Pith. sign in

REVIEW 3 major objections 4 minor 39 references

Training-only module lifts medical image segmentation by modeling anatomical variation

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 · glm-5.2

2026-07-09 11:25 UTC pith:TGDLCLYV

load-bearing objection VCDP: Variation-Conditioned Distributional Proxy Learning for Semi-Supervised Medical Image Segmentation the 3 major comments →

arxiv 2607.07416 v1 pith:TGDLCLYV submitted 2026-07-08 cs.CV

VCDP: Variation-Conditioned Distributional Proxy Learning for Semi-Supervised Medical Image Segmentation

Zimu Zhang , Yiheng Zhong , Zhuoru Zhang , Yingzhen Hu , Yanan He , Fanliang Meng , Xiaofeng Liu This is my paper
classification cs.CV
keywords semi-supervised segmentationmedical image segmentationproxy learningintra-class variationfeature-space regularizationdistributional proxyvariation prototype3D medical imaging
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.

The paper proposes VCDP, a plug-and-play training-only regularization module for semi-supervised 3D medical image segmentation. The core claim is that existing semi-supervised methods focus on prediction-level consistency but leave the feature space poorly organized, especially for small organs and structures with high anatomical variability across patients. VCDP addresses this by representing each anatomical class with two complementary objects: a learnable Gaussian distribution (the distributional proxy) that captures shared class-level semantics, and a set of learnable variation prototypes (five per class) that capture fine-grained intra-class sub-modes such as shape differences, boundary ambiguity, and local appearance changes. These are fused into a single variation-conditioned compatibility score that guides every voxel embedding toward both the correct global organ identity and the most relevant local variation pattern. The module is attached to decoder features during training and removed entirely at inference, adding no computational cost at test time. The paper demonstrates that this feature-space regularization improves most evaluated baselines on Synapse and AMOS benchmarks, with the largest gains on small, ambiguous, and highly variable organs such as the pancreas, esophagus, and adrenal glands.

Core claim

The central mechanism is the variation-conditioned compatibility score (Eq. 6), which fuses distributional similarity (cosine similarity to stochastically sampled Gaussian proxy representatives) with variation-aware similarity (a log-sum-exp soft aggregation over class-specific variation prototypes). This score produces class-wise soft assignments for all voxel embeddings, providing dense feature-space regularization on both labeled and unlabeled data. The paper shows that the Gaussian proxy and variation prototypes are complementary: each alone improves over baselines, but together they yield the best results, confirming that modeling both global class semantics and fine-grained intra-class

What carries the argument

Variation-conditioned compatibility score fusing Gaussian distributional proxy similarity with soft variation prototype aggregation; stop-gradient on proxy means to prevent drift; labeled proxy calibration path aligning Gaussian means with labeled feature anchors

Load-bearing premise

The paper assumes that the learned Gaussian proxy means and variation prototypes, trained on limited labeled and unlabeled data, generalize to unseen anatomical variations at test time despite being removed from the model at inference, relying entirely on the backbone having internalized their feature-space organization.

What would settle it

If attaching VCDP to an already well-organized feature space (e.g., a fully supervised model with strong contrastive pretraining) produced no improvement, the module's benefit would be limited to compensating for insufficient baseline feature organization rather than providing fundamentally new representational capacity.

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

If this is right

  • VCDP could be applied beyond medical imaging to any dense prediction task with high intra-class variation and limited labels, such as satellite imagery or autonomous driving scene parsing.
  • The variation prototype mechanism suggests that fixed K=5 prototypes per class may be suboptimal; adaptive or hierarchical prototype counts could yield further gains for classes with particularly complex sub-structures.
  • Since the module is training-only and architecture-agnostic, it could be stacked with other feature-space regularization methods, though interaction effects remain untested.
  • The soft assignment approach provides dense feature supervision on unlabeled data without pseudo-labeling, which could complement rather than replace existing pseudo-label strategies.
  • The principle of decoupling global class identity from local variation patterns could inform representation learning in few-shot and continual learning settings where intra-class diversity is high.

Where Pith is reading between the lines

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

  • The fact that VCDP produces the largest gains on the weakest baselines (e.g., +20 Dice for SS-Net on AMOS) suggests the module compensates for missing feature-space organization; well-organized feature spaces may see diminishing returns, implying a ceiling effect the paper does not address.
  • The removal of all auxiliary components at inference means the backbone must internalize the feature-space structure indirectly through gradients; whether the learned representations are as robust as they would be with persistent prototypes at test time is an open question the paper does not answer.
  • The variation prototypes are learned but not analyzed for semantic interpretability; one could test whether they correspond to recognizable anatomical sub-modes (e.g., different pancreas shapes) by clustering held-out embeddings and examining the resulting groups.

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

3 major / 4 minor

Summary. The paper proposes Variation-Conditioned Distributional Proxy Learning (VCDP), a training-only regularization module for semi-supervised 3D medical image segmentation. VCDP attaches to decoder features and represents each anatomical class using a learnable Gaussian proxy (for shared semantics) and multiple variation prototypes (for intra-class structural sub-modes). A unified compatibility score (Eq. 6) fuses distributional similarity and variation-aware similarity to guide voxel embeddings via soft assignments (Eqs. 8-12), while a calibration loss (Eq. 15) stabilizes the Gaussian means using labeled data. The module is removed at inference, introducing no additional test-time cost. Experiments on Synapse and AMOS show improvements when VCDP is plugged into several semi-supervised baselines (CPS, MagicNet, DHC, GenSSL, SS-Net, Adsh, DCMamba).

Significance. The paper addresses a relevant problem of feature-space organization for anatomically variable structures in semi-supervised medical image segmentation. The mathematical formulation (Eqs. 1-16) is internally consistent, and the module is designed to be plug-and-play with zero inference cost, which is a practical strength. The ablation study (Table IV) cleanly isolates the contributions of the Gaussian and variation components. Anonymous code is provided, which aids reproducibility. However, the empirical validation has significant shortcomings that undermine the central claims, as detailed below.

major comments (3)
  1. Tables I-III: No error bars, standard deviations, or significance tests are reported for any experiment. The Synapse test set contains only 6 scans (Section IV-A). With N=6, Dice variance across cases is typically large, and gains of 0.47-1.25 Dice (e.g., MagicNet on AMOS, CPS on Synapse) could easily fall within noise. Without multiple seeds and significance testing, the broad improvement claim is not adequately substantiated.
  2. Tables II and III: Several organ-level regressions are severe and undiscussed. On Synapse (Table II), SS-Net+VCDP drops LK from 43.32 to 26.61, and DHC+VCDP drops St from 35.90 to 32.00 and IVC from 55.00 to 23.30. On AMOS (Table III), DCMamba+VCDP drops LAG from 14.26 to 0.00. These regressions undermine the claim that VCDP provides complementary feature-space supervision that broadly improves representation learning, and they must be explicitly analyzed.
  3. Table I and Section IV-B: The largest gains are concentrated on the weakest baselines (SS-Net +20.39, GenSSL +14.29 on AMOS), while two baselines on AMOS (DHC and DCMamba) show net Dice decreases. The paper does not discuss whether these large gains reflect the specific variation-conditioned representation learning claimed, or merely a general training stabilization effect that any auxiliary feature regularization could provide. A comparison with a simpler prototype-based regularization baseline would help isolate the contribution of the variation-conditioned design.
minor comments (4)
  1. Section III-B, Eq. (2): The text states that sigma_c is 'kept fixed during training,' but it is listed as a learnable parameter in the formulation. Clarify whether sigma_c is learned or fixed, and if fixed, state its value.
  2. Figure 2: The text in the figure panels is small and difficult to read. Consider enlarging the labels for the Gaussian mean, variation prototypes, and compatibility score components.
  3. Section IV-A: The values of hyperparameters K, S, tau, lambda_var, lambda_reg, and lambda_cal are not specified in the main text. These should be reported for reproducibility, even if they are available in the code.
  4. Table I: The HD95 for MagicNet on Synapse is reported as 1.64, which seems unusually low for this dataset. Please verify this value.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful and constructive review. The referee raises three major concerns: (1) lack of error bars, standard deviations, or significance tests given the small Synapse test set (N=6); (2) severe per-organ regressions in Tables II–III that are undiscussed; and (3) concentration of gains on weak baselines without a simpler prototype-based baseline comparison to isolate the variation-conditioned design. We agree with all three points and will revise the manuscript accordingly. Specifically, we will add multi-seed experiments with standard deviations and significance tests, explicitly analyze per-organ regressions (including potential failure modes of the variation prototype mechanism), and include a comparison with a single-prototype regularization baseline. We believe these revisions substantially strengthen the empirical validation.

read point-by-point responses

  1. Referee: Tables I-III: No error bars, standard deviations, or significance tests are reported for any experiment. The Synapse test set contains only 6 scans (Section IV-A). With N=6, Dice variance across cases is typically large, and gains of 0.47-1.25 Dice could easily fall within noise. Without multiple seeds and significance testing, the broad improvement claim is not adequately substantiated.

    Authors: The referee is correct. The absence of standard deviations and significance tests is a genuine weakness, particularly given the small Synapse test set (N=6). We will address this by running all experiments with at least three independent random seeds and reporting mean ± standard deviation for all metrics in Tables I–III. We will also include paired t-tests (or Wilcoxon signed-rank tests where normality assumptions are questionable given N=6) to assess statistical significance. We acknowledge that some of the smaller gains (e.g., MagicNet +0.47 on AMOS, CPS +1.25 on Synapse) may not reach statistical significance, and we will state this transparently. For the larger gains (e.g., SS-Net +20.39 on AMOS, DCMamba +9.68 on Synapse), we expect these to remain significant, but we will let the tests speak for themselves. The revised manuscript will qualify claims of 'broad improvement' where significance is not established. revision: yes

  2. Referee: Tables II and III: Several organ-level regressions are severe and undiscussed. On Synapse (Table II), SS-Net+VCDP drops LK from 43.32 to 26.61, and DHC+VCDP drops St from 35.90 to 32.00 and IVC from 55.00 to 23.30. On AMOS (Table III), DCMamba+VCDP drops LAG from 14.26 to 0.00. These regressions undermine the claim that VCDP provides complementary feature-space supervision that broadly improves representation learning, and they must be explicitly analyzed.

    Authors: We agree that these regressions should have been discussed. We will add a dedicated subsection analyzing per-organ failures. Our preliminary investigation suggests two failure modes: (1) When a baseline already produces very low scores for an organ (e.g., DHC's IVC at 55.00 or DCMamba's LAG at 14.26), the variation prototypes may capture spurious patterns from poorly represented or near-absent classes in the labeled mini-batch, causing the soft assignment to misroute embeddings. (2) For organs with extreme anatomical variability (e.g., stomach, IVC), the fixed number of variation prototypes (K=5) may be insufficient or may compete with the Gaussian proxy's distributional signal, creating conflicting gradients. The LAG regression to 0.00 in DCMamba+VCDP is particularly concerning and likely indicates that the variation prototypes for LAG collapsed to a degenerate solution, since DCMamba's baseline LAG performance (14.26) was already marginal. We will add this analysis to the revised manuscript and discuss potential mitigations, such as class-adaptive prototype counts or prototype reset mechanisms, as future work. We will also temper the claim of 'broad improvement' to acknowledge that VCDP can introduce regressions on specific organs when baseline performance is already low. revision: yes

  3. Referee: Table I and Section IV-B: The largest gains are concentrated on the weakest baselines (SS-Net +20.39, GenSSL +14.29 on AMOS), while two baselines on AMOS (DHC and DCMamba) show net Dice decreases. The paper does not discuss whether these large gains reflect the specific variation-conditioned representation learning claimed, or merely a general training stabilization effect that any auxiliary feature regularization could provide. A comparison with a simpler prototype-based regularization baseline would help isolate the contribution of the variation-conditioned design.

    Authors: This is a fair and important point. The current ablation (Table IV) compares Gaussian-only, variation-only, and combined configurations, but does not include a comparison with a simpler single-prototype regularization baseline (i.e., one deterministic prototype per class with soft assignment, without distributional modeling or variation prototypes). We will add this baseline to the ablation study. Specifically, we will implement a single-prototype variant that replaces the Gaussian proxy and variation prototypes with a single learnable class center and uses the same soft assignment and calibration losses. This will allow us to isolate whether the gains come from the variation-conditioned design specifically or from generic feature-space regularization. We also agree that the concentration of gains on weak baselines needs discussion: VCDP provides the most benefit when the baseline feature space is poorly organized (as with SS-Net and GenSSL on AMOS), but offers diminishing returns—or can even interfere—when the baseline already has strong feature representations (as with DHC and DCMamba on AMOS). We will add this discussion to Section IV-B and frame it as a scope limitation of the method. revision: yes

Circularity Check

0 steps flagged

No circularity found: VCDP's compatibility score and losses are defined from learnable parameters optimized by external labeled-data anchors, not from their own outputs.

full rationale

The paper's central derivation chain is not circular. The variation-conditioned compatibility score (Eq. 6: g(z_i, c) = s_dist(z_i, c) + λ_var * s_var(z_i, c)) is computed from learnable parameters (μ_c, σ_c, v_{c,k}) and voxel embeddings z_i, all of which are optimized via the alignment loss (Eq. 9), discrimination loss (Eq. 11), and calibration loss (Eq. 15). The calibration loss uses labeled voxel feature anchors a_c (Eq. 14) as an external signal — these are computed from ground-truth labels y_i (Eq. 13), not from the model's own predictions. The stop-gradient on μ_c in the dense path (Eq. 7) explicitly prevents the dense regularization from circularly updating the Gaussian proxy means via unlabeled embeddings. The variation prototypes {v_{c,k}} are free learnable parameters updated by gradient descent on the alignment and discrimination objectives, not defined in terms of the compatibility score they feed into. No self-citation chain is load-bearing for the mathematical formulation: the method is self-contained in its equations. The concerns raised by the skeptic (small test sets, lack of significance testing, organ-level regressions) are correctness and empirical-validity issues, not circularity. The paper does not fit a parameter to data and then rename the fit as a prediction, nor does it define any quantity in terms of the result it claims to derive.

Axiom & Free-Parameter Ledger

9 free parameters · 4 axioms · 1 invented entities

The axiom ledger reveals that VCDP introduces at least 6 unspecified hyperparameters (S, τ, λ_var, λ_reg, λ_cal, σ_c) and one ad hoc choice (K=5) without sensitivity analysis. The variation prototypes are the primary invented entity, and the paper provides no independent evidence that they capture real anatomical sub-modes rather than serving as optimization aids. The core domain assumptions are reasonable but unvalidated.

free parameters (9)
  • K (number of variation prototypes per class) = 5
    Set to 5 in implementation (Section III-B). No justification or sensitivity analysis provided for this choice.
  • S (number of stochastic proxy samples) = Not reported in text
    Used in Eq. 3-4 for distributional similarity estimation. Value not specified in the paper.
  • τ (temperature for soft aggregation) = Not reported in text
    Controls sharpness in Eq. 5. Value not specified.
  • λ_var (variation balance weight) = Not reported in text
    Balances distributional and variation terms in Eq. 6. Value not specified.
  • λ_reg (regularization loss weight) = Not reported in text
    Weights L_reg in final objective Eq. 16. Value not specified.
  • λ_cal (calibration loss weight) = Not reported in text
    Weights L_cal in final objective Eq. 16. Value not specified.
  • σ_c (Gaussian dispersion scale) = Fixed, value not reported
    Stated as fixed during training (Section III-B) but the value is not given.
  • μ_c (Gaussian proxy means) = Learned during training
    Per-class learnable parameters, one per class per embedding dimension.
  • v_{c,k} (variation prototypes) = Learned during training
    K=5 per class, learnable embedding-space vectors.
axioms (4)
  • domain assumption Voxel embeddings from the decoder feature map can be meaningfully regularized via class-level proxies in the projected embedding space.
    Foundational premise of the entire approach (Section III-A). Assumes the projection head ϕ(·) preserves enough anatomical information for proxy-based regularization to be meaningful.
  • ad hoc to paper K=5 variation prototypes are sufficient to capture intra-class anatomical sub-modes.
    Section III-B fixes K=5 without justification. No sensitivity analysis or principled selection criterion is provided.
  • domain assumption The soft assignment q_i(c) derived from the model's own compatibility scores provides a useful supervision signal for embedding regularization.
    Eqs. 8-9 use the model's own compatibility scores as soft targets. This is a standard assumption in self-supervised learning but means the supervision is partially self-generated.
  • domain assumption Labeled voxel feature anchors a_c are reliable enough to calibrate Gaussian proxy means without introducing bias from small labeled sets.
    Eq. 14-15. With only 20% labeled data on Synapse (4 scans) and 5% on AMOS, the per-batch labeled anchors may be noisy or class-imbalanced.
invented entities (1)
  • Variation prototypes {v_{c,k}} no independent evidence
    purpose: Per-class learnable vectors that capture fine-grained intra-class structural sub-modes (e.g., local appearance changes, shape variations, boundary ambiguity).
    The paper does not provide external evidence that the learned prototypes correspond to interpretable anatomical sub-modes. No visualization, clustering analysis, or correlation with known anatomical variations is presented. The prototypes could be functioning as additional free parameters that improve optimization without capturing meaningful variation structure.

pith-pipeline@v1.1.0-glm · 17017 in / 3535 out tokens · 583338 ms · 2026-07-09T11:25:52.771874+00:00 · methodology

0 comments
read the original abstract

Semi-supervised 3D medical image segmentation reduces the need for dense voxel-level annotations by exploiting unlabeled volumes. Although existing methods such as consistency regularization, pseudo-labeling, and co-training improve prediction-level robustness, they often provide insufficient feature-space organization for anatomically complex structures, especially small organs and ambiguous boundary regions with large intra-class variations. To address this issue, we propose Variation-Conditioned Distributional Proxy Learning (VCDP), a plug-and-play training-only regularization module for semi-supervised 3D medical image segmentation. VCDP represents each class with a learnable Gaussian distribution for shared class semantics and multiple variation prototypes for fine-grained intra-class patterns. A unified variation-conditioned compatibility score is further formulated to fuse distributional similarity and soft variation aggregation, guiding voxel embeddings to align with both global organ identity and local anatomical variations. VCDP is attached to decoder features during training and removed during inference, introducing no additional inference cost. Experiments on multi-organ segmentation benchmarks show that VCDP improves most evaluated baselines, particularly for small, ambiguous, and highly variable organs. Our anonymous code is released at https://anonymous.4open.science/r/VCDP_code-41ED.

Figures

Figures reproduced from arXiv: 2607.07416 by Fanliang Meng, Xiaofeng Liu, Yanan He, Yiheng Zhong, Yingzhen Hu, Zhuoru Zhang, Zimu Zhang.

Figure 1
Figure 1. Figure 1: Motivation of the proposed VCDP framework. (a) Co [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed VCDP framework for semi-supervised medical image segmentation. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison on the 20% labeled Synapse [PITH_FULL_IMAGE:figures/full_fig_p007_3.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

39 extracted references · 39 canonical work pages · 7 internal anchors

  1. [1]

    Pixelseg: Pixel-by-pixel stochastic semantic segmentation for ambiguous medical images,

    W. Zhang, X. Zhang, S. Huang, Y . Lu, and K. Wang, “Pixelseg: Pixel-by-pixel stochastic semantic segmentation for ambiguous medical images,” inProceedings of the 30th ACM International Conference on Multimedia, 2022, pp. 4742–4750

    work page 2022

  2. [2]

    Mdf-net: A multi-scale dynamic fusion network for breast tumor segmentation of ultrasound images,

    W. Qi, H.-C. Wu, and S.-C. Chan, “Mdf-net: A multi-scale dynamic fusion network for breast tumor segmentation of ultrasound images,” IEEE Transactions on Image Processing, vol. 32, pp. 4842–4855, 2023

    work page 2023

  3. [3]

    Uncertainty-aware mean teacher framework with inception and squeeze-and-excitation block for miccai flare22 challenge,

    H. Meng, H. Zhao, Z. Yu, Q. Li, and J. Niu, “Uncertainty-aware mean teacher framework with inception and squeeze-and-excitation block for miccai flare22 challenge,” inMICCAI Challenge on Fast and Low- Resource Semi-supervised Abdominal Organ Segmentation. Springer, 2022, pp. 245–259

    work page 2022

  4. [4]

    Gcl: Gradient- guided contrastive learning for medical image segmentation with multi- perspective meta labels,

    Y . Wu, J. Chen, J. Yan, Y . Zhu, D. Z. Chen, and J. Wu, “Gcl: Gradient- guided contrastive learning for medical image segmentation with multi- perspective meta labels,” inProceedings of the 31st ACM International Conference on Multimedia, 2023, pp. 463–471

    work page 2023

  5. [5]

    Pg-sam: A fine-grained prior-guided sam framework for prompt-free medical image segmentation,

    Y . Zhong, Z. Luo, C. Liu, F. Tang, Y . Hu, Z. Peng, M. Hu, J. Su, Z. Ge, and I. Razzak, “Pg-sam: A fine-grained prior-guided sam framework for prompt-free medical image segmentation,” in2025 IEEE International Conference on Bioinformatics and Biomedicine (BIBM). IEEE, 2025, pp. 3369–3376

    work page 2025

  6. [6]

    Semi-supervised semantic segmentation via prototypical contrastive learning,

    Z. Chen and Z. Lian, “Semi-supervised semantic segmentation via prototypical contrastive learning,” inProceedings of the 30th ACM International Conference on Multimedia, 2022, pp. 6696–6705

    work page 2022

  7. [7]

    Rda: Reciprocal distribution alignment for robust semi-supervised learning,

    Y . Duan, L. Qi, L. Wang, L. Zhou, and Y . Shi, “Rda: Reciprocal distribution alignment for robust semi-supervised learning,” inEuropean Conference on Computer Vision. Springer, 2022, pp. 533–549

    work page 2022

  8. [8]

    Eye-guided dual-path network for multi-organ segmentation of abdomen,

    C. Wang, D. Zhang, and R. Ge, “Eye-guided dual-path network for multi-organ segmentation of abdomen,” inInternational Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 2023, pp. 23–32

    work page 2023

  9. [9]

    Semi-supervised learning for multi-label video action detection,

    H. Zhang, X. Zhao, and D. Wang, “Semi-supervised learning for multi-label video action detection,” inProceedings of the 30th ACM International Conference on Multimedia, 2022, pp. 2124–2134

    work page 2022

  10. [10]

    Virtual adversarial training: a regularization method for supervised and semi-supervised learning,

    T. Miyato, S.-i. Maeda, M. Koyama, and S. Ishii, “Virtual adversarial training: a regularization method for supervised and semi-supervised learning,”IEEE transactions on pattern analysis and machine intelli- gence, vol. 41, no. 8, pp. 1979–1993, 2018

    work page 1979

  11. [11]

    Regularization with stochastic transformations and perturbations for deep semi-supervised learning,

    M. Sajjadi, M. Javanmardi, and T. Tasdizen, “Regularization with stochastic transformations and perturbations for deep semi-supervised learning,”Advances in neural information processing systems, vol. 29, 2016

    work page 2016

  12. [12]

    Adaptive Bidirectional Displacement for Semi-Supervised Medical Image Segmentation

    H. Chi, J. Pang, B. Zhang, and W. Liu, “Adaptive bidirectional displacement for semi-supervised medical image segmentation,” 2024. [Online]. Available: https://arxiv.org/abs/2405.00378

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  13. [13]

    Fedmlp: Federated multi-label medical image classification under task heterogeneity,

    Z. Sun, N. Wu, J. Shi, L. Yu, X. Yang, K.-T. Cheng, and Z. Yan, “Fedmlp: Federated multi-label medical image classification under task heterogeneity,” 2024. [Online]. Available: https://arxiv.org/abs/2406. 18995

    work page 2024

  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,” 2023. [Online]. Available: https://arxiv.org/abs/2305.00673

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  15. [15]

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

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

    work page 2020

  16. [16]

    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,” inInternational conference on medical image computing and computer- assisted intervention. Springer, 2019, pp. 605–613

    work page 2019

  17. [17]

    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,” inInternational Conference on Med- ical Image Computing and Computer-Assisted Intervention. Springer, 2020, pp. 552–561

    work page 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,” inProceedings of the AAAI conference on artificial intelligence, vol. 35, no. 10, 2021, pp. 8801– 8809

    work page 2021

  19. [19]

    Dycon: Dynamic uncertainty-aware consistency and contrastive learning for semi-supervised medical image segmentation,

    M. Assefa, M. Naseer, I. I. Ganapathi, S. S. Ali, M. L. Seghier, and N. Werghi, “Dycon: Dynamic uncertainty-aware consistency and contrastive learning for semi-supervised medical image segmentation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

    work page 2025

  20. [20]

    Annotation ambiguity aware semi-supervised medical image segmentation,

    S. Kumari and P. Singh, “Annotation ambiguity aware semi-supervised medical image segmentation,” inProceedings of the computer vision and pattern recognition conference, 2025, pp. 10 404–10 413

    work page 2025

  21. [21]

    2015 miccai multi-atlas labeling beyond the cranial vault workshop and challenge,

    B. Landman, Z. Xu, J. E. Iglesias, M. Styner, T. R. Langerak, and A. Klein, “2015 miccai multi-atlas labeling beyond the cranial vault workshop and challenge,” inMICCAI Multi-Atlas Labeling Beyond Cranial Vault Workshop and Challenge, 2015

    work page 2015

  22. [22]

    Proxy anchor loss for deep metric learning,

    S. Kim, D. Kim, M. Cho, and S. Kwak, “Proxy anchor loss for deep metric learning,” 2020. [Online]. Available: https://arxiv.org/abs/2003. 13911

    work page 2020

  23. [23]

    Rethinking Semantic Segmentation: A Prototype View

    T. Zhou, W. Wang, E. Konukoglu, and L. V . Gool, “Rethinking semantic segmentation: A prototype view,” 2022. [Online]. Available: https://arxiv.org/abs/2203.15102

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  24. [24]

    Variation-aware proxy learning for semantic segmentation,

    H. Bae and B. C. Song, “Variation-aware proxy learning for semantic segmentation,”Neurocomputing, vol. 659, p. 131783, 2026. [Online]. Available: https://www.sciencedirect.com/science/article/pii/ S0925231225024555

    work page 2026

  25. [25]

    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,”

    work page

  26. [26]

    Exploring Smoothness and Class-Separation for Semi-supervised Medical Image Segmentation

    [Online]. Available: https://arxiv.org/abs/2203.01324

    work page internal anchor Pith review Pith/arXiv arXiv

  27. [27]

    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

    work page 2017

  28. [28]

    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

    work page 2021

  29. [29]

    Magicnet: Semi-supervised multi-organ segmentation via magic-cube partition and recovery,

    D. Chen, Y . Bai, W. Shen, Q. Li, L. Yu, and Y . Wang, “Magicnet: Semi-supervised multi-organ segmentation via magic-cube partition and recovery,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 23 869–23 878

    work page 2023

  30. [30]

    Towards generic semi-supervised framework for volumetric medical image segmentation,

    H. Wang and X. Li, “Towards generic semi-supervised framework for volumetric medical image segmentation,”Advances in Neural Informa- tion Processing Systems, vol. 36, pp. 1833–1848, 2023

    work page 2023

  31. [31]

    Dhc: Dual-debiased heterogeneous co-training framework for class-imbalanced semi-supervised medical image segmentation,

    ——, “Dhc: Dual-debiased heterogeneous co-training framework for class-imbalanced semi-supervised medical image segmentation,” inIn- ternational conference on medical image computing and computer- assisted intervention. Springer, 2023, pp. 582–591

    work page 2023

  32. [32]

    ProxyNCA++: Revisiting and Revitalizing Proxy Neighborhood Component Analysis

    E. W. Teh, T. DeVries, and G. W. Taylor, “Proxynca++: Revisiting and revitalizing proxy neighborhood component analysis,” 2020. [Online]. Available: https://arxiv.org/abs/2004.01113

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  33. [33]

    Seg- contrast: 3d point cloud feature representation learning through self- supervised segment discrimination,

    L. Nunes, R. Marcuzzi, X. Chen, J. Behley, and C. Stachniss, “Seg- contrast: 3d point cloud feature representation learning through self- supervised segment discrimination,”IEEE Robotics and Automation Letters, vol. 7, no. 2, pp. 2116–2123, 2022

    work page 2022

  34. [34]

    Disco- 3d: Discovering and segmenting sub-concepts from open-vocabulary queries in nerf,

    D. Petit, S. Bourgeois, V . Gay-Bellile, F. Chabot, and L. Barthe, “Disco- 3d: Discovering and segmenting sub-concepts from open-vocabulary queries in nerf,” inProceedings of the IEEE/CVF International Con- ference on Computer Vision, 2025, pp. 20 043–20 052

    work page 2025

  35. [35]

    Multi-Modal Prototypes for Open-World Semantic Segmentation

    Y . Yang, C. Ma, C. Ju, F. Zhang, J. Yao, Y . Zhang, and Y . Wang, “Multi-modal prototypes for open-world semantic segmentation,” 2024. [Online]. Available: https://arxiv.org/abs/2307.02003

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  36. [36]

    Class-imbalanced semi-supervised learning with adaptive thresholding,

    L.-Z. Guo and Y .-F. Li, “Class-imbalanced semi-supervised learning with adaptive thresholding,” inInternational conference on machine learning. PMLR, 2022, pp. 8082–8094

    work page 2022

  37. [37]

    Diversity-enhanced collaborative mamba for semi-supervised medical image segmentation,

    S. Li, J. Zhang, L. Qi, L. Zhou, Y . Shi, and Y . Gao, “Diversity-enhanced collaborative mamba for semi-supervised medical image segmentation,”

    work page

  38. [38]

    Diversity-enhanced Collaborative Mamba for Semi-supervised Medical Image Segmentation

    [Online]. Available: https://arxiv.org/abs/2508.13712

    work page internal anchor Pith review Pith/arXiv arXiv

  39. [39]

    Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation,

    Y . Ji, H. Bai, C. Ge, J. Yang, Y . Zhu, R. Zhang, Z. Li, L. Zhanng, W. Ma, X. Wanet al., “Amos: A large-scale abdominal multi-organ benchmark for versatile medical image segmentation,”Advances in neural information processing systems, vol. 35, pp. 36 722–36 732, 2022

    work page 2022