement Module (OFDM) to capture OFD features for reliable SSL by disentangling unlabeled data into background and foreground representations with a reliable structural distribution, thereby effectively reducing error accumulation and alleviating confirmation bias among unlabeled data. Specifically, OFD-Net explicitly employs a Disentanglement Guidance Module (DGM) to inject the resulting structural priors of foreground-background into the decoder by deformable convolution processing, and outputs predictions with clearer foreground representations. Based on DGM and the OFDM, we further develop a reliability-aware pseudo-label learning mechanism that evaluates unlabeled supervision according to the structural consistency between the main prediction and the disentangled foreground-background responses, and then down-weights unreliable regions during training. Extensive experiments on four public medical image segmentation benchmarks, namely ISIC-2016, Kvasir-SEG, Synapse, and ACDC, validate the effectiveness of OFD-Net. These results confirm that orthogonal foreground-background disentanglement enables OFD-Net to establish an efficient and reliable training paradigm within a teacher-free single-network framework.

Figures

Figures reproduced from arXiv: 2607.16705 by Chen Yi, Cong-xuan zhang, Huan-huan Shi, Qin Lu, Shao-feng Jiang, Zhen Chen, Zhe-yang Jing.

Figure 1
Figure 1. Figure 1: Conceptual comparison between conventional semi-supervised seg [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall architecture of the proposed single-network OFD-Net, built on a SegFormer-B4 backbone with orthogonal foreground-background [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Visual results of different methods on skin lesion and polyp segmentation with 10% labeled data. The first row shows a tiny skin lesion case, and the [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visual comparison of multi-organ segmentation results on the Synapse dataset. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visualization of reliability maps with and without the orthogonal [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Performance improvements after introducing [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Visualization of response maps under different perturbation settings [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Sensitivity analysis of the pseudo-label confidence threshold and the [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

40 extracted references · 3 linked inside Pith

  1. [1]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” inInternational Conference on Medical image computing and computer-assisted intervention. Springer, 2015, pp. 234–241

    2015

  2. [2]

    Transunet: Transformers make strong encoders for medical image segmentation,

    J. Chen, Y . Lu, Q. Yu, X. Luo, E. Adeli, Y . Wang, L. Lu, A. L. Yuille, and Y . Zhou, “Transunet: Transformers make strong encoders for medical image segmentation,”arXiv preprint arXiv:2102.04306, 2021

    Pith/arXiv arXiv 2021

  3. [3]

    Swin-unet: Unet-like pure transformer for medical image segmenta- tion,

    H. Cao, Y . Wang, J. Chen, D. Jiang, X. Zhang, Q. Tian, and M. Wang, “Swin-unet: Unet-like pure transformer for medical image segmenta- tion,” inEuropean conference on computer vision. Springer, 2022, pp. 205–218

    2022

  4. [4]

    nn- former: V olumetric medical image segmentation via a 3d transformer,

    H.-Y . Zhou, J. Guo, Y . Zhang, X. Han, L. Yu, L. Wang, and Y . Yu, “nn- former: V olumetric medical image segmentation via a 3d transformer,” IEEE Transactions on Image Processing, vol. 32, pp. 4036–4045, 2023

    2023

  5. [5]

    Semi-supervised semantic segmentation with cross pseudo supervision,

    X. Chen, Y . Yuan, G. Zeng, and J. Wang, “Semi-supervised semantic segmentation with cross pseudo supervision,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2021, pp. 2613–2622

    2021

  6. [6]

    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,”Advances in neural information processing systems, vol. 30, 2017

    2017

  7. [7]

    Enhancing pseudo label quality for semi-supervised domain-generalized medical image segmentation,

    H. Yao, X. Hu, and X. Li, “Enhancing pseudo label quality for semi-supervised domain-generalized medical image segmentation,” in Proceedings of the AAAI conference on artificial intelligence, vol. 36, no. 3, 2022, pp. 3099–3107

    2022

  8. [8]

    Semi-supervised left atrium segmentation with mutual consistency training,

    Y . Wu, M. Xu, Z. Ge, J. Cai, and L. Zhang, “Semi-supervised left atrium segmentation with mutual consistency training,” inInternational Confer- ence on Medical Image Computing and Computer-Assisted Intervention. Springer, 2021, pp. 297–306. 14

    2021

  9. [9]

    Mcf: Mutual correction framework for semi-supervised medical image segmentation,

    Y . Wang, B. Xiao, X. Bi, W. Li, and X. Gao, “Mcf: Mutual correction framework for semi-supervised medical image segmentation,” inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2023, pp. 15 651–15 660

    2023

  10. [10]

    Pseudo-labeling and confirmation bias in deep semi-supervised learn- ing,

    E. Arazo, D. Ortego, P. Albert, N. E. O’Connor, and K. McGuinness, “Pseudo-labeling and confirmation bias in deep semi-supervised learn- ing,” inInternational Joint Conference on Neural Networks (IJCNN), 2020, pp. 1–8

    2020

  11. [11]

    Disentangled repre- sentation learning in cardiac image analysis,

    A. Chartsias, T. Joyce, G. Papanastasiou, S. Semple, M. Williams, D. E. Newby, R. Dharmakumar, and S. A. Tsaftaris, “Disentangled repre- sentation learning in cardiac image analysis,”Medical Image Analysis, vol. 58, p. 101535, 2019

    2019

  12. [12]

    Pnp-adanet: Plug-and-play adversarial domain adaptation network at unpaired cross-modality cardiac segmentation,

    Q. Dou, C. Ouyang, C. Chen, H. Chen, B. Glocker, X. Zhuang, and P.-A. Heng, “Pnp-adanet: Plug-and-play adversarial domain adaptation network at unpaired cross-modality cardiac segmentation,”Ieee Access, vol. 7, pp. 99 065–99 076, 2019

    2019

  13. [13]

    Adaptfrcnet: Semi-supervised adaptation of pre-trained model with frequency and region consistency for medical image segmentation,

    A. He, Y . Wu, Z. Wang, T. Li, and H. Fu, “Adaptfrcnet: Semi-supervised adaptation of pre-trained model with frequency and region consistency for medical image segmentation,”Medical Image Analysis, vol. 103, p. 103626, 2025

    2025

  14. [14]

    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,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2023, pp. 11 514–11 524

    2023

  15. [15]

    Realistic evaluation of deep semi-supervised learning algorithms,

    A. Oliver, A. Odena, C. A. Raffel, E. D. Cubuk, and I. Goodfellow, “Realistic evaluation of deep semi-supervised learning algorithms,” in Advances in Neural Information Processing Systems, S. Bengio, H. Wal- lach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett, Eds., vol. 31. Curran Associates, Inc., 2018

    2018

  16. [16]

    When confidence fails: Revisiting pseudo-label selection in semi-supervised semantic segmentation,

    P. Liu and J. Liu, “When confidence fails: Revisiting pseudo-label selection in semi-supervised semantic segmentation,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 21 874–21 884

    2025

  17. [17]

    Semi-supervised semantic segmen- tation with cross-consistency training,

    Y . Ouali, C. Hudelot, and M. Tami, “Semi-supervised semantic segmen- tation with cross-consistency training,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2020

    2020

  18. [18]

    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,”Proceedings of the AAAI Conference on Artificial Intelligence, vol. 35, no. 10, pp. 8801–8809, May 2021

    2021

  19. [19]

    Semi-supervised medical image segmentation using cross-model pseudo-supervision with shape aware- ness and local context constraints,

    J. Liu, C. Desrosiers, and Y . Zhou, “Semi-supervised medical image segmentation using cross-model pseudo-supervision with shape aware- ness and local context constraints,” inMedical Image Computing and Computer Assisted Intervention – MICCAI 2022: 25th International Conference, Singapore, September 18–22, 2022, Proceedings, Part VIII. Berlin, Heidelberg:...

    2022

  20. [20]

    Efficient semi-supervised gross target volume of nasopharyngeal carcinoma segmentation via uncertainty rectified pyramid consistency,

    X. Luo, W. Liao, J. Chen, T. Song, Y . Chen, S. Zhang, N. Chen, G. Wang, and S. Zhang, “Efficient semi-supervised gross target volume of nasopharyngeal carcinoma segmentation via uncertainty rectified pyramid consistency,” inMedical Image Computing and Computer Assisted Intervention – MICCAI 2021, M. de Bruijne, P. C. Cattin, S. Cotin, N. Padoy, S. Speide...

    2021

  21. [21]

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

    S. Li, C. Zhang, and X. He, “Shape-aware semi-supervised 3d seman- tic segmentation for medical images,” inMedical Image Computing and Computer Assisted Intervention – MICCAI 2020, A. L. Martel, P. Abolmaesumi, D. Stoyanov, D. Mateus, M. A. Zuluaga, S. K. Zhou, D. Racoceanu, and L. Joskowicz, Eds. Cham: Springer International Publishing, 2020, pp. 552–561

    2020

  22. [22]

    St++: Make self- training work better for semi-supervised semantic segmentation,

    L. Yang, W. Zhuo, L. Qi, Y . Shi, and Y . Gao, “St++: Make self- training work better for semi-supervised semantic segmentation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2022, pp. 4268–4277

    2022

  23. [23]

    Learnable prompting sam-induced knowledge distillation for semi- supervised medical image segmentation,

    K. Huang, T. Zhou, H. Fu, Y . Zhang, Y . Zhou, C. Gong, and D. Liang, “Learnable prompting sam-induced knowledge distillation for semi- supervised medical image segmentation,”IEEE Transactions on Medical Imaging, vol. 44, no. 5, pp. 2295–2306, 2025

    2025

  24. [24]

    Allspark: Reborn labeled features from unlabeled in transformer for semi-supervised semantic segmenta- tion,

    H. Wang, Q. Zhang, Y . Li, and X. Li, “Allspark: Reborn labeled features from unlabeled in transformer for semi-supervised semantic segmenta- tion,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 3627–3636

    2024

  25. [25]

    A semantic knowledge complementarity based decoupling framework for semi-supervised class-imbalanced medical image segmentation,

    Z. Zhang, G. Yin, B. Zhang, W. Liu, X. Zhou, and W. Wang, “A semantic knowledge complementarity based decoupling framework for semi-supervised class-imbalanced medical image segmentation,” inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2025, pp. 25 940–25 949

    2025

  26. [26]

    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,” inMedical Image Computing and Computer Assisted Intervention – MICCAI 2019, D. Shen, T. Liu, T. M. Peters, L. H. Staib, C. Essert, S. Zhou, P.-T. Yap, and A. Khan, Eds. Cham: Springer International Publishing, 2019, p...

    2019

  27. [27]

    Daw: Exploring the better weighting function for semi-supervised semantic segmentation,

    R. Sun, H. Mai, T. Zhang, and F. Wu, “Daw: Exploring the better weighting function for semi-supervised semantic segmentation,” inAd- vances in Neural Information Processing Systems, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, Eds., vol. 36. Curran Associates, Inc., 2023, pp. 61 792–61 805

    2023

  28. [28]

    Rethinking semi-supervised medical image segmentation: A variance-reduction perspective,

    C. You, W. Dai, Y . Min, F. Liu, D. Clifton, S. K. Zhou, L. Staib, and J. Duncan, “Rethinking semi-supervised medical image segmentation: A variance-reduction perspective,” inAdvances in Neural Information Processing Systems, A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, Eds., vol. 36. Curran Associates, Inc., 2023, pp. 9984–10 021

    2023

  29. [29]

    Meta pseudo labels,

    H. Pham, Z. Dai, Q. Xie, and Q. V . Le, “Meta pseudo labels,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2021, pp. 11 557–11 568

    2021

  30. [30]

    Metassl: A general heterogeneous loss for semi-supervised medical image segmentation,

    W. Zhao, L. Zhong, X. Liao, W. Liao, S. Zhang, S. Zhang, and G. Wang, “Metassl: A general heterogeneous loss for semi-supervised medical image segmentation,”IEEE Transactions on Medical Imaging, pp. 1– 1, 2025

    2025

  31. [31]

    In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection frame- work for semi-supervised learning,

    M. N. Rizve, K. Duarte, Y . S. Rawat, and M. Shah, “In defense of pseudo-labeling: An uncertainty-aware pseudo-label selection frame- work for semi-supervised learning,”ArXiv, vol. abs/2101.06329, 2021

    Pith/arXiv arXiv 2021

  32. [32]

    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,” inAdvances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M. Balcan, and H. Lin, Eds., vol. 33. Curran Associates, Inc., 2020, pp. 596–608

    2020

  33. [33]

    D. A. Gutman, N. C. F. Codella, M. E. Celebi, B. Helba, M. A. Marchetti, N. K. Mishra, and A. Halpern, “Skin lesion analysis toward melanoma detection: A challenge at the international symposium on biomedical imaging (ISBI) 2016, hosted by the international skin imaging collaboration (ISIC),”CoRR, vol. abs/1605.01397, 2016

    Pith/arXiv arXiv 2016

  34. [34]

    Kvasir-seg: A segmented polyp dataset,

    D. Jha, P. H. Smedsrud, M. A. Riegler, P. Halvorsen, T. de Lange, D. Johansen, and H. D. Johansen, “Kvasir-seg: A segmented polyp dataset,” inMultiMedia Modeling, Y . M. Ro, W.-H. Cheng, J. Kim, W.- T. Chu, P. Cui, J.-W. Choi, M.-C. Hu, and W. De Neve, Eds. Cham: Springer International Publishing, 2020, pp. 451–462

    2020

  35. [35]

    Miccai multi-atlas labeling beyond the cranial vault–workshop and challenge,

    B. Landman, Z. Xu, J. Igelsias, M. Styner, T. Langerak, and A. Klein, “Miccai multi-atlas labeling beyond the cranial vault–workshop and challenge,” inProc. MICCAI multi-atlas labeling beyond cranial vault—workshop challenge, vol. 5. Munich, Germany, 2015, p. 12

    2015

  36. [36]

    Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved?

    O. Bernard, A. Lalande, C. Zotti, F. Cervenansky, X. Yang, P.-A. Heng, I. Cetin, K. Lekadir, O. Camara, M. A. G. Ballesteret al., “Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved?”IEEE transactions on medical imaging, vol. 37, no. 11, pp. 2514–2525, 2018

    2018

  37. [37]

    Segformer: Simple and efficient design for semantic segmentation with transformers,

    E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, and P. Luo, “Segformer: Simple and efficient design for semantic segmentation with transformers,” inAdvances in Neural Information Processing Systems, M. Ranzato, A. Beygelzimer, Y . Dauphin, P. Liang, and J. W. Vaughan, Eds., vol. 34. Curran Associates, Inc., 2021, pp. 12 077–12 090

    2021

  38. [38]

    Dmt: Dynamic mutual training for semi-supervised learning,

    Z. Feng, Q. Zhou, Q. Gu, X. Tan, G. Cheng, X. Lu, J. Shi, and L. Ma, “Dmt: Dynamic mutual training for semi-supervised learning,”Pattern Recognition, vol. 130, p. 108777, 2022

    2022

  39. [39]

    Frcnet: Frequency and re- gion consistency for semi-supervised medical image segmentation,

    A. He, T. Li, Y . Wu, K. Zou, and H. Fu, “Frcnet: Frequency and re- gion consistency for semi-supervised medical image segmentation,” inMedical Image Computing and Computer Assisted Intervention – MICCAI 2024, M. G. Linguraru, Q. Dou, A. Feragen, S. Giannarou, B. Glocker, K. Lekadir, and J. A. Schnabel, Eds. Cham: Springer Nature Switzerland, 2024, pp. 305–315

    2024

  40. [40]

    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 International conference on medical image computing and computer- assisted intervention. Springer, 2022, pp. 34–43

    2022