Pith. sign in

REVIEW 4 major objections 5 minor 55 references

Inpainting is All You Need: A Diffusion-based Augmentation Method for Semi-supervised Medical Image Segmentation

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Using diffusion inpainting to repaint backgrounds around labeled organs generates synthetic image-label pairs that train better segmenters than semi-supervised or self-supervised methods when labels are scarce.

desk verdict AugPaint is a plausible, cheap augmentation trick with consistent gains on four medical datasets, but its central label-fidelity guarantee is asserted rather than measured and needs one direct experiment before the results are trusted. read the letter →

arxiv 2506.23038 v1 pith:Q33EU55K submitted 2025-06-28 cs.CV

Xinrong Hu , Yiyu Shi This is my paper
classification cs.CV
keywords medicalimagesegmentationdataaugmentationlatentdiffusionmodelsinpaintingsemi-supervisedlearninglimitedannotationssyntheticimage-labelpairslabel-efficient
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 argues that inpainting, not full-image generation, is the right way to synthesize training data for medical image segmentation under limited labels. The proposed AugPaint method takes a labeled image, keeps the pixels inside the labeled foreground region, and uses a latent diffusion model to paint a new background around it, yielding a new image that is guaranteed (by construction) to share the same label mask. The authors claim that these synthetic pairs, used as supervision for a U-Net, outperform state-of-the-art self-supervised and semi-supervised approaches on four medical datasets covering cardiac MRI, brain MRI, abdominal CT, and skin lesions. If correct, this offers a simple, plug-and-play augmentation that converts a handful of annotations into a much larger training set without retraining the diffusion model.

What carries the argument

The central object is AugPaint, an inpainting procedure built on a latent diffusion model (LDM) that operates in a compressed latent space. Given an image-label pair, the labeled foreground is converted into a bounding-box mask, resized to the latent resolution, and at every denoising step the latent pixels inside the mask are overwritten with the noised ground-truth latent code while the background is sampled from the unconditional denoiser. The mechanism that carries the argument is the combination of (1) box-style masking instead of label-shape masking, which preserves organ boundaries and prevents the model from expanding organ shapes, and (2) the use of an unconditional LDM with DDIM accelerated sampling, which maintains diversity and avoids the reconstruction collapse of conditional inpainting. The resulting decoded images are paired with the original label mask, giving exact pixel-level supervision without relying on pseudo-labels or cross-attention localization.

What would settle it

Generate a batch of AugPaint samples and run a fully supervised, high-accuracy segmentation model on them; if a substantial fraction of the synthetic images have predicted segmentations that do not match the paired label masks, the label-fidelity premise fails and the claimed supervision signal would be corrupted.

Watch

Extended reading notes

Core claim

The central claim is that training a segmentation model on data augmented by AugPaint—latent diffusion inpainting that preserves the labeled foreground and repaints the background—yields higher dice scores than training with existing label-efficient methods, under very limited labeled cases (1, 2, or 5 cases for volumetric scans; 10, 20, or 50 for skin lesions). The paper reports consistent gains on all four datasets, with average dice improvements over training from scratch of 20.40 on ACDC, 6.13 on BraTS, 21.07 on Synapse, and 4.77 on ISIC 2018. The paper also shows that an unconditional latent diffusion model, rather than a conditional one, avoids the collapse of inpainting into simple reconstruction of the input image, preserving diversity in the augmented dataset. A filtering step, which keeps the top half of generated images by a confidence score from a segmentation model trained on the labeled data, is used to remove poor-quality or label-mismatched samples.

Load-bearing premise

The binary label mask, resized to the latent space, lines up with the encoded foreground regions, so overwriting the latent background while keeping the masked latent pixels preserves the true segmentation in the decoded image.

Editorial extensions

If this is right

  • AugPaint can be combined with existing self-supervised or semi-supervised methods to push segmentation performance further than either approach alone.
  • The augmentation benefit transfers across segmentation architectures, improving U-Net, DeepLabV3, and TransUNet when trained on the same synthetic dataset.
  • Generating roughly 5–10 inpainted samples per labeled image is a practical balance between computational cost and segmentation accuracy, since too many samples degrade performance.
  • The benefit of AugPaint is largest when labels are extremely scarce and shrinks as more labeled data becomes available, because the diffusion model is trained only on the available dataset.
  • Box-style masking is critical: using the exact label shape as the condition mask can actually hurt performance on datasets like ISIC 2018, where generated samples mismatch the labels.

Reading between the lines

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

  • A testable extension is to apply the same box-preserving inpainting to other dense prediction tasks (e.g., instance segmentation or panoptic segmentation) where foreground shape is roughly known but background diversity is high.
  • The paper's label-fidelity guarantee rests on an empirical alignment assumption; if that assumption fails for higher compression rates or larger latent downsampling factors, the method would need a correction step to re-align masks in latent space.
  • The filtering mechanism—keeping the top half by a confidence score—could be replaced by a learned uncertainty or out-of-distribution detector, which the authors themselves flag as a limitation; improving this component may allow more aggressive sampling without contamination.
  • The approach suggests a general recipe for semi-supervised segmentation: instead of generating entire images from labels, keep the labeled part intact and synthesize only the unlabeled context, which turns annotation scarcity into a background-completion problem.
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

4 major / 5 minor

Summary. The paper proposes AugPaint, a data augmentation method for semi-supervised medical image segmentation that uses a latent diffusion model to inpaint the background region outside a bounding box around labeled foreground structures. Each original image-label pair thus yields multiple synthetic image-label pairs, which are used to augment the scarce labeled set when training a downstream segmentation network. The method is evaluated on four medical imaging datasets (ACDC cardiac MRI, BraTS brain MRI, Synapse abdominal CT, and ISIC 2018 skin lesions) under limited-label settings (1, 2, and 5 labeled cases, or 10/20/50 for ISIC). The reported dice-score tables show consistent improvements over several self-supervised and semi-supervised baselines, and further gains when AugPaint is combined with those baselines. The paper also includes ablations on mask type, number of generated samples, segmentation architecture, and a mask-flipping variant for BraTS.

Significance. If the results hold, AugPaint is a practically attractive augmentation method: the unconditional version requires no retraining of the diffusion model for inpainting, it is substantially faster than RePaint, and the reported gains across four diverse modalities and multiple architectures are consistent. The paper also demonstrates a simple plug-and-play combination with existing label-efficient methods, which is a useful property. The main strength is the breadth of experiments and the clear framing of the label-fidelity claim as the method's differentiator. However, that claim is asserted rather than directly measured, and the absence of statistical significance testing and of comparisons with cited diffusion-based augmentation baselines currently leaves the strength of the central claim partially unverified.

major comments (4)
  1. [§3.2, Algorithm 1] The load-bearing assumption of label fidelity is asserted but not quantitatively verified. The pipeline outputs D(ẑ0), and the autoencoder trained at f=4 is not guaranteed to preserve the exact segmentation boundary of the preserved region; a decoded image can have a foreground that does not match y0. Moreover, because the mask is a bounding box, the outer background is entirely sampled, so the generator can paint a new organ-like structure there; the original label y0 then incorrectly omits that structure. The filter in §3.3 scores confidence only 'within the labeled areas', so it cannot catch a spurious foreground outside the label. Please measure label correctness directly (e.g., compute dice between y0 and the prediction of a strong independently trained segmenter on the generated images, or perform a human study on a random sample), and report the per-dataset mismatch rate. Without this measurement, the reported dice improvements in Tables 1–2 could reflect distributional mimicry or easy-sample selection rather than correct synthetic supervision.
  2. [Tables 1–2] The headline claim that AugPaint 'outperforms state-of-the-art' is not supported by any statistical significance test. All results are reported as mean ± std over three independent splits, and in several settings the differences are within one standard deviation (e.g., BraTS with 1 case: AugPaint 35.19±7.96 vs. PCL 31.92±9.03; ACDC with 2 cases: AugPaint 46.15±6.98 vs. UniMatch 39.23±8.63, though this one is larger). With only three splits, paired tests (e.g., Wilcoxon signed-rank or paired t-test over the per-split dice values) are feasible and should be reported for the main comparisons. This is essential to establish that the empirical advantage is not noise.
  3. [§2, §4.3] The paper cites several diffusion-based data augmentation methods for medical image segmentation (Refs [15,16,23,50]) but does not compare against any of them; the only generative baseline is ControlNet. Given that the paper's positioning is explicitly against generative augmentation methods and that the abstract claims superiority over 'state-of-the-art label-efficient methodologies,' the absence of these most-related baselines is a significant gap. At least one of the cited diffusion augmentation approaches (or an equivalent implementation under the same limited-label protocol) should be included in Tables 1–2, or the authors should explain concretely why such comparisons are not feasible or not meaningful.
  4. [§3.3, Limitations] The confidence-based filtering mechanism creates a potential selection bias: the filter model is trained on the very same limited labels used to generate synthetic data, so it may preferentially retain images that resemble the training set rather than images with correct labels. The Limitations section itself calls the filter 'relatively rudimentary.' The Rand Aug baseline does use the same filtering rule, which partially controls for the filter's effect, but Rand Aug does not face label-mismatch risk, so the interaction between filtering and synthetic label correctness is not isolated. Please add an ablation with (a) no filtering, (b) filtering with a model trained on the full labeled set, and (c) an oracle filter that uses ground-truth labels of the generated images. This will clarify how much of the reported gain is due to the generative mechanism versus the filtering strategy.
minor comments (5)
  1. [Table 3] The ACDC '2 cases' row reports AugPaint alone as 44.15, while Table 1 reports 46.15±6.98 for the same setting. One of these is a typo; please correct and verify all repeated numbers across tables.
  2. [Algorithm 1] The notation in line 6 uses 'xt' without definition; the algorithm mixes x0, z0, and zt. Please use a consistent notation (e.g., zt = √α_t z0 + ...) and define τ_t and ̄z0 in the text or caption.
  3. [§4.2] The runtime statement 'It takes average5.42sto generate 8 samples' is missing a space and should report the per-sample latency with the same hardware for both AugPaint and RePaint to substantiate the '20× faster' claim.
  4. [Table 6] The column headers '1(10) case' and '5(50) cases' are ambiguous; clarify that the numbers in parentheses refer to ISIC labeled counts and specify which dataset each sub-table uses.
  5. [Figure 2] In the caption, '×(T-1)' is unexplained; it should read 'repeated T-1 times' or similar.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; the central performance claim is an empirical comparison, and the label-fidelity premise is an unverified assumption rather than a circular derivation.

full rationale

The paper's central claim is empirical: synthetic pairs produced by AugPaint improve downstream dice scores against SSL and semi-supervised baselines, supported by held-out test measurements in Tables 1-3. This is not a derivation that reduces to its own inputs. The only load-bearing premise not directly measured is label fidelity: Section 3.2 asserts that 'our empirical results demonstrate that the resized mask aligns effectively with corresponding areas in z_t, which ensures that labeled areas from the original input x_0 are accurately preserved in the latent representation.' This is an unverified assumption and a genuine correctness risk, but it is not circular—if it fails, the method's conclusions would be invalidated empirically rather than exposed as tautological. The self-references ([21], [51], [53]) are used for related work and baselines, not to justify the central contribution, and no uniqueness theorem or ansatz is imported from them. The filtering mechanism in Section 3.3 trains a model on the same limited labels and selects the top half by confidence; the Limitations section itself calls this filter 'relatively rudimentary and simply relies on a model trained with few labeled data.' That is a disclosed methodological weakness and a potential self-reinforcing data-selection loop, but the final dice scores are still evaluated on independent test data, so it is not a case where a fitted parameter is renamed as a prediction. Overall, no step in the paper's derivation chain is circular; score 1 reflects the presence of the unverified label-fidelity assumption and minor self-citations that are not load-bearing.

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

The method has no fitted physical constants and introduces no new objects. Its reliance is on empirical hyperparameters and dataset assumptions about anatomical consistency and latent-space alignment.

free parameters (5)
  • number of inpainted samples per labeled image = 5 (main), 5-10 recommended
    Chosen by hand; Fig 5(a) shows dice peaks around 30 and declines after, so the choice affects results.
  • filter retention rate = 0.5 (top half)
    Chosen by hand; no ablation of filtering strength is provided.
  • DDIM sampling steps = 50
    Standard choice; trade-off between quality and speed.
  • mask type = bounding box (not label shape)
    Chosen by design; Table 6 shows it outperforms label-shape masking, but it changes the conditioning information.
  • mask flip probability (BraTS only) = not specified
    Randomly flip masked image; probability is not reported.
assumptions (5)
  • domain assumption A latent diffusion model trained on all available images (labeled and unlabeled) learns a distribution p(x) that enables realistic inpainting of background areas
    AugPaint's quality rests on this; invoked throughout §3.
  • domain assumption Resizing a binary mask to the latent dimensions preserves alignment with the encoded foreground
    Invoked in §3.2: 'our empirical results demonstrate that the resized mask aligns effectively with corresponding areas in z_t'.
  • domain assumption Medical datasets have small foreground relative to background and consistent background, making background inpainting feasible
    Stated in §1 as motivation for the method.
  • ad hoc to paper A segmentation model trained on the scarce labeled data yields confidence scores that reliably separate good from bad synthetic samples
    Filtering step in §3.3; the limitations section calls it 'relatively rudimentary'.
  • standard math The DDIM update equations in Algorithm 1 correctly implement accelerated reverse diffusion
    Borrowed from [35, 41]; the algorithm as typeset contains typos, e.g., the tau_t line references 'xt'.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Inpainting is All You Need: A Diffusion-based Augmentation Method for Semi-supervised Medical Image Segmentation." pith.science (2026). https://pith.science/paper/Q33EU55K

@misc{pith2026250623038,
  author       = {Pith},
  title        = {Pith review of: Inpainting is All You Need: A Diffusion-based Augmentation Method for Semi-supervised Medical Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q33EU55K}},
  note         = {Machine review of arXiv:2506.23038}
}
read the original abstract

Collecting pixel-level labels for medical datasets can be a laborious and expensive process, and enhancing segmentation performance with a scarcity of labeled data is a crucial challenge. This work introduces AugPaint, a data augmentation framework that utilizes inpainting to generate image-label pairs from limited labeled data. AugPaint leverages latent diffusion models, known for their ability to generate high-quality in-domain images with low overhead, and adapts the sampling process for the inpainting task without need for retraining. Specifically, given a pair of image and label mask, we crop the area labeled with the foreground and condition on it during reversed denoising process for every noise level. Masked background area would gradually be filled in, and all generated images are paired with the label mask. This approach ensures the accuracy of match between synthetic images and label masks, setting it apart from existing dataset generation methods. The generated images serve as valuable supervision for training downstream segmentation models, effectively addressing the challenge of limited annotations. We conducted extensive evaluations of our data augmentation method on four public medical image segmentation datasets, including CT, MRI, and skin imaging. Results across all datasets demonstrate that AugPaint outperforms state-of-the-art label-efficient methodologies, significantly improving segmentation performance.

Figures

Figures reproduced from arXiv: 2506.23038 by the authors.

Figure 1
Figure 1. Samples generated with our AugPaint on four medical image segmentation datasets. Based on labeled data, we only keep boxes [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Illustration of sampling process via our AugPaint. Given a unconditional latent diffusion model trained on all data, we modify the [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Comparisons of inpainting samples between with con [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Inpainting with flipped masks on BraTS dataset. The [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: The influence of number of sampled images and number [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

55 extracted references · 41 canonical work pages

  1. [1]

    Accessed: 2021-09-06

    “multi-atlas labeling beyond the cranial vault.https: //www.synapse.org/#!Synapse:syn3193805/ wiki/89480. Accessed: 2021-09-06. 5

    work page 2021

  2. [2]

    Diffusion-based data augmentation for skin disease classification: Impact across original medical datasets to fully synthetic images

    Mohamed Akrout, B ´alint Gyepesi, P ´eter Holl ´o, Adrienn Po´or, Bl´aga Kincs ˝o, Stephen Solis, Katrina Cirone, Jeremy Kawahara, Dekker Slade, Latif Abid, et al. Diffusion-based data augmentation for skin disease classification: Impact across original medical datasets to fully synthetic images. arXiv preprint arXiv:2301.04802, 2023. 1

    arXiv 2023

  3. [3]

    The medical segmentation decathlon.Nature communications, 13(1):1–13, 2022

    Michela Antonelli, Annika Reinke, Spyridon Bakas, Key- van Farahani, Annette Kopp-Schneider, Bennett A Landman, Geert Litjens, Bjoern Menze, Olaf Ronneberger, Ronald M Summers, et al. The medical segmentation decathlon.Nature communications, 13(1):1–13, 2022. 5

    work page 2022

  4. [4]

    Bidirectional copy-paste for semi-supervised medical image segmentation

    Yunhao Bai, Duowen Chen, Qingli Li, Wei Shen, and Yan Wang. Bidirectional copy-paste for semi-supervised medical image segmentation. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 11514–11524, 2023. 3

    work page 2023

  5. [5]

    Red blood cell image generation for data augmentation using con- ditional generative adversarial networks

    Oleksandr Bailo, DongShik Ham, and Young Min Shin. Red blood cell image generation for data augmentation using con- ditional generative adversarial networks. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition workshops, pages 0–0, 2019. 2

    work page 2019

  6. [6]

    Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved?IEEE transactions on medical imaging, 37(11):2514–2525, 2018

    Olivier Bernard, Alain Lalande, Clement Zotti, Freder- ick Cervenansky, Xin Yang, Pheng-Ann Heng, Irem Cetin, Karim Lekadir, Oscar Camara, Miguel Angel Gonzalez Ballester, et al. Deep learning techniques for automatic mri cardiac multi-structures segmentation and diagnosis: is the problem solved?IEEE transactions on medical imaging, 37(11):2514–2525, 2018. 5

    work page 2018

  7. [7]

    Extracting training data from diffu- sion models

    Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagiel- ski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ip- polito, and Eric Wallace. Extracting training data from diffu- sion models. In32nd USENIX Security Symposium (USENIX Security 23), pages 5253–5270, 2023. 5

    work page 2023

  8. [8]

    Krishna Chaitanya, Ertunc Erdil, Neerav Karani, and Ender Konukoglu. Contrastive learning of global and local fea- tures for medical image segmentation with limited annota- tions.Advances in Neural Information Processing Systems, 33:12546–12558, 2020. 1, 3, 5, 6

    work page 2020

Show all 55 references
  1. [9]

    Semi-supervised task-driven data augmentation for medical image segmentation.Medical Image Analysis, 68:101934, 2021

    Krishna Chaitanya, Neerav Karani, Christian F Baumgart- ner, Ertunc Erdil, Anton Becker, Olivio Donati, and Ender Konukoglu. Semi-supervised task-driven data augmentation for medical image segmentation.Medical Image Analysis, 68:101934, 2021. 2, 6

    2021

  2. [10]

    Transunet: Transformers make strong encoders for medi- cal image segmentation.arXiv preprint arXiv:2102.04306,

    Jieneng Chen, Yongyi Lu, Qihang Yu, Xiangde Luo, Ehsan Adeli, Yan Wang, Le Lu, Alan L Yuille, and Yuyin Zhou. Transunet: Transformers make strong encoders for medi- cal image segmentation.arXiv preprint arXiv:2102.04306,

    arXiv

  3. [11]

    Rethinking atrous convolution for seman- tic image segmentation.arXiv preprint arXiv:1706.05587,

    Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for seman- tic image segmentation.arXiv preprint arXiv:1706.05587,

    arXiv

  4. [12]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Ge- offrey Hinton. A simple framework for contrastive learning of visual representations. InInternational conference on ma- chine learning, pages 1597–1607. PMLR, 2020. 1

    2020

  5. [13]

    Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the interna- tional skin imaging collaboration (isic).arXiv preprint arXiv:1902.03368, 2019

    Noel Codella, Veronica Rotemberg, Philipp Tschandl, M Emre Celebi, Stephen Dusza, David Gutman, Brian Helba, Aadi Kalloo, Konstantinos Liopyris, Michael Marchetti, et al. Skin lesion analysis toward melanoma detection 2018: A challenge hosted by the interna- tional skin imagin...

    2018 arXiv

  6. [14]

    Diffusion models beat gans on image synthesis.Advances in neural informa- tion processing systems, 34:8780–8794, 2021

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis.Advances in neural informa- tion processing systems, 34:8780–8794, 2021. 1

    2021

  7. [15]

    Arsdm: Colonoscopy images synthesis with adaptive refinement se- mantic diffusion models

    Yuhao Du, Yuncheng Jiang, Shuangyi Tan, Xusheng Wu, Qi Dou, Zhen Li, Guanbin Li, and Xiang Wan. Arsdm: Colonoscopy images synthesis with adaptive refinement se- mantic diffusion models. InInternational Conference on Medical Image Computing and Computer-Assisted Interven- tion,...

    2023

  8. [16]

    Can segmentation models be trained with fully synthetically generated data? InInter- national Workshop on Simulation and Synthesis in Medical Imaging, pages 79–90

    Virginia Fernandez, Walter Hugo Lopez Pinaya, Pedro Borges, Petru-Daniel Tudosiu, Mark S Graham, Tom Ver- cauteren, and M Jorge Cardoso. Can segmentation models be trained with fully synthetically generated data? InInter- national Workshop on Simulation and Synthesis in Medica...

    2022

  9. [17]

    Medgen3d: A deep generative framework for paired 3d image and mask generation

    Kun Han, Yifeng Xiong, Chenyu You, Pooya Khosravi, Shanlin Sun, Xiangyi Yan, James S Duncan, and Xiaohui Xie. Medgen3d: A deep generative framework for paired 3d image and mask generation. InInternational Conference on Medical Image Computing and Computer-Assisted Interven- ti...

    2023

  10. [18]

    Unetr: Transformers for 3d medical image segmentation

    Ali Hatamizadeh, Yucheng Tang, Vishwesh Nath, Dong Yang, Andriy Myronenko, Bennett Landman, Holger R Roth, and Daguang Xu. Unetr: Transformers for 3d medical image segmentation. InProceedings of the IEEE/CVF Win- ter Conference on Applications of Computer Vision, pages 574–584...

    2022

  11. [19]

    Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574, 2022

    Ruifei He, Shuyang Sun, Xin Yu, Chuhui Xue, Wenqing Zhang, Philip Torr, Song Bai, and Xiaojuan Qi. Is synthetic data from generative models ready for image recognition? arXiv preprint arXiv:2210.07574, 2022. 2

    2022 arXiv

  12. [20]

    Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models.Advances in neural information processing systems, 33:6840–6851, 2020. 1

    2020

  13. [21]

    Semi- supervised contrastive learning for label-efficient medical image segmentation

    Xinrong Hu, Dewen Zeng, Xiaowei Xu, and Yiyu Shi. Semi- supervised contrastive learning for label-efficient medical image segmentation. InInternational Conference on Med- ical Image Computing and Computer-Assisted Intervention, pages 481–490. Springer, 2021. 1, 3

    2021

  14. [22]

    Deep learning approaches for data augmentation in medical imaging: A review.Journal of Imaging, 9(4):81, 2023

    Aghiles Kebaili, J ´erˆome Lapuyade-Lahorgue, and Su Ruan. Deep learning approaches for data augmentation in medical imaging: A review.Journal of Imaging, 9(4):81, 2023. 2

    2023

  15. [23]

    Medical diffusion– denoising diffusion probabilistic models for 3d medical im- age generation.arXiv preprint arXiv:2211.03364, 2022

    Firas Khader, Gustav Mueller-Franzes, Soroosh Tayebi Arasteh, Tianyu Han, Christoph Haarburger, Maximilian Schulze-Hagen, Philipp Schad, Sandy Engelhardt, Bettina Baessler, Sebastian Foersch, et al. Medical diffusion– denoising diffusion probabilistic models for 3d medical im-...

    2022 arXiv

  16. [24]

    Semantic segmentation with generative models: Semi-supervised learning and strong out-of-domain generalization

    Daiqing Li, Junlin Yang, Karsten Kreis, Antonio Torralba, and Sanja Fidler. Semantic segmentation with generative models: Semi-supervised learning and strong out-of-domain generalization. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, page...

    2021

  17. [25]

    Repaint: Inpainting using denoising diffusion probabilistic models

    Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11461–11471, 2022. 1, 4, 6

    2022

  18. [26]

    Semi-supervised medical image segmen- tation via cross teaching between cnn and transformer

    Xiangde Luo, Minhao Hu, Tao Song, Guotai Wang, and Shaoting Zhang. Semi-supervised medical image segmen- tation via cross teaching between cnn and transformer. In International Conference on Medical Imaging with Deep Learning, pages 820–833. PMLR, 2022. 3, 5, 6

    2022

  19. [27]

    Learning data aug- mentation for brain tumor segmentation with coarse-to-fine generative adversarial networks

    Tony CW Mok and Albert CS Chung. Learning data aug- mentation for brain tumor segmentation with coarse-to-fine generative adversarial networks. InBrainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries: 4th International Workshop, BrainLes 2018, Held in Co...

    2018

  20. [28]

    Data augmentation for brain-tumor segmentation: a review

    Jakub Nalepa, Michal Marcinkiewicz, and Michal Kawulok. Data augmentation for brain-tumor segmentation: a review. Frontiers in computational neuroscience, 13:83, 2019. 2

    2019

  21. [29]

    Generative adversarial network based synthesis for supervised medical image segmentation

    Thomas Neff, Christian Payer, Darko Stern, and Martin Urschler. Generative adversarial network based synthesis for supervised medical image segmentation. InProc. OAGM and ARW joint Workshop, volume 3, page 4, 2017. 1, 2

    2017

  22. [30]

    Dataset diffusion: Diffusion-based synthetic dataset gener- ation for pixel-level semantic segmentation.arXiv preprint arXiv:2309.14303, 2023

    Quang Nguyen, Truong Vu, Anh Tran, and Khoi Nguyen. Dataset diffusion: Diffusion-based synthetic dataset gener- ation for pixel-level semantic segmentation.arXiv preprint arXiv:2309.14303, 2023. 2

    2023 arXiv

  23. [31]

    Improved denoising diffusion probabilistic models

    Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. InInternational Conference on Machine Learning, pages 8162–8171. PMLR,

  24. [32]

    Un- supervised medical image translation with adversarial diffu- sion models.IEEE Transactions on Medical Imaging, 2023

    Muzaffer ¨Ozbey, Onat Dalmaz, Salman UH Dar, Hasan A Bedel, S ¸aban¨Ozturk, Alper G ¨ung¨or, and Tolga C ¸ ukur. Un- supervised medical image translation with adversarial diffu- sion models.IEEE Transactions on Medical Imaging, 2023. 1

    2023

  25. [33]

    Self-paced contrastive learning for semi-supervised medical image segmentation with meta- labels.Advances in Neural Information Processing Systems, 34:16686–16699, 2021

    Jizong Peng, Ping Wang, Christian Desrosiers, and Marco Pedersoli. Self-paced contrastive learning for semi-supervised medical image segmentation with meta- labels.Advances in Neural Information Processing Systems, 34:16686–16699, 2021. 3

    2021

  26. [34]

    Brain imaging generation with latent diffusion models

    Walter HL Pinaya, Petru-Daniel Tudosiu, Jessica Dafflon, Pedro F Da Costa, Virginia Fernandez, Parashkev Nachev, Sebastien Ourselin, and M Jorge Cardoso. Brain imaging generation with latent diffusion models. InMICCAI Work- shop on Deep Generative Models, pages 117–126. Springer,

  27. [35]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 2, 6

    2022

  28. [36]

    Palette: Image-to-image diffusion models

    Chitwan Saharia, William Chan, Huiwen Chang, Chris Lee, Jonathan Ho, Tim Salimans, David Fleet, and Mohammad Norouzi. Palette: Image-to-image diffusion models. In ACM SIGGRAPH 2022 Conference Proceedings, pages 1– 10, 2022. 1

    2022

  29. [37]

    A novel data augmentation method using style-based gan for robust pul- monary nodule segmentation

    Haoqi Shi, Junguo Lu, and Qianjun Zhou. A novel data augmentation method using style-based gan for robust pul- monary nodule segmentation. In2020 Chinese Control and Decision Conference (CCDC), pages 2486–2491. IEEE,

  30. [38]

    Medical image synthesis for data augmentation and anonymization using generative adversarial networks

    Hoo-Chang Shin, Neil A Tenenholtz, Jameson K Rogers, Christopher G Schwarz, Matthew L Senjem, Jeffrey L Gunter, Katherine P Andriole, and Mark Michalski. Medical image synthesis for data augmentation and anonymization using generative adversarial networks. InSimulation and Syn...

    2018

  31. [39]

    Fixmatch: Simplifying semi-supervised learning with consistency and confidence

    Kihyuk Sohn, David Berthelot, Nicholas Carlini, Zizhao Zhang, Han Zhang, Colin A Raffel, Ekin Dogus Cubuk, Alexey Kurakin, and Chun-Liang Li. Fixmatch: Simplifying semi-supervised learning with consistency and confidence. Advances in neural information processing systems, 33:5...

    2020

  32. [40]

    Diffusion art or digital forgery? investigating data replication in diffusion models

    Gowthami Somepalli, Vasu Singla, Micah Goldblum, Jonas Geiping, and Tom Goldstein. Diffusion art or digital forgery? investigating data replication in diffusion models. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6048–6058, 2023. 5

    2023

  33. [41]

    Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models.arXiv preprint arXiv:2010.02502, 2020. 4

    2010 arXiv

  34. [42]

    Effective data augmentation with diffusion models.arXiv preprint arXiv:2302.07944, 2023

    Brandon Trabucco, Kyle Doherty, Max Gurinas, and Ruslan Salakhutdinov. Effective data augmentation with diffusion models.arXiv preprint arXiv:2302.07944, 2023. 2

    2023 arXiv

  35. [43]

    Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using dif- fusion models.arXiv preprint arXiv:2303.11681, 2023

    Weijia Wu, Yuzhong Zhao, Mike Zheng Shou, Hong Zhou, and Chunhua Shen. Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using dif- fusion models.arXiv preprint arXiv:2303.11681, 2023. 2

    2023 arXiv

  36. [44]

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

    Yicheng Wu, Zhonghua Wu, Qianyi Wu, Zongyuan Ge, and Jianfei Cai. Exploring smoothness and class-separation for semi-supervised medical image segmentation. InIn- ternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 34–43. Springer,

  37. [45]

    Uncertainty-aware multi-view co-training for semi-supervised medical image segmentation and do- main adaptation.Medical image analysis, 65:101766, 2020

    Yingda Xia, Dong Yang, Zhiding Yu, Fengze Liu, Jinzheng Cai, Lequan Yu, Zhuotun Zhu, Daguang Xu, Alan Yuille, and Holger Roth. Uncertainty-aware multi-view co-training for semi-supervised medical image segmentation and do- main adaptation.Medical image analysis, 65:101766, 2020. 3

    2020

  38. [46]

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

    Lihe Yang, Lei Qi, Litong Feng, Wayne Zhang, and Yinghuan Shi. Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7236–7246, 2023. 1, 5, 6

    2023

  39. [47]

    Bootstrapping semi-supervised medical image segmentation with anatomical-aware contrastive dis- tillation

    Chenyu You, Weicheng Dai, Yifei Min, Lawrence Staib, and James S Duncan. Bootstrapping semi-supervised medical image segmentation with anatomical-aware contrastive dis- tillation. InInternational Conference on Information Pro- cessing in Medical Imaging, pages 641–653. Springe...

    2023

  40. [48]

    Momentum contrastive voxel-wise representation learning for semi-supervised volumetric medical image seg- mentation

    Chenyu You, Ruihan Zhao, Lawrence H Staib, and James S Duncan. Momentum contrastive voxel-wise representation learning for semi-supervised volumetric medical image seg- mentation. InInternational Conference on Medical Image Computing and Computer-Assisted Intervention, pages 639–

  41. [49]

    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. InMedi- cal Image Computing and Computer Assisted Intervention– MICCAI 2019: 22nd International Conference, Shenzhen, ...

    2019

  42. [50]

    Diffusion-based data augmentation for nuclei image segmentation

    Xinyi Yu, Guanbin Li, Wei Lou, Siqi Liu, Xiang Wan, Yan Chen, and Haofeng Li. Diffusion-based data augmentation for nuclei image segmentation. InInternational Conference on Medical Image Computing and Computer-Assisted Inter- vention, pages 592–602. Springer, 2023. 2, 3

    2023

  43. [51]

    Positional contrastive learning for volumetric medical image segmentation

    Dewen Zeng, Yawen Wu, Xinrong Hu, Xiaowei Xu, Haiyun Yuan, Meiping Huang, Jian Zhuang, Jingtong Hu, and Yiyu Shi. Positional contrastive learning for volumetric medical image segmentation. InInternational Conference on Med- ical Image Computing and Computer-Assisted Interventi...

    2021

  44. [52]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3836–3847, 2023. 5, 6

    2023

  45. [53]

    Unsupervised feature clustering im- proves contrastive representation learning for medical image segmentation

    Yejia Zhang, Xinrong Hu, Nishchal Sapkota, Yiyu Shi, and Danny Z Chen. Unsupervised feature clustering im- proves contrastive representation learning for medical image segmentation. In2022 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), pages 1820–1823....

    2022

  46. [54]

    Datasetgan: Efficient labeled data factory with minimal human effort

    Yuxuan Zhang, Huan Ling, Jun Gao, Kangxue Yin, Jean- Francois Lafleche, Adela Barriuso, Antonio Torralba, and Sanja Fidler. Datasetgan: Efficient labeled data factory with minimal human effort. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition...

    2021

  47. [55]

    Data augmentation using learned transformations for one-shot medical image segmentation

    Amy Zhao, Guha Balakrishnan, Fredo Durand, John V Gut- tag, and Adrian V Dalca. Data augmentation using learned transformations for one-shot medical image segmentation. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 8543–8553, 2019. 2

    2019

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.