Pith. sign in

REVIEW 3 major objections 6 minor 40 references

Multimodal Prototype Alignment for Semi-supervised Pathology Image Segmentation

T0 review · 3 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read MPAMatch claims that adding pixel-level contrastive alignment to image and text prototypes improves semi-supervised pathology segmentation, reporting consistent gains over the strongest baselines on four public benchmarks.

desk verdict A plausible SSL/pathology segmentation integration whose central SOTA claim is currently under-supported: the same-backbone gains are small and no variance or protocol details are reported. read the letter →

arxiv 2508.19574 v1 pith:42YSIKR5 submitted 2025-08-27 cs.CV cs.AI

Mingxi Fu , Fanglei Fu , Xitong Ling , Huaitian Yuan , Tian Guan , Yonghong He , Lianghui Zhu This is my paper · ORCID
classification cs.CVcs.AI
keywords semi-supervisedsegmentationpathologyimageprototypelearningmultimodalalignmenttext-guidedconsistencyregularizationvision-languagemodelshistopathology
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper is trying to establish that semi-supervised pathology image segmentation gets better when unlabeled pixels are supervised not only by weak-to-strong consistency between augmented views, but also by class prototypes drawn from two modalities. On the image side, prototypes come from online clustering of pixel features; on the text side, they come from class descriptions generated by prompting a language model and embedded by a pathology-tuned vision-language model. Both prototype sets are aligned to pixel embeddings with contrastive losses, giving unlabeled pixels coarse structural guidance and semantic boundary guidance. The paper reports that this multimodal prototype alignment beats prior semi-supervised segmentation approaches on GlaS, EBHI-SEG-GLAND, EBHI-SEG-CANCER, and KPI, while requiring no additional pathologist annotation. The practical stake is lower annotation cost: stronger segmentation from a small labeled set plus unlabeled images.

What carries the argument

The load-bearing mechanism is the two-way prototype-pixel contrastive alignment, applied to both visual and textual prototypes. For each semantic class the model keeps K prototypes; pixels are assigned to the nearest prototype by cosine similarity, and two losses train the assignments: PAL maximizes the probability that a pixel belongs to its assigned prototype, and PCL pushes pixels away from prototypes of other classes. Text prototypes enter the same machinery: descriptive phrases for each class are embedded by a pretrained vision-language text encoder, and learnable prompt tokens are appended so the text prototypes can be optimized end-to-end. This gives unlabeled pixels coarse structural

What would settle it

Re-run the four benchmark comparisons with several random seeds under the identical labeled/unlabeled split, training schedule, augmentation, and backbone initialization, reporting per-seed mDice with standard deviations. If MPAMatch's margin over the strongest baseline falls within the run-to-run spread on GlaS or reverses on KPI, the claimed consistent superiority is not established.

Watch

Extended reading notes

Core claim

Stated on the paper's own terms: MPAMatch is a semi-supervised segmentation framework that adds a dual prototype-pixel contrastive loss to a consistency-regularization base. Visual prototypes are formed by online clustering of decoder features, and text prototypes are formed by embedding language-model-generated class descriptions with a pathology vision-language encoder and appending learnable prompt tokens. Each pixel embedding is compared against both prototype sets, and two losses—prototype alignment (PAL) and prototype contrast (PCL)—pull same-class pixels together and push other-class prototypes away. The paper claims this is the first use of text-prototype supervision in segmentation

Load-bearing premise

The load-bearing premise is that the baseline methods in Table I were trained under exactly the same conditions as MPAMatch; the paper reports no variance or significance tests, so margins of a few tenths of a Dice point could disappear under reruns.

Editorial extensions

If this is right

  • If the reported results hold, semi-supervised pathology segmentation can be improved over strong consistency-based baselines by roughly 0.2 to 1.7 Dice points without any extra manual annotation.
  • Text-prototype supervision should transfer to other dense prediction tasks whenever a short description per class is available, because the mechanism only needs a class-description list and a pretrained text encoder.
  • Swapping the segmentation encoder for a pathology-pretrained foundation-model encoder is a separable component that existing semi-supervised frameworks can adopt independently of the prototype losses.
  • The framework's best configuration with a moderate confidence threshold and a single learnable prompt token suggests the extra supervision is most useful when pseudo-labels are clean but not overly sparse.
  • The reported behavior—performance stabilizing once unlabeled data reaches 15–20% of training data—supports the practical target of cutting annotation effort rather than eliminating it.

Reading between the lines

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

  • The reported gains are not cleanly attributed: the paper does not ablate the encoder swap separately from the prototype losses, so part of the margin over baselines may come from the stronger pathology-pretrained backbone rather than from multimodal alignment.
  • Text prototypes are likely to add the most value on classes with distinct semantic descriptions; on visually similar tissue classes, the visual prototypes probably carry most of the discriminative signal.
  • The same dual-prototype loss could be tested on other label-scarce medical imaging domains (radiology, dermatology, cytology) by reusing the class-description prompting step, provided a suitable text encoder exists.
  • The ablation on prompt type implies that performance depends on prompt quality; a practical deployment would need a protocol for writing and validating class descriptions rather than assuming any text works.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes MPAMatch, a semi-supervised learning framework for pathology image segmentation. It builds on UniMatch-style consistency regularization and introduces multimodal prototype alignment: image prototypes are formed by online clustering of pixel features, while text prototypes are derived from textual descriptions encoded by a pathology vision-language model augmented with learnable CoOp tokens. Two prototype-based losses (PAL and PCL) are added to the supervised and unsupervised losses. The encoder is TransUNet with the ViT replaced by the UNI pathology foundation model. The method is evaluated on GlaS, EBHI-SEG-GLAND, EBHI-SEG-CANCER, and KPI, with reported mDice/mIoU/mCPA numbers in Table I, plus ablations on text prompts, number of Coop tokens, unlabeled-data ratio, and pseudo-label threshold.

Significance. If the empirical superiority claim were fully supported, the paper would make a useful contribution by introducing text-guided prototype alignment into semi-supervised pixel-level segmentation, an idea that is plausible and potentially relevant to computational pathology. The method is clearly specified in terms of loss functions, the choice of public datasets is appropriate, and the qualitative visualizations help illustrate the intended effect. However, the current experimental evidence is too weak to establish the central claim: the gains over a directly comparable baseline are small, no variance or significance analysis is reported, the labeled/unlabeled protocol is not a demanding low-label regime, and the ablations do not isolate the proposed components. The paper's value is therefore conditional on substantial additional experimentation.

major comments (3)
  1. [Sec. IV-B, IV-C, Table I] The central claim of consistent superiority over SOTA is not supported by the reported evidence. Table I contains a single run per method with no error bars, seeds, or statistical significance tests. The direct same-backbone comparison is MPAMatch vs. UniMatch+TransUNi, with mDice gains of only 1.22, 0.93, 0.29, and 1.70 on GlaS, EBHI-G, EBHI-C, and KPI, respectively. On EBHI-G, FixMatch+TransUNet (94.55) is within 0.18 mDice of MPAMatch (94.73). The paper does not state that all baselines were trained under identical conditions (same augmentation strengths, pseudo-label threshold, learning-rate schedule, and pretrained weights), so even the modest differences may reflect protocol mismatch rather than method performance. The authors must provide multi-seed results with variance and significance tests, and explicitly specify the common training protocol for all baselines.
  2. [Sec. IV-E, Figs. 3-4, Tables II-III] The ablations do not isolate the contribution of the proposed multimodal prototype alignment. No experiment removes L_proto entirely, or compares variants with only image prototypes vs. only text prototypes, or with vs. without the prototype fusion module. The reported ablations vary text-prompt type, Coop token count, unlabeled-data ratio, and pseudo-label threshold, none of which directly tests whether the prototype losses (PAL/PCL, Eq. 10-12) or textual supervision are responsible for the gains. In addition, the claim that MPAMatch "even outperform[s] fully-supervised models trained with the same total number of samples" is not backed by any fully-supervised baseline in the manuscript. These missing experiments are necessary to attribute the observed improvements to the paper's contributions.
  3. [Sec. IV-B, Fig. 4] The experimental setting is not a demanding low-label regime. The text states that the training set is divided into labeled and unlabeled samples at a 7:2 ratio, meaning approximately 78% of the training data is used with pixel-level labels. This is far from the typical SSL regime (e.g., 1/16 or 1/8 labeled), and with 70% of all data labeled, the observed 0.3-1.7 mDice improvement over a strong same-backbone baseline is unsurprisingly small. The paper should either evaluate under a genuinely low-label protocol (e.g., 1/16 or 1/8) or explicitly justify why this particular split is clinically relevant. The ablation in Fig. 4, where the unlabeled ratio is varied, is also unclear: the text says the total number of training images is fixed, but the ratio axis and its relationship to the 7:2 setting are not specified, and no fully-supervised comparison is shown.
minor comments (6)
  1. [Eq. (8)] The dimension of the concatenated vector [e_c^k; p_c^(k,1); ...; p_c^(k,L)] is (L+1)*d, but the mean operator is ambiguous: is the mean taken over the L+1 token embeddings? If so, the result is d-dimensional, which should be stated explicitly.
  2. [Eq. (3) and Table III] The pseudo-label threshold τ in Eq. (3) is not specified as a probability in [0,1], but Table III reports thresholds of 90, 95, 99. State clearly that these are percentages and how they map to τ.
  3. [Various] Notation is inconsistent: 'Transuni' appears in Table I while the text uses 'TransUNi' (also 'TransUNet'). The abstract uses 'GLAS' while the body uses 'GlaS'. Please standardize.
  4. [Sec. III-B] The prototype update rule is described only as 'continuously updated using the average of all assigned embeddings in the current batch.' Provide the exact update equation (e.g., with a momentum parameter) and state whether prototypes are updated with labeled, unlabeled, or both pixel features. This is needed for reproducibility.
  5. [Sec. III-B and Sec. IV] The claim that text prototype supervision is introduced 'for the first time' in segmentation is not supported by a comparative literature search. Please either provide evidence from a systematic search or soften the claim.
  6. [Fig. 4] The figure captions for Figs. 3 and 4 are minimal. For Fig. 4, specify the x-axis units (percentage of unlabeled data), the datasets, and whether the total number of training images is held constant across points. Consider adding error bars.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's SOTA claim rests on held-out test labels; pseudo-label and prototype losses are standard self-training regularizers, not derivation-equivalent predictions.

full rationale

MPAMatch's central claim is an empirical comparison on four held-out test splits. Training uses self-generated pseudo-labels (Eq. 3) and self-derived prototype assignments (Eqs. 4-5, 10), which are common SSL/clustering regularizers and do not make the test evaluation circular: final predictions are scored against external ground truth, not against the model's own outputs. The text-prototype branch is initialized from an external foundation model (CONCH), so the multimodal supervision is not purely tautological. No load-bearing self-citation or imported uniqueness theorem appears; the only overlapping-author citation [37] is a related-work mention. The paper's weaknesses—absent variance/error bars, unspecified baseline protocols, and an unsupported fully-supervised comparison—are correctness risks, not circularity.

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

The method introduces no new physical or conceptual entities; its reliance is on pretrained foundation models (UNI, CONCH) and several hand-tuned or ablated hyperparameters. The free-parameter count is moderate, but several values are unreported, which weakens reproducibility.

free parameters (6)
  • alpha, beta, gamma = 0.25, 0.5, 0.25
    Loss weighting coefficients in Eq. (13); hand-chosen, not tuned.
  • pseudo-label threshold tau = 95 (final)
    Selected from {90,95,99} via ablation in Table III, then used for final results.
  • number of Coop tokens L = 1 (final)
    Selected from {1,2,3,4,5,6} via ablation in Fig. 3, then used for final results.
  • number of prototypes per class K
    Required in Eq. (4); no value or ablation is reported in the paper.
  • prototype loss weights alpha1, alpha2
    Weights for PAL and PCL in Eq. (12) are named but their values are not given.
  • unsupervised loss weights lambda, mu
    Weights in Eq. (3) are introduced but their values are not reported.
assumptions (4)
  • domain assumption UNI's self-supervised features are effective for downstream gland and tissue segmentation when used as the TransUNet encoder.
    The paper replaces ViT with UNI and treats this as beneficial; no ablation isolates the effect of this substitution.
  • domain assumption CONCH text embeddings of LLM-generated class descriptions provide meaningful semantic prototypes.
    Text prototypes initialize the semantic prototypes; no analysis of embedding quality or alignment with pathology semantics is provided.
  • domain assumption Pseudo-labels from weakly augmented predictions are reliable enough to supervise prototype assignment and consistency on unlabeled data.
    Standard SSL assumption; the paper only tunes a threshold and does not measure pseudo-label noise in the pathology setting.
  • domain assumption Cross-attention between pixel features and prototypes improves segmentation.
    The design is stated in Sec. III-B; no ablation removes the cross-attention block to verify its contribution.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multimodal Prototype Alignment for Semi-supervised Pathology Image Segmentation." pith.science (2026). https://pith.science/paper/42YSIKR5

@misc{pith2026250819574,
  author       = {Pith},
  title        = {Pith review of: Multimodal Prototype Alignment for Semi-supervised Pathology Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/42YSIKR5}},
  note         = {Machine review of arXiv:2508.19574}
}
read the original abstract

Pathological image segmentation faces numerous challenges, particularly due to ambiguous semantic boundaries and the high cost of pixel-level annotations. Although recent semi-supervised methods based on consistency regularization (e.g., UniMatch) have made notable progress, they mainly rely on perturbation-based consistency within the image modality, making it difficult to capture high-level semantic priors, especially in structurally complex pathology images. To address these limitations, we propose MPAMatch - a novel segmentation framework that performs pixel-level contrastive learning under a multimodal prototype-guided supervision paradigm. The core innovation of MPAMatch lies in the dual contrastive learning scheme between image prototypes and pixel labels, and between text prototypes and pixel labels, providing supervision at both structural and semantic levels. This coarse-to-fine supervisory strategy not only enhances the discriminative capability on unlabeled samples but also introduces the text prototype supervision into segmentation for the first time, significantly improving semantic boundary modeling. In addition, we reconstruct the classic segmentation architecture (TransUNet) by replacing its ViT backbone with a pathology-pretrained foundation model (Uni), enabling more effective extraction of pathology-relevant features. Extensive experiments on GLAS, EBHI-SEG-GLAND, EBHI-SEG-CANCER, and KPI show MPAMatch's superiority over state-of-the-art methods, validating its dual advantages in structural and semantic modeling.

Figures

Figures reproduced from arXiv: 2508.19574 by the authors.

Figure 1
Figure 1. Overall process: (a) The framework of Transunet+UNI. (b) The framework of semi-supervised segmentation method. (c) Multimodal prototype learning [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Qualitative comparison of segmentation masks generated by FixMatch, UniMatch, and our proposed MPAMatch framework. A, B, C, and D represent [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Influence of different numbers of coop token. [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Influence of different percentages of unlabeled images in DICE. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

40 extracted references · 30 canonical work pages

  1. [1]

    Accurate diagnostic tissue segmentation and concurrent disease subtyping with small datasets,

    S. J. Frank, “Accurate diagnostic tissue segmentation and concurrent disease subtyping with small datasets,” Journal of Pathology Informatics, vol. 14, p. 100174, 2023

    work page 2023

  2. [2]

    Multi-scale domain-adversarial multiple-instance cnn for cancer subtype classifica- tion with unannotated histopathological images,

    N. Hashimoto, D. Fukushima, R. Koga, Y . Takagi, K. Ko, K. Kohno, M. Nakaguro, S. Nakamura, H. Hontani, and I. Takeuchi, “Multi-scale domain-adversarial multiple-instance cnn for cancer subtype classifica- tion with unannotated histopathological images,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2020, pp. 3852–3861

    work page 2020

  3. [3]

    Fully convolutional networks for semantic segmentation,

    J. Long, E. Shelhamer, and T. Darrell, “Fully convolutional networks for semantic segmentation,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2015, pp. 3431–3440

    2015

  4. [4]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international con- ference, Munich, Germany, October 5-9, 2015, proceedings, part III 18 . Springer, 2015, pp. 234–241

    2015

  5. [5]

    Encoder- decoder with atrous separable convolution for semantic image segmen- tation,

    L.-C. Chen, Y . Zhu, G. Papandreou, F. Schroff, and H. Adam, “Encoder- decoder with atrous separable convolution for semantic image segmen- tation,” in Proceedings of the European conference on computer vision (ECCV), 2018, pp. 801–818

    2018

  6. [6]

    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

    arXiv 2021

  7. [7]

    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,” in European conference on computer vision . Springer, 2022, pp. 205–218

    2022

  8. [8]

    Clinical-grade computational pathology using weakly supervised deep learning on whole slide images,

    G. Campanella, M. G. Hanna, L. Geneslaw, A. Miraflor, V . Werneck Krauss Silva, K. J. Busam, E. Brogi, V . E. Reuter, D. S. Klimstra, and T. J. Fuchs, “Clinical-grade computational pathology using weakly supervised deep learning on whole slide images,” Nature medicine , vol. 25, no. 8, pp. 1301–1309, 2019

    work page 2019

Show all 40 references
  1. [9]

    Deep semi-supervised learning for medical image segmentation: A review,

    K. Han, V . S. Sheng, Y . Song, Y . Liu, C. Qiu, S. Ma, and Z. Liu, “Deep semi-supervised learning for medical image segmentation: A review,” Expert Systems with Applications , vol. 245, p. 123052, 2024

    2024

  2. [10]

    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,” in Advances in Neural Information Processing Systems (NeurIPS), vol. 30, 2017

    2017

  3. [11]

    Adversarial learning for semi-supervised semantic segmentation,

    W.-C. Hung, Y .-H. Tsai, Y .-T. Liou, Y .-Y . Lin, and M.-H. Yang, “Adversarial learning for semi-supervised semantic segmentation,” arXiv preprint arXiv:1802.07934, 2018

    2018 arXiv

  4. [12]

    Billion- scale semi-supervised learning for image classification,

    I. Z. Yalniz, H. J ´egou, K. Chen, M. Paluri, and D. Mahajan, “Billion- scale semi-supervised learning for image classification,” arXiv preprint arXiv:1905.00546, 2019

    1905 arXiv

  5. [13]

    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,” Advances in neural information processing systems , vol. 33, pp. 596–608, 2020

    2020

  6. [14]

    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,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 7236–7246

    2023

  7. [15]

    Clims: Cross language image matching for weakly supervised semantic segmentation,

    J. Xie, X. Hou, K. Ye, and L. Shen, “Clims: Cross language image matching for weakly supervised semantic segmentation,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2022, pp. 4483–4492

    2022

  8. [16]

    Prototypical multiple instance learning for predicting lymph node metastasis of breast cancer from whole-slide pathological images,

    J.-G. Yu, Z. Wu, Y . Ming, S. Deng, Y . Li, C. Ou, C. He, B. Wang, P. Zhang, and Y . Wang, “Prototypical multiple instance learning for predicting lymph node metastasis of breast cancer from whole-slide pathological images,” Medical Image Analysis, vol. 85, p. 102748, 2023

    2023

  9. [17]

    Pamil: Prototype attention-based multiple instance learning for whole slide image classification,

    J. Liu, A. Mao, Y . Niu, X. Zhang, T. Gong, C. Li, and Z. Gao, “Pamil: Prototype attention-based multiple instance learning for whole slide image classification,” in International Conference on Medical Image Computing and Computer-Assisted Intervention . Springer, 2024, pp. 362–372

    2024

  10. [18]

    Learning transferable visual models from natural language supervision,

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark et al. , “Learning transferable visual models from natural language supervision,” in International conference on machine learning . PmLR, 2021, pp. 8748–8763

    2021

  11. [19]

    A visual–language foundation model for pathology image analysis using medical twitter,

    Z. Huang, F. Bianchi, M. Yuksekgonul, T. J. Montine, and J. Zou, “A visual–language foundation model for pathology image analysis using medical twitter,” Nature medicine, vol. 29, no. 9, pp. 2307–2316, 2023

    2023

  12. [20]

    A visual-language foundation model for computational pathology,

    M. Y . Lu, B. Chen, D. F. Williamson, R. J. Chen, I. Liang, T. Ding, G. Jaume, I. Odintsov, L. P. Le, G. Gerber et al. , “A visual-language foundation model for computational pathology,” Nature Medicine , vol. 30, no. 3, pp. 863–874, 2024

    2024

  13. [21]

    Learning to prompt for vision- language models,

    K. Zhou, J. Yang, C. C. Loy, and Z. Liu, “Learning to prompt for vision- language models,” International Journal of Computer Vision , vol. 130, no. 9, pp. 2337–2348, 2022

    2022

  14. [22]

    Towards a general-purpose foundation model for computational pathology,

    R. J. Chen, T. Ding, M. Y . Lu, D. F. Williamson, G. Jaume, A. H. Song, B. Chen, A. Zhang, D. Shao, M. Shaban et al. , “Towards a general-purpose foundation model for computational pathology,” Nature medicine, vol. 30, no. 3, pp. 850–862, 2024

    2024

  15. [23]

    Semi-supervised deep transfer learn- ing for benign-malignant diagnosis of pulmonary nodules in chest ct images,

    F. Shi, B. Chen, Q. Cao et al. , “Semi-supervised deep transfer learn- ing for benign-malignant diagnosis of pulmonary nodules in chest ct images,” IEEE Transactions on Medical Imaging , vol. 41, no. 4, pp. 771–781, 2022

    2022

  16. [24]

    Semixup: In- and out-of-manifold regularization for deep semi-supervised knee os- teoarthritis severity grading from plain radiographs,

    H. H. Nguyen, S. Saarakkala, M. B. Blaschko et al. , “Semixup: In- and out-of-manifold regularization for deep semi-supervised knee os- teoarthritis severity grading from plain radiographs,” IEEE Transactions on Medical Imaging , vol. 39, no. 12, pp. 4346–4356, 2020

    2020

  17. [25]

    Shadow-consistent semi- supervised learning for prostate ultrasound segmentation,

    X. Xu, T. Sanford, B. Turkbey et al. , “Shadow-consistent semi- supervised learning for prostate ultrasound segmentation,” IEEE Trans- actions on Medical Imaging , vol. 41, no. 6, pp. 1331–1345, 2022

    2022

  18. [26]

    Few-shot learning by a cascaded framework with shape-constrained pseudo label assessment for whole heart segmentation,

    W. Wang, Q. Xia, Z. Hu et al. , “Few-shot learning by a cascaded framework with shape-constrained pseudo label assessment for whole heart segmentation,” IEEE Transactions on Medical Imaging , vol. 40, no. 10, pp. 2629–2641, 2021

    2021

  19. [27]

    Semantic segmen- tation with generative models: Semi-supervised learning and strong out- of-domain generalization,

    D. Li, J. Yang, K. Kreis, A. Torralba, and S. Fidler, “Semantic segmen- tation with generative models: Semi-supervised learning and strong out- of-domain generalization,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 8300–8311

    2021

  20. [28]

    Revisiting consistency regularization for semi-supervised learning,

    Y . Fan, A. Kukleva, D. Dai, and B. Schiele, “Revisiting consistency regularization for semi-supervised learning,” International Journal of Computer Vision, vol. 131, no. 3, pp. 626–643, 2023

    2023

  21. [29]

    Semi-supervised deep learning via transformation consistency regular- ization for remote sensing image semantic segmentation,

    B. Zhang, Y . Zhang, Y . Li, Y . Wan, H. Guo, Z. Zheng, and K. Yang, “Semi-supervised deep learning via transformation consistency regular- ization for remote sensing image semantic segmentation,” IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing...

    2022

  22. [30]

    Semi- supervised medical image segmentation using adversarial consistency learning and dynamic convolution network,

    T. Lei, D. Zhang, X. Du, X. Wang, Y . Wan, and A. K. Nandi, “Semi- supervised medical image segmentation using adversarial consistency learning and dynamic convolution network,” IEEE transactions on medical imaging , vol. 42, no. 5, pp. 1265–1277, 2022

    2022

  23. [31]

    A three-stage self-training framework for semi-supervised semantic segmentation,

    R. Ke, A. I. Aviles-Rivero, S. Pandey, S. Reddy, and C.-B. Sch ¨onlieb, “A three-stage self-training framework for semi-supervised semantic segmentation,” IEEE Transactions on Image Processing , vol. 31, pp. 1805–1815, 2022

    2022

  24. [32]

    Local contrastive loss with pseudo-label based self-training for semi-supervised medical image segmentation,

    K. Chaitanya, E. Erdil, N. Karani, and E. Konukoglu, “Local contrastive loss with pseudo-label based self-training for semi-supervised medical image segmentation,” Medical image analysis , vol. 87, p. 102792, 2023

    2023

  25. [33]

    A semi-supervised learning approach for tissue semantic segmentation in whole slide images,

    R. Rashmi, G. Sudhamsh, and S. Girisha, “A semi-supervised learning approach for tissue semantic segmentation in whole slide images,” IEEE Access, 2024

    2024

  26. [34]

    Semi-supervised breast cancer pathology image segmentation based on fine-grained clas- sification guidance,

    K. Sun, Y . Zheng, X. Yang, X. Chen, and W. Jia, “Semi-supervised breast cancer pathology image segmentation based on fine-grained clas- sification guidance,” Medical & Biological Engineering & Computing , vol. 62, no. 3, pp. 901–912, 2024

    2024

  27. [35]

    Multi-layer pseudo-supervision for histopathology tissue semantic segmentation using patch-level classification labels,

    C. Han, J. Lin, J. Mai, Y . Wang, Q. Zhang, B. Zhao, X. Chen, X. Pan, Z. Shi, Z. Xu et al., “Multi-layer pseudo-supervision for histopathology tissue semantic segmentation using patch-level classification labels,” Medical Image Analysis , vol. 80, p. 102487, 2022

    2022

  28. [36]

    A semi-supervised learning for segmentation of gigapixel histopathol- ogy images from brain tissues,

    Z. Lai, C. Wang, Z. Hu, B. N. Dugger, S.-C. Cheung, and C.-N. Chuah, “A semi-supervised learning for segmentation of gigapixel histopathol- ogy images from brain tissues,” in 2021 43rd annual international conference of the IEEE engineering in Medicine & Biology Society (EMBC)...

    2021

  29. [37]

    Diagnostic text-guided representation learning in hierarchical classification for pathological whole slide image,

    J. Li, Q. Sun, R. Yan, Y . Wang, Y . Fu, Y . Wei, T. Guan, H. Shi, Y . He, and A. Han, “Diagnostic text-guided representation learning in hierarchical classification for pathological whole slide image,” arXiv preprint arXiv:2411.10709, 2024

    2024 arXiv

  30. [38]

    mtree: Multi-level text-guided representation end-to-end learning for whole slide image analysis,

    Q. Liu, R. Deng, C. Cui, T. Yao, V . Nath, Y . Tang, and Y . Huo, “mtree: Multi-level text-guided representation end-to-end learning for whole slide image analysis,” arXiv preprint arXiv:2405.17824 , 2024

    2024 arXiv

  31. [39]

    Pathology-knowledge enhanced multi-instance prompt learning for few- shot whole slide image classification,

    L. Qu, D. Yang, D. Huang, Q. Guo, R. Luo, S. Zhang, and X. Wang, “Pathology-knowledge enhanced multi-instance prompt learning for few- shot whole slide image classification,” in European Conference on Computer Vision. Springer, 2024, pp. 196–212

    2024

  32. [40]

    Unet 3+: A full-scale connected unet for medical image segmentation,

    H. Huang, L. Lin, R. Tong, H. Hu, Q. Zhang, Y . Iwamoto, X. Han, Y .-W. Chen, and J. Wu, “Unet 3+: A full-scale connected unet for medical image segmentation,” in ICASSP 2020-2020 IEEE international conference on acoustics, speech and signal processing (ICASSP) . IEEE, 2020, p...

    2020

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.