Pith. sign in

REVIEW 3 major objections 8 minor 48 references

Leveraging Textual Anatomical Knowledge for Class-Imbalanced Semi-Supervised Multi-Organ Segmentation

T0 review · 3 major / 8 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read Injecting one-time textual descriptions of organ positions and shapes into the segmentation head substantially improves class-imbalanced semi-supervised multi-organ segmentation.

desk verdict Genuinely new use of LLM-generated anatomical priors for SSL segmentation, with solid ablations, but single-run numbers and test-set hyperparameter tuning make the headline gains unconfirmed. read the letter →

arxiv 2501.13470 v2 pith:N3QJKHXP submitted 2025-01-23 cs.CV

Yuliang Gu , Weilun Tsao , Bo Du , Thierry Géraud , Yongchao Xu This is my paper · ORCID
classification cs.CV
keywords semi-supervisedlearningmulti-organsegmentationclassimbalancetextualanatomicalknowledgecross-modalcontrastivevision-languagemodels3Dmedicalimaging
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper claims that class-imbalanced semi-supervised multi-organ segmentation can be improved substantially by handing the model explicit textual priors about where each organ sits and what shape it takes. The priors, generated once by a multimodal large language model and encoded with a biomedical vision-language encoder, are injected as parameters of the segmentation head and simultaneously used as anchors in a cross-modal contrastive loss that aligns visual features with the same anatomical descriptions. On two abdominal CT benchmarks with 2 to 20 percent labeled data, the reported average Dice is 4.2 to 8.3 points above the previous strongest class-imbalance-aware method, with the largest per-class gains on small organs such as the duodenum, pancreas, and adrenal glands. If correct, this shows that anatomical knowledge can be added cheaply and generically, without hand-crafted topology or geometry modules.

What carries the argument

The load-bearing object is the textual anatomical knowledge (TAK), a one-time set of per-organ descriptions of inter-organ relative positions and organ shape. These descriptions are encoded by a biomedical image-text encoder into embeddings $T^p_k$ and $T^s_k$; for each class $k$, the concatenation of $T^p_k$, $T^s_k$, and the global pooled visual feature $F$ is fed to an MLP controller that generates the segmentation head parameters $\theta_k$. A second mechanism, the cross-modal contrastive alignment module, treats text embeddings and visual features of the same class at the same scale as positive pairs and features of different classes as negative pairs, so the pixel-level decoder is regularized to honor the anatomical statements.

What would settle it

Run the full pipeline with the same training data and hyperparameters but replace the validated anatomical descriptions with scrambled or false ones, for example describing the duodenum as lying superior to the liver; if average Dice on the small organs does not drop well below the correct-prior run, the reported gains are not caused by the truth of the anatomical priors.

Watch

Extended reading notes

Core claim

The paper's central discovery is that free-text anatomical facts, encoded into class embeddings, can act as both the parameters of a segmentation head and as contrastive anchors for visual features. For each organ class, the position description and shape description are encoded separately, concatenated with the pooled image feature, and passed through an MLP that outputs the head weights for that class. At the same time, multi-scale student features are pushed toward the text embeddings of their class and pulled away from other classes, with uncertain voxels from unlabeled images excluded by an entropy threshold. Under a mean-teacher pseudo-labeling setup, the method reports an average Dice of 60.84 over the gradient-aware baseline's 53.84 at 2 percent labeled data on the first benchmark, 70.20 over 63.50 at 5 percent, 65.75 over 57.45 at 10 percent on the second benchmark, and 72.66 over 68.43 at 20 percent.

Load-bearing premise

The method assumes that the large language model's textual descriptions of organ positions and shapes are anatomically accurate for the target scans and that a biomedical image-text encoder can align those descriptions with 3D CT visual features well enough for them to serve as useful segmentation-head parameters and contrastive anchors.

Editorial extensions

If this is right

  • At 2% labeled data on the first benchmark, average Dice improves from 53.84 to 60.84, a +7.00 gain, with small organs improving by 8.92 points on average.
  • At 5% labeled data on the same benchmark, average Dice reaches 70.20, with small organs gaining 7.62 points over the gradient-aware baseline.
  • On the second benchmark, 10% labeled data yields an 8.30-point average gain and 20% labeled data a 4.23-point gain, and the largest per-class gains are again on small organs.
  • The text priors are generated once before training, so the extra cost is confined to a small MLP and a contrastive loss; the reported model stays close to the baseline in FLOPs and parameters, at about 151 GFLOPs and 18.5M parameters.

Reading between the lines

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

  • A natural extension the paper leaves untested is cross-dataset transfer: because the textual priors are class-level and generated once, the same descriptions could be reused on a new CT dataset without re-annotation, provided the organ definitions and imaging protocols match.
  • The method's sensitivity to wording is a testable prediction: paraphrasing the same anatomical facts should preserve the gains, while false or vague descriptions should degrade small-organ Dice if the contrastive alignment is truly doing the work.
  • The same parameter-injection plus contrastive-anchor recipe could plausibly carry other stable medical priors, such as pathological co-occurrence or physiological dynamics, into segmentation models, although the paper only demonstrates spatial and shape priors.
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 / 8 minor

Summary. The paper proposes TAK, a semi-supervised multi-organ segmentation method that injects textual anatomical knowledge into a mean-teacher framework. GPT-4o generates textual descriptions of inter-organ spatial relationships and organ shape priors; these are encoded with BiomedCLIP and used, together with the global image feature, by a text-based controller to generate segmentation-head parameters. A cross-modal contrastive loss aligns visual features with the text embeddings, with entropy-based filtering on pseudo-labels. Using MagicNet as the vision backbone, the method is evaluated on AMOS (2% and 5% labeled data) and Synapse (10% and 20% labeled data), reporting average Dice gains over GA of +7.00, +6.70, +8.30, and +4.23 points respectively, with larger gains on small organs. Ablations cover prompt types, the contrastive module, the contrastive loss coefficient, the number of sampled visual features, and the choice of text encoder.

Significance. If the reported gains are robust, the paper makes a useful contribution: it offers a clean mechanism for turning LLM-generated anatomical priors into usable segmentation-head parameters and contrastive anchors, without adding substantial inference cost (Table 6 shows only a small FLOP/parameter increase over GA). The experimental setup is largely fair: public AMOS/Synapse splits, a single consistent backbone, a wide set of 14 baselines, and ablations that isolate the proposed components. The claim is also non-circular: the improvement is measured against external baselines, not against a quantity fitted by the method itself. The main weakness is statistical: all tables report single runs with no variance, and the headline hyperparameters are tuned on the same dataset and label ratio used to report the main result. Given that the 2% AMOS setting corresponds to roughly four labeled volumes, the central quantitative claim is not yet established at the level a medical-imaging journal should require.

major comments (3)
  1. [§4.3, Tables 1–4] Every reported Dice number is a single run with no standard deviation, no number of seeds, and no repeated sampling of the labeled subset. This matters most at 2% labeled AMOS, where the labeled set contains approximately four volumes: with small organs present in only some of those volumes, the +7.00 Dice advantage over GA could easily be a favorable labeled-subset or seed effect rather than a property of the method. The load-bearing claim of the paper is empirical, so the authors should report mean and standard deviation over at least three random seeds, and ideally multiple random choices of the labeled subset under the same protocol, with paired differences against GA. Without this, the central conclusion that textual anatomical knowledge is responsible for the gains is not supported.
  2. [§4.4, Tables 7–8] The hyperparameters λc and λN are selected by ablations on 5% labeled AMOS and then used to produce the main 5% AMOS result in Table 2, where TAK is reported to beat GA by +6.70 Dice. Tuning on the same dataset and label ratio that is then used for the headline comparison introduces an optimistic selection bias, which is especially relevant because the gains are of the same order as the spread across λc values in Table 7 (68.16 to 70.20) and across λN values in Table 8 (68.16 to 70.20). The authors should either tune on a separate validation split and apply the fixed hyperparameters to the reported test setting, or perform a nested evaluation; at minimum, they should report the sensitivity of GA to the same tuned hyperparameters and show that the TAK–GA gap is not an artifact of choosing favorable coefficients.
  3. [§3.2] The textual anatomical knowledge generation is central to the method, but the main text describes it only at a high level: the second agent performs multi-modal validation against randomly sampled labeled data and filters low-confidence claims, and doctors review the output, yet no details are given about the exact prompts, the validation criteria, the number of descriptions retained or discarded, or the nature of the doctor review, with all details deferred to the supplementary. For reproducibility and for assessing the claim that the priors are generic rather than tailored to the test set, the paper should include the complete set of generated descriptions and the full generation protocol in the main paper or in a permanently released appendix.
minor comments (8)
  1. [§4.4, Table 5] The marker layout of Table 5 is ambiguous: the rows and columns do not make it clear which configuration corresponds to 'Name plus Position plus Shape with the Contrast module' versus 'without', and the column header contains the typo 'Positon'. Please redraw the table with explicit check marks for each prompt component and for the contrast module, and clarify in the caption which rows are compared in the text.
  2. [§1 and §4] The abstract says the source code 'will be available', while §4.2 says it 'is available'; please unify the wording and ensure the GitHub link is included in the final version.
  3. [§4.2] The sentence 'we use MagicNet as the backbone of the vision branch' is followed by 'with decay of 10e−4'; please specify that this is the weight decay and give the exact value (10e−4 appears to mean 0.001, which should be stated unambiguously).
  4. [§3.3, Eq. (1)] The entropy formula uses p̂_ij without defining the voxel index j; please define the notation and state whether the background class is included in the sum over C classes.
  5. [§3.3, Eq. (2)] The contrastive loss is described as acting on multi-scale features, but the positive and negative sets are not fully specified: please clarify whether negative pairs are restricted to the same feature scale (i1 = i2), how the background class is handled in unlabeled data, and how the sets P(f) and N(f) are constructed when text embeddings are the anchor.
  6. [§1, Fig. 1] The convex hull volume ratio is mentioned as a shape-complexity measure but never defined quantitatively; please provide the formula or a precise definition in the text or caption.
  7. [§3.2] The sentence 'The detail of this Textual Anatomical Knowledge (TAK) Generation process is illustrated in Sec. 3.3' appears to reference the wrong section; the generation process is described in §3.2 and the integration in §3.3, so please correct the cross-reference.
  8. [§4.4, Table 9] The text says 'all models significantly outperform the baseline', but no statistical test or error bars are reported; 'significantly' should be removed or supported by paired significance testing.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: textual priors are fixed external inputs, comparisons are against external SOTA on held-out test splits, and ablations isolate the contribution of each proposed component.

full rationale

The central claim is that injecting GPT-4o-generated textual anatomical knowledge, encoded with BiomedCLIP and integrated via a text-driven segmentation head plus cross-modal contrastive alignment, improves class-imbalanced semi-supervised multi-organ segmentation. The reported gains are assessed against external baselines (GA, DHC, MagicNet, etc.) on held-out AMOS and Synapse test sets. The textual anatomical priors are generated once before training (Sec. 3.2) and are not fitted to the test data; the second-agent validation on randomly sampled labeled data and doctor review is a quality filter, not a fitted parameter. Ablations in Tables 5, 7, 8, and 9 vary the prompt content, the contrastive module, the loss coefficient, the sample number, and the text encoder while keeping the same evaluation protocol, which isolates the contribution of the proposed components rather than reusing a fitted quantity as a prediction. No equation in the paper defines the claimed improvement in terms of an input parameter, and no load-bearing argument reduces to a self-citation. The authors' own prior works [12, 13] appear only in related work and are not used to justify the main result. Concerns about single-run, no-error-bar evaluation at very low labeling ratios are correctness and robustness risks, not circularity. Therefore no significant circularity is present.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical or mathematical entities; it relies on standard SSL machinery, a pretrained text encoder, and generated textual priors. The main fitted quantities are the two contrastive-learning hyperparameters, a temperature whose value is omitted, and the learned parameters of the segmentation head and projection MLPs, which are trained normally.

free parameters (3)
  • λc (contrastive loss coefficient) = 0.1
    Chosen by ablation on the 5% labeled AMOS setup (Table 7); Dice ranges from 68.16 to 70.20 across tested values.
  • λN (number of sampled visual features per class in contrastive loss) = 40
    Chosen by ablation in Table 8; peak Dice at 40.
  • τ (temperature in contrastive loss) = not stated
    Equation 2 includes a temperature τ for the contrastive loss, but its value is not reported.
assumptions (4)
  • domain assumption GPT-4o-generated textual anatomical priors are anatomically accurate and generalizable
    Section 3.2; the method's prior injection depends on the fidelity of these descriptions, which are only validated on a small labeled set and by doctor review.
  • domain assumption Entropy-based uncertainty filtering removes unreliable pseudo-label voxels
    Section 3.3, Eq. 1; the contrastive learning on unlabeled data assumes this filter improves rather than biases the sampled visual features.
  • domain assumption BiomedCLIP text embeddings of long descriptive sentences align with 3D CT visual features after MLP projection
    Section 3.3; both the text-driven head and the contrastive alignment assume a meaningful shared embedding space.
  • domain assumption Convex hull volume ratio is a valid proxy for shape complexity
    Used in Fig. 1 and Sec. 1 to motivate the need for shape priors; no independent validation of this proxy is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Leveraging Textual Anatomical Knowledge for Class-Imbalanced Semi-Supervised Multi-Organ Segmentation." pith.science (2026). https://pith.science/paper/N3QJKHXP

@misc{pith2026250113470,
  author       = {Pith},
  title        = {Pith review of: Leveraging Textual Anatomical Knowledge for Class-Imbalanced Semi-Supervised Multi-Organ Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N3QJKHXP}},
  note         = {Machine review of arXiv:2501.13470}
}
read the original abstract

Annotating 3D medical images demands substantial time and expertise, driving the adoption of semi-supervised learning (SSL) for segmentation tasks. However, the complex anatomical structures of organs often lead to significant class imbalances, posing major challenges for deploying SSL in real-world scenarios. Despite the availability of valuable prior information, such as inter-organ relative positions and organ shape priors, existing SSL methods have yet to fully leverage these insights. To address this gap, we propose a novel approach that integrates textual anatomical knowledge (TAK) into the segmentation model. Specifically, we use GPT-4o to generate textual descriptions of anatomical priors, which are then encoded using a CLIP-based model. These encoded priors are injected into the segmentation model as parameters of the segmentation head. Additionally, contrastive learning is employed to enhance the alignment between textual priors and visual features. Extensive experiments demonstrate the superior performance of our method, significantly surpassing state-of-the-art approaches. The source code will be available at: https://github.com/Lunn88/TAK-Semi.

Figures

Figures reproduced from arXiv: 2501.13470 by the authors.

Figure 1
Figure 1. A scatter plot of the voxel proportion and shape [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Examples of textual anatomical knowledge and their [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of our method. We propose Textual Anatomical Knowledge Generation Module and Cross-Modal Contrastive [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Some qualitative segmentation results of our TAK [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 44 canonical work pages

  1. [1]

    Semi-supervised semantic segmentation with pixel-level contrastive learning from a class-wise memory bank

    Inigo Alonso, Alberto Sabater, David Ferstl, Luis Mon- tesano, and Ana C Murillo. Semi-supervised semantic segmentation with pixel-level contrastive learning from a class-wise memory bank. In Int. Conf. Comput. Vis. , pages 8219–8228, 2021. 2

    work page 2021

  2. [2]

    Pseudo-label guided contrastive learning for semi-supervised medical image segmentation

    Hritam Basak and Zhaozheng Yin. Pseudo-label guided contrastive learning for semi-supervised medical image segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 19786–19797, 2023. 1, 2

    work page 2023

  3. [3]

    Address- ing class imbalance in semi-supervised image segmenta- tion: A study on cardiac mri

    Hritam Basak, Sagnik Ghosal, and Ram Sarkar. Address- ing class imbalance in semi-supervised image segmenta- tion: A study on cardiac mri. In Proc. of MICCAI, pages 224–233, 2022. 6, 7

    work page 2022

  4. [4]

    Rep- resentation and fusion of heterogeneous fuzzy informa- tion in the 3d space for model-based structural recogni- tion—application to 3d brain imaging

    Isabelle Bloch, Thierry G ´eraud, and Henri Maˆıtre. Rep- resentation and fusion of heterogeneous fuzzy informa- tion in the 3d space for model-based structural recogni- tion—application to 3d brain imaging. Artificial Intelli- gence, 148(1-2):141–175, 2003. 3

    work page 2003

  5. [5]

    What is the effect of importance weighting in deep learning? In Proc

    Jonathon Byrd and Zachary Lipton. What is the effect of importance weighting in deep learning? In Proc. of Intl. Conf. on Machine Learning, pages 872–881, 2019. 2

    work page 2019

  6. [6]

    Orthogonal annotation benefits barely- supervised medical image segmentation

    Heng Cai, Shumeng Li, Lei Qi, Qian Yu, Yinghuan Shi, and Yang Gao. Orthogonal annotation benefits barely- supervised medical image segmentation. In IEEE Conf. Comput. Vis. Pattern Recog., pages 3302–3311, 2023. 6, 7

    work page 2023

  7. [7]

    Learning imbalanced datasets with label-distribution-aware margin loss

    Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, and Tengyu Ma. Learning imbalanced datasets with label-distribution-aware margin loss. Adv. Neural In- form. Process. Syst., 32, 2019. 2

    work page 2019

  8. [8]

    Magicnet: Semi-supervised multi- organ segmentation via magic-cube partition and recov- ery

    Duowen Chen, Yunhao Bai, Wei Shen, Qingli Li, Lequan Yu, and Yan Wang. Magicnet: Semi-supervised multi- organ segmentation via magic-cube partition and recov- ery. In IEEE Conf. Comput. Vis. Pattern Recog. , pages 23869–23878, 2023. 2, 3, 5, 6, 7

    work page 2023

Show all 48 references
  1. [9]

    An embarrassingly simple baseline for imbalanced semi- supervised learning

    Hao Chen, Yue Fan, Yidong Wang, Jindong Wang, Bernt Schiele, Xing Xie, Marios Savvides, and Bhiksha Raj. An embarrassingly simple baseline for imbalanced semi- supervised learning. arXiv preprint arXiv:2211.11086 ,

    arXiv

  2. [10]

    Semi-supervised semantic segmentation with cross pseudo supervision

    Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2613–2622, 2021. 1, 6, 7

    2021

  3. [11]

    Semi-supervised and unsupervised deep visual learning: A survey

    Yanbei Chen, Massimiliano Mancini, Xiatian Zhu, and Zeynep Akata. Semi-supervised and unsupervised deep visual learning: A survey. IEEE Trans. on Pattern Anal. and Mach. Intell., 46(3):1327–1347, 2022. 1

    2022

  4. [12]

    Shape transforma- tion driven by active contour for class-imbalanced semi- supervised medical image segmentation

    Yuliang Gu, Yepeng Liu, Zhichao Sun, Jinchi Zhu, Yongchao Xu, and Laurent Najman. Shape transforma- tion driven by active contour for class-imbalanced semi- supervised medical image segmentation. In 2024 IEEE BIBM, pages 1966–1973, 2024. 2

    2024

  5. [13]

    Dual structure- aware image filterings for semi-supervised medical im- age segmentation

    Yuliang Gu, Zhichao Sun, Tian Chen, Xin Xiao, Yepeng Liu, Yongchao Xu, and Laurent Najman. Dual structure- aware image filterings for semi-supervised medical im- age segmentation. Medical Image Analysis, 99:103364,

  6. [14]

    Class-imbalanced semi- supervised learning with adaptive thresholding

    Lan-Zhe Guo and Yu-Feng Li. Class-imbalanced semi- supervised learning with adaptive thresholding. In Proc. of Intl. Conf. on Machine Learning , pages 8082–8094,

  7. [15]

    Learning topo- logical interactions for multi-class medical image seg- mentation

    Saumya Gupta, Xiaoling Hu, James Kaan, Michael Jin, Mutshipay Mpoy, Katherine Chung, Gagandeep Singh, Mary Saltz, Tahsin Kurc, Joel Saltz, et al. Learning topo- logical interactions for multi-class medical image seg- mentation. In Eur. Conf. Comput. Vis., pages 701–718,

  8. [16]

    Topology-aware uncertainty for image segmentation

    Saumya Gupta, Yikai Zhang, Xiaoling Hu, Prateek Prasanna, and Chao Chen. Topology-aware uncertainty for image segmentation. Adv. Neural Inform. Process. Syst., 36, 2024. 3

    2024

  9. [17]

    Cat: Coordinat- ing anatomical-textual prompts for multi-organ and tu- mor segmentation

    Zhongzhen Huang, Yankai Jiang, Rongzhao Zhang, Shaoting Zhang, and Xiaofan Zhang. Cat: Coordinat- ing anatomical-textual prompts for multi-organ and tu- mor segmentation. Adv. Neural Inform. Process. Syst. ,

  10. [18]

    Amos: A large-scale abdominal multi-organ benchmark for versatile medical image seg- mentation

    Yuanfeng Ji, Haotian Bai, Jie Yang, Chongjian Ge, Ye Zhu, Ruimao Zhang, Zhen Li, Lingyan Zhang, Wanling Ma, Xiang Wan, et al. Amos: A large-scale abdominal multi-organ benchmark for versatile medical image seg- mentation. arXiv preprint arXiv:2206.08023, 2022. 5, 6, 7, 8

    2022 arXiv

  11. [19]

    Zept: Zero-shot pan- tumor segmentation via query-disentangling and self- prompting

    Yankai Jiang, Zhongzhen Huang, Rongzhao Zhang, Xi- aofan Zhang, and Shaoting Zhang. Zept: Zero-shot pan- tumor segmentation via query-disentangling and self- prompting. In IEEE Conf. Comput. Vis. Pattern Recog., pages 11386–11397, 2024. 3

    2024

  12. [20]

    Unimed-clip: Towards a unified image-text pretraining paradigm for diverse medical imaging modalities

    Muhammad Uzair Khattak, Shahina Kunhimon, Muza- mmal Naseer, Salman Khan, and Fahad Shahbaz Khan. Unimed-clip: Towards a unified image-text pretraining paradigm for diverse medical imaging modalities. arXiv preprint arXiv:2412.10372, 2024. 3, 8

    2024 arXiv

  13. [21]

    Miccai multi-atlas labeling be- yond the cranial vault–workshop and challenge

    B Landman, Z Xu, J Eugenio Igelsias, M Styner, T Langerak, and A Klein. Miccai multi-atlas labeling be- yond the cranial vault–workshop and challenge. In Proc. MICCAI: Multi-Atlas Labeling Beyond Cranial Vault- Workshop Challenge, 2015. 5, 6, 7

    2015

  14. [22]

    Grounded language- image pre-training

    Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, et al. Grounded language- image pre-training. In IEEE Conf. Comput. Vis. Pattern Recog., pages 10965–10975, 2022. 3

    2022

  15. [23]

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

    Shuailin Li, Chuyu Zhang, and Xuming He. Shape-aware semi-supervised 3d semantic segmentation for medical images. In Proc. of MICCAI, pages 552–561, 2020. 3

    2020

  16. [24]

    Pmc-clip: Contrastive language-image pre-training using biomedi- cal documents

    Weixiong Lin, Ziheng Zhao, Xiaoman Zhang, Chaoyi Wu, Ya Zhang, Yanfeng Wang, and Weidi Xie. Pmc-clip: Contrastive language-image pre-training using biomedi- cal documents. In International Conference on Medical Image Computing and Computer-Assisted Intervention , pages 525–536...

    2023

  17. [25]

    Calibrating label distribution for class- imbalanced barely-supervised knee segmentation

    Yiqun Lin, Huifeng Yao, Zezhong Li, Guoyan Zheng, and Xiaomeng Li. Calibrating label distribution for class- imbalanced barely-supervised knee segmentation. In Proc. of MICCAI, pages 109–118, 2022. 2, 3, 6, 7 10

    2022

  18. [26]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Adv. Neural Inform. Pro- cess. Syst., 36, 2024. 3

    2024

  19. [27]

    Clip-driven universal model for organ segmentation and tumor de- tection

    Jie Liu, Yixiao Zhang, Jie-Neng Chen, Junfei Xiao, Yongyi Lu, Bennett A Landman, Yixuan Yuan, Alan Yuille, Yucheng Tang, and Zongwei Zhou. Clip-driven universal model for organ segmentation and tumor de- tection. In Int. Conf. Comput. Vis., pages 21152–21164,

  20. [28]

    Large-scale long-tailed recognition in an open world

    Ziwei Liu, Zhongqi Miao, Xiaohang Zhan, Jiayun Wang, Boqing Gong, and Stella X Yu. Large-scale long-tailed recognition in an open world. In IEEE Conf. Comput. Vis. Pattern Recog., pages 2537–2546, 2019. 2

    2019

  21. [29]

    Semi-supervised medical image segmentation through dual-task consistency

    Xiangde Luo, Jieneng Chen, Tao Song, and Guotai Wang. Semi-supervised medical image segmentation through dual-task consistency. InProc. of the AAAI Conf. on Artificial Intelligence, pages 8801–8809, 2021. 3

    2021

  22. [30]

    Pseudo- label guided image synthesis for semi-supervised covid- 19 pneumonia infection segmentation

    Fei Lyu, Mang Ye, Jonathan Frederik Carlsen, Kenny Erleben, Sune Darkner, and Pong C Yuen. Pseudo- label guided image synthesis for semi-supervised covid- 19 pneumonia infection segmentation. IEEE Trans. on Medical Imaging, 42(3):797–809, 2022. 1, 2

    2022

  23. [31]

    Uncertainty-guided dual- views for semi-supervised volumetric medical image segmentation

    Himashi Peiris, Munawar Hayat, Zhaolin Chen, Gary Egan, and Mehrtash Harandi. Uncertainty-guided dual- views for semi-supervised volumetric medical image segmentation. Nature Machine Intelligence , 5(7):724– 738, 2023. 6, 7, 8

    2023

  24. [32]

    Gradient-aware for class-imbalanced semi-supervised medical image seg- mentation

    Wenbo Qi, Jiafei Wu, and SC Chan. Gradient-aware for class-imbalanced semi-supervised medical image seg- mentation. In Eur. Conf. Comput. Vis. , 2024. 1, 2, 3, 5, 6, 7, 8

    2024

  25. [33]

    Semi-supervised ct lesion segmen- tation using uncertainty-based data pairing and swapmix

    Pengchong Qiao, Han Li, Guoli Song, Hu Han, Zhiqiang Gao, Yonghong Tian, Yongsheng Liang, Xi Li, S Kevin Zhou, and Jie Chen. Semi-supervised ct lesion segmen- tation using uncertainty-based data pairing and swapmix. IEEE Trans. on Medical Imaging, 2022. 1, 2

    2022

  26. [34]

    Deep co-training for semi-supervised image recognition

    Siyuan Qiao, Wei Shen, Zhishuai Zhang, Bo Wang, and Alan Yuille. Deep co-training for semi-supervised image recognition. In Eur. Conf. Comput. Vis., pages 135–152,

  27. [35]

    Learning transferable visual models from natural lan- guage supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sas- try, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural lan- guage supervision. In Proc. of Intl. Conf. on Machine Learning, pa...

    2021

  28. [36]

    Learning transferable visual models from natural lan- guage supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sas- try, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural lan- guage supervision. In Proc. of Intl. Conf. on Machine Learning, pa...

    2021

  29. [37]

    Meta-weight-net: Learn- ing an explicit mapping for sample weighting

    Jun Shu, Qi Xie, Lixuan Yi, Qian Zhao, Sanping Zhou, Zongben Xu, and Deyu Meng. Meta-weight-net: Learn- ing an explicit mapping for sample weighting. Adv. Neu- ral Inform. Process. Syst., 32, 2019. 2

    2019

  30. [38]

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

    Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Adv. Neural Inform. Process. Syst., 30, 2017. 2

    2017

  31. [39]

    Semi-supervised segmentation of radiation-induced pul- monary fibrosis from lung ct scans with multi-scale guided dense attention

    Guotai Wang, Shuwei Zhai, Giovanni Lasio, Baoshe Zhang, Byong Yi, Shifeng Chen, Thomas J Macvittie, Dimitris Metaxas, Jinghao Zhou, and Shaoting Zhang. Semi-supervised segmentation of radiation-induced pul- monary fibrosis from lung ct scans with multi-scale guided dense atten...

    2021

  32. [40]

    DHC: Dual- debiased heterogeneous co-training framework for class- imbalanced semi-supervised medical image segmenta- tion

    Haonan Wang and Xiaomeng Li. DHC: Dual- debiased heterogeneous co-training framework for class- imbalanced semi-supervised medical image segmenta- tion. In Proc. of MICCAI, pages 582–591, 2023. 1, 2, 3, 5, 6, 7, 8

    2023

  33. [41]

    Towards generic semi- supervised framework for volumetric medical image seg- mentation

    Haonan Wang and Xiaomeng Li. Towards generic semi- supervised framework for volumetric medical image seg- mentation. Adv. Neural Inform. Process. Syst., 36, 2024. 6, 7

    2024

  34. [42]

    Debiased learning from naturally imbalanced pseudo- labels

    Xudong Wang, Zhirong Wu, Long Lian, and Stella X Yu. Debiased learning from naturally imbalanced pseudo- labels. In IEEE Conf. Comput. Vis. Pattern Recog., pages 14647–14657, 2022. 6, 7

    2022

  35. [43]

    Hunting sparsity: Density-guided contrastive learning for semi-supervised semantic segmentation

    Xiaoyang Wang, Bingfeng Zhang, Limin Yu, and Jimin Xiao. Hunting sparsity: Density-guided contrastive learning for semi-supervised semantic segmentation. In IEEE Conf. Comput. Vis. Pattern Recog. , pages 3114– 3123, 2023. 2

    2023

  36. [44]

    Crest: A class-rebalancing self-training framework for imbalanced semi-supervised learning

    Chen Wei, Kihyuk Sohn, Clayton Mellina, Alan Yuille, and Fan Yang. Crest: A class-rebalancing self-training framework for imbalanced semi-supervised learning. In IEEE Conf. Comput. Vis. Pattern Recog. , pages 10857– 10866, 2021. 3, 6, 7

    2021

  37. [45]

    FUSSNet: Fus- ing two sources of uncertainty for semi-supervised medi- cal image segmentation

    Jinyi Xiang, Peng Qiu, and Yang Yang. FUSSNet: Fus- ing two sources of uncertainty for semi-supervised medi- cal image segmentation. InProc. of MICCAI, pages 481– 491, 2022. 1, 2

    2022

  38. [46]

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

    Lequan Yu, Shujun Wang, Xiaomeng Li, Chi-Wing Fu, and Pheng-Ann Heng. Uncertainty-aware self- ensembling model for semi-supervised 3d left atrium segmentation. In Proc. of MICCAI , pages 605–613,

  39. [47]

    Biomedclip: a multi- modal biomedical foundation model pretrained from fif- teen million scientific image-text pairs

    Sheng Zhang, Yanbo Xu, Naoto Usuyama, Hanwen Xu, Jaspreet Bagga, Robert Tinn, Sam Preston, Rajesh Rao, Mu Wei, Naveen Valluri, et al. Biomedclip: a multi- modal biomedical foundation model pretrained from fif- teen million scientific image-text pairs. arXiv preprint arXiv:2303...

    2023 arXiv

  40. [48]

    Continual learning for abdominal multi-organ and tumor segmentation

    Yixiao Zhang, Xinyi Li, Huimiao Chen, Alan L Yuille, Yaoyao Liu, and Zongwei Zhou. Continual learning for abdominal multi-organ and tumor segmentation. InProc. of MICCAI, pages 35–45, 2023. 3 11

    2023

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.