Adaptively Distilled ControlNet: Accelerated Training and Superior Sampling for Medical Image Synthesis

Kunpeng Qiu; Yongxin Guo; Zhiying Zhou

REVIEW 5 major objections 8 minor 36 references

Adaptively Distilled ControlNet: Accelerated Training and Superior Sampling for Medical Image Synthesis

T0 review · 5 major / 8 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 A mask-only diffusion model, trained by distillation from a mask-image teacher, generates synthetic medical images that improve downstream segmentation beyond real-data baselines.

desk verdict A coherent, useful incremental contribution to mask-conditioned medical image synthesis, but the central causal claim about mask-lesion alignment is asserted, not measured. read the letter →

arxiv 2507.23652 v1 pith:N7DTEJ3G submitted 2025-07-31 cs.CV

Kunpeng Qiu , Zhiying Zhou , Yongxin Guo This is my paper

classification cs.CV

keywords medicalimagesynthesisdiffusionmodelsControlNetknowledgedistillationlesion-maskalignmentsyntheticdataaugmentationsegmentationadaptiveloss

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 tries to show that a mask-only diffusion model, trained with the help of a mask-image-conditioned teacher, can generate synthetic medical images that serve as better training data for downstream segmentation than the real images alone. The method, Adaptively Distilled ControlNet, claims to accelerate training convergence, preserve sampling speed, and improve lesion-mask alignment through an adaptive distillation loss. If the reported gains hold, synthetic data from mask inputs could replace or augment real annotated images in privacy-sensitive medical settings.

What carries the argument

The load-bearing mechanism is a dual-branch teacher-student architecture sharing a frozen VAE encoder and a forward noise process. The teacher branch conditions on both the mask and the real image, while the student conditions on the mask only; the student's predicted noise is regularized toward the teacher's via a stop-gradient distillation loss. An adaptive weight matrix, derived from the lesion-background pixel ratio, upweights the distillation loss in lesion regions. During sampling only the student runs, so generation proceeds at ControlNet speed using arbitrary masks.

What would settle it

Compute the Dice coefficient (or IoU) between each input mask and the binarized lesion region of the corresponding generated image for both Adaptively Distilled ControlNet and ControlNet on KiTS19 and Polyps. If alignment is not significantly higher for the proposed method, the paper's claimed causal link between the adaptive distillation loss, alignment, and segmentation gains is unsupported.

Watch

Extended reading notes

Core claim

The central claim is that distilling a mask-image-conditioned ControlNet teacher into a mask-only student, with a spatially adaptive loss that upweights lesion regions, yields synthetic medical images whose mask-lesion alignment is good enough to improve downstream segmentation models. On KiTS19, TransUNet trained with the synthetic plus real data improves mDice by 2.4% and mIoU by 4.2% over the real-data baseline; on Polyps, SANet improves mDice by 2.6% and mIoU by 3.5%. The paper attributes these improvements to the adaptive distillation loss, which is shown in an ablation to outperform both the standard distillation loss and the no-distillation baseline.

Load-bearing premise

The paper assumes that the downstream segmentation improvements are caused by better mask-lesion alignment, but it never directly measures alignment; the gains could come from other differences in the synthetic image distribution, such as texture or shape.

Editorial extensions

If this is right

If the method is correct, segmentation models can be trained with synthetic images generated purely from masks, without needing real patient images at sampling time, which preserves privacy.
The task-agnostic design suggests the same training recipe transfers to other modalities beyond CT and endoscopy, such as MRI or X-ray, without architecture changes.
Because only the student model runs at inference, the approach keeps ControlNet's sampling speed, making it usable in compute-limited settings.
The ablation indicates that the adaptive distillation loss is the component responsible for the gains, so removing it should revert the improvements.

Reading between the lines

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

One could directly test the paper's causal story by computing the Dice or IoU between input masks and generated lesion regions; if alignment is not higher than ControlNet's, the attribution to mask-lesion alignment would need revision.
The adaptive weight, defined by a simple lesion-background ratio, could be extended to multi-class masks or to weighting by lesion size, and could be evaluated on datasets with more extreme class imbalance.
A promising follow-up would be to add a perceptual or adversarial loss to narrow the small FID gap to ControlNet while keeping the alignment benefit, which the paper itself does not do.

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.

Desk Editor's Note

read the letter

The paper is a legitimate incremental contribution: a teacher-student distillation framework where a mask-only student is regularized by an image-and-mask-conditioned teacher, with adaptive weighting on lesion regions. The combination is new in medical synthesis, and the payoff is real: downstream segmentation improves modestly but consistently across two datasets and three backbones, and sampling only needs masks, which is a genuine privacy win. The ablation shows the adaptive loss helps over both no distillation and uniform distillation, and training is cheap (3,000 steps). Credit is due for reporting the slightly worse FID rather than burying it, and for releasing code.

The soft spots are concentrated in the evidence for the paper's headline interpretation. The claim that segmentation gains come from superior mask-lesion alignment is never directly tested. No IoU/Dice between generated lesions and input masks, no edge error, no independent alignment metric. Table 1 shows FID slightly worse than ControlNet, and CLIP-I better, but neither measures spatial fidelity. The stress-test point is fair: since the teacher sees the real image and the student matches its noise prediction, L2 distillation drives the student toward the conditional mean teacher output, which could blur boundaries rather than sharpen them. So the alignment superiority is not a necessary consequence of the objective; it is an unvalidated assumption. That should be fixable with one simple experiment: compute mask-lesion IoU or Dice on generated samples for each method.

Two smaller issues. The 'accelerated training' claim rests on a qualitative figure and a sentence; a quantitative training loss curve or steps-to-target comparison would back it up. And there are no error bars or significance tests anywhere, which matters when the reported gains are a few points on segmentation metrics; some of the per-dataset improvements in Table 3 are within plausible noise. The Polyps improvement over ArSDM is modest overall, and the KiTS19 gains are clearer, so I would not over-read the numbers.

The citation pattern looks honest; the self-citation to the authors' CVPR work is not a load-bearing loop. The method does not fit parameters to the outcome metric, so circularity is not a concern.

Who is this for? Researchers who need a mask-only medical image generator for privacy-preserving data augmentation and who want a training recipe that does not require image conditions at sampling time. They will get a usable method and a fair comparison baseline. It deserves a serious referee, but not acceptance as-is: the causal claim needs a direct alignment metric, the significance issue needs at least error bars or a small number of seeds, and the convergence claim needs quantitative support. I would send it to peer review with a request for those revisions.

Referee Report

5 major / 8 minor

Summary. This paper proposes Adaptively Distilled ControlNet, a mask-conditioned diffusion framework for medical image synthesis. A teacher branch conditioned on image-mask pairs (cmix = ci + cm) distills into a mask-only student branch through an adaptive loss LAda that reweights the teacher-student noise-prediction discrepancy according to lesion-background pixel ratios; at inference only the student runs. Experiments on KiTS19 and Polyps compare FID/CLIP-I and downstream segmentation performance, reporting improved mDice/mIoU over real-data and ControlNet-augmented training. The paper also claims accelerated training convergence and superior lesion-mask alignment, but the latter is never directly measured.

Significance. If the reported segmentation gains are reproducible, the framework is practically useful: it moves the heavy image-conditioned teacher to training time only, keeping sampling privacy-preserving and at ControlNet speed, and it shows consistent downstream gains across two modalities and four segmentation architectures. The adaptive weighting is derived from mask statistics rather than tuned to the outcome metric, and downstream evaluation is on held-out data, so the core comparison is not circular. The availability of code is a further strength. However, the headline interpretive claims—superior alignment and accelerated training—are currently supported only by qualitative or indirect evidence, FID is slightly worse than the ControlNet baseline, and no variance estimates are reported; these gaps materially limit the strength of the conclusions.

major comments (5)

[Section 3.2, Eq. (7)] The central claim that the adaptive distillation objective yields superior mask-lesion alignment is not established by the derivation. Since the teacher condition is cmix = ci + cm (Eq. 3), the teacher's target depends on the specific real image paired with the mask; for a fixed mask the student is trained to match a distribution of teacher noise predictions under an L2 loss with a stop-gradient. The optimal student prediction is then the conditional mean of the teacher predictions, which can smooth or blur lesion boundaries rather than sharpen them. The paper does not provide any direct measurement of alignment, such as IoU or boundary F1 between the generated lesion and the conditioning mask, so the alignment superiority remains an unvalidated assumption rather than a consequence of Eq. (7). Please add a direct alignment metric and, if possible, an experiment that isolates alignment from other distributional changes.
[Section 4.4, Table 1] The causal interpretation of the downstream gains is not supported by the reported image-quality metrics. Table 1 shows that the proposed method's FID is slightly worse than ControlNet on both datasets (66.587 vs. 65.609 on Polyps; 70.786 vs. 69.240 on KiTS19), and CLIP-I measures global semantic similarity rather than spatial mask fidelity. The observation in Sec. 4.4 that there is no consistency between image quality and segmentation performance is an indirect argument, not a measurement of lesion-mask alignment. Please report direct alignment statistics or a controlled experiment (for example, training segmentation models on synthetic-only data with identical masks and sample counts across generators) before attributing the improvements to alignment.
[Tables 2-4] All segmentation and ablation results are reported as single runs without error bars, multiple seeds, or significance tests. Improvements of 1-3% mDice can be within run-to-run variance for segmentation models, especially for nnUNet with five-fold cross-validation; the absence of variance estimates makes it impossible to assess whether the reported differences are meaningful. Please provide mean and standard deviation over at least three independent training runs for the main comparisons and the ablation, or include statistical significance tests.
[Section 4.2, 4.3, Fig. 2] The 'accelerated training' claim is supported only by the qualitative statement in Sec. 4.3 that the student fits the data within about 300 steps, illustrated in Fig. 2. No quantitative training curves, iterations-to-target-FID, or wall-clock comparisons against ControlNet are provided. Please add quantitative convergence or timing measurements to substantiate this headline claim.
[Section 5, Table 4] The ablation of the adaptive distillation loss is conducted only on KiTS19; no corresponding ablation is reported for the Polyps dataset. Since the paper claims effectiveness across two distinct modalities and the adaptive loss is the key proposed component, the contribution of LAda to the RGB/endoscopy setting is unverified. Please add the same ablation on Polyps or justify its omission.

minor comments (8)

[Abstract] The phrase 'predicted noise alignment in parameter space' is inaccurate, because Eq. (7) aligns predicted noise values in latent space rather than model parameters; please rephrase.
[Section 4.2] The input resolution appears as '3842', which is likely a typo for 384x384; please correct it.
[Section 4.4] The statement that FID 'overfits with limited data' is not substantiated by the cited reference or an analysis; please either provide evidence or soften the claim.
[Figure 2] The axes are unlabeled, which makes the claimed convergence speed impossible to interpret quantitatively; please add axis labels and units.
[Section 4.2] The use of 5% prompt dropout and classifier-free guidance implies text prompts, but the exact prompts used for the two datasets are not specified; please provide them.
[Section 1] The term 'field distillation framework' is undefined and nonstandard; please use standard terminology or define the term.
[Table 4] The settings 'w/(Standard)' and 'w/(Adaptive)' are not defined in the text; please clarify how the standard distillation loss differs from the adaptive one.
[Tables 2-3] The comparison set is limited to SinGAN, ArSDM, T2I-Adapter, and ControlNet; given that the manuscript cites more recent synthesis methods, please add one or two recent baselines to support the 'state-of-the-art' claim.

Circularity Check

0 steps flagged · score 2.0 of 10

No material circularity: the teacher-student distillation objective, adaptive weights, and downstream evaluation are self-contained. The one self-citation ([18]) is a supporting reference, not a load-bearing reduction.

full rationale

The central derivation is a distillation loop defined by the paper's own equations: the teacher is conditioned on the fused mask-image features cmix = ci + cm (Eq. 3), the student is conditioned only on the mask, and LAda (Eq. 7) aligns student noise predictions to a stop-gradient teacher using weights wAda computed from mask pixel statistics (Eq. 6). No outcome metric, such as downstream mDice/mIoU, appears in the training objective or in the definition of wAda. The downstream segmentation experiments are evaluated on held-out, non-overlapping test cases (KiTS19: 10 non-overlapping cases; Polyps: five public test datasets not used for generation training), so the reported gains are not fitted inputs. The paper's claim that improved mask-lesion alignment drives segmentation gains is an interpretation supported indirectly in Sec. 4.4 ('there is no consistency between image quality and segmentation performance, indirectly highlighting that our method's superior mask-lesion alignment is key'), and Table 1 shows FID slightly worse than ControlNet. This is an empirical gap or causal-attribution risk, but not a circular reduction: alignment superiority is not defined as the segmentation metric nor fed back into training. The only self-citation appears in Sec. 1, where [18] is cited for the general statement that 'conditional inputs act as implicit regularizers to ensure stable optimization and enhanced image quality.' That citation is not load-bearing, because the method's equations, ablations, and external comparisons stand independently, and [18] is a separate peer-reviewed paper with its own evaluation. Thus no circular step is present; the self-citation is minor and non-load-bearing, yielding a score of 2.

Assumptions & free parameters 7 free parameters · 4 assumptions · 0 invented entities

The method introduces no new latent entity or physical constant. The central result rests on standard diffusion model training, the ControlNet architecture, and conventional hyperparameters. The adaptive weighting is derived directly from masks. The weakest assumptions are domain-level: that downstream segmentation accuracy captures synthetic image utility, and that the teacher's noise prediction is a sound target for the student. No fitted constants are used beyond training hyperparameters.

free parameters (7)

Learning rate = 1e-5
AdamW optimizer hyperparameter, chosen without a sensitivity study (Section 4.2).
Weight decay = 1e-2
AdamW optimizer hyperparameter, fixed during training (Section 4.2).
Training iterations = 3000
Fixed training budget; the convergence claim is based on this budget (Section 4.2).
Classifier-free guidance scale = 9
Sampling hyperparameter, CFG=9 (Section 4.2).
DDIM steps = 50
Deterministic sampling steps (Section 4.2).
Prompt dropout probability = 0.05
Probability of dropping the prompt during training (Section 4.2).
Loss weight for LAda = 1 (implicit)
The three loss terms in Eq. (4) are summed with equal weight; no weighting schedule is reported, which may affect the distillation balance.

assumptions (4)

standard math Standard diffusion model training objective (Eq. 1) produces a valid denoising network.
Foundation of the method, stated in Section 2.
domain assumption Downstream segmentation performance is a valid proxy for synthetic image utility.
The paper evaluates the method primarily through segmentation gains (Tables 2-3).
domain assumption The teacher's image-conditioned noise prediction is a meaningful target for the mask-only student.
This is the core of the distillation mechanism (Eq. 7), assumed without theoretical justification.
domain assumption CLIP-I is a valid semantic similarity metric for medical images.
Used in Table 1 to support quality claims when FID is worse.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Adaptively Distilled ControlNet: Accelerated Training and Superior Sampling for Medical Image Synthesis." pith.science (2026). https://pith.science/paper/N7DTEJ3G

@misc{pith2026250723652,
  author       = {Pith},
  title        = {Pith review of: Adaptively Distilled ControlNet: Accelerated Training and Superior Sampling for Medical Image Synthesis},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/N7DTEJ3G}},
  note         = {Machine review of arXiv:2507.23652}
}

read the original abstract

Medical image annotation is constrained by privacy concerns and labor-intensive labeling, significantly limiting the performance and generalization of segmentation models. While mask-controllable diffusion models excel in synthesis, they struggle with precise lesion-mask alignment. We propose \textbf{Adaptively Distilled ControlNet}, a task-agnostic framework that accelerates training and optimization through dual-model distillation. Specifically, during training, a teacher model, conditioned on mask-image pairs, regularizes a mask-only student model via predicted noise alignment in parameter space, further enhanced by adaptive regularization based on lesion-background ratios. During sampling, only the student model is used, enabling privacy-preserving medical image generation. Comprehensive evaluations on two distinct medical datasets demonstrate state-of-the-art performance: TransUNet improves mDice/mIoU by 2.4%/4.2% on KiTS19, while SANet achieves 2.6%/3.5% gains on Polyps, highlighting its effectiveness and superiority. Code is available at GitHub.

Figures

Figures reproduced from arXiv: 2507.23652 by the authors.

**Figure 1.** (a) Illustration of our method during the training phase. (b) During sampling, only the student model is utilized with arbitrary masks. ϵθ(xt, t) directly predicts the noise, reducing the training objective to: Lsimple = Ext,t,ϵ ∥ϵθ(xt, t) − ϵ∥ 2 2 [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Visualizing the difference between ControlNet and our method in training convergence and data fitting [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Examples of real and synthetic kidney tumor images generated by each method [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (1 more)

**Figure 4.** Figure 4: Examples of real and synthetic polyp images generated by each method. where sg(·) indicates stop-gradient operation. 4 Experiment 4.1 Dataset and Evaluation Metrics We evaluate our method on two publicly available medical datasets: Polyps [12,1] (RGB) and KiTS19 [6] (C…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 28 canonical work pages

[1]

saliency maps from physicians

Bernal, J., Sánchez, F.J., Fernández-Esparrach, G., Gil, D., Rodríguez, C., Vilar- iño, F.: Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. COMPUT MED IMAG GRAP 43, 99–111 (2015)

work page 2015
[2]

Medical Image Analysis (2024)

Chen, J., Mei, J., Li, X., Lu, Y., Yu, Q., Wei, Q., Luo, X., Xie, Y., Adeli, E., Wang, Y., et al.: Transunet: Rethinking the u-net architecture design for medical image segmentation through the lens of transformers. Medical Image Analysis (2024)

work page 2024
[3]

In: NeurIPS (2021)

Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: NeurIPS (2021)

work page 2021
[4]

CAAI AIR2, 9150015 (2021)

Dong, B., Wang, W., Fan, D.P., Li, J., Fu, H., Shao, L.: Polyp-pvt: Polyp segmen- tation with pyramid vision transformers. CAAI AIR2, 9150015 (2021)

work page 2021
[5]

In: MICCAI (2023)

Du, Y., Jiang, Y., Tan, S., Wu, X., Dou, Q., Li, Z., Li, G., Wan, X.: Arsdm: colonoscopy images synthesis with adaptive refinement semantic diffusion models. In: MICCAI (2023)

work page 2023
[6]

arXiv preprint arXiv:1904.00445 (2019)

Heller, N., Sathianathen, N., Kalapara, A., Walczak, E., Moore, K., Kaluzniak, H., Rosenberg, J., Blake, P., Rengel, Z., Oestreich, M., et al.: The kits19 challenge data: 300 kidney tumor cases with clinical context, ct semantic segmentations, and surgical outcomes. arXiv preprint arXiv:1904.00445 (2019)

arXiv 2019
[7]

In: NeurIPS (2017)

Heusel,M.,Ramsauer,H.,Unterthiner,T.,Nessler,B.,Hochreiter,S.:Ganstrained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017)

work page 2017
[8]

In: NeurIPS (2020)

Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)

2020

Show all 36 references

[9]

In: NeurIPS Workshop (2022)

Ho, J., Salimans, T.: Classifier-free diffusion guidance. In: NeurIPS Workshop (2022)

2022
[10]

In: AAAI (2024)

Hu, T., Zhang, J., Yi, R., Du, Y., Chen, X., Liu, L., Wang, Y., Wang, C.: Anoma- lydiffusion: Few-shot anomaly image generation with diffusion model. In: AAAI (2024)

2024
[11]

Nature methods (2021)

Isensee, F., Jaeger, P.F., Kohl, S.A., Petersen, J., Maier-Hein, K.H.: nnu-net: a self-configuring method for deep learning-based biomedical image segmentation. Nature methods (2021)

2021
[12]

In: MMM (2020)

Jha, D., Smedsrud, P.H., Riegler, M.A., Halvorsen, P., De Lange, T., Johansen, D., Johansen, H.D.: Kvasir-seg: A segmented polyp dataset. In: MMM (2020)

2020
[13]

In: ECCV (2024)

Li, M., Yang, T., Kuang, H., Wu, J., Wang, Z., Xiao, X., Chen, C.: Control- net_plus_plus: Improving conditional controls with efficient consistency feedback. In: ECCV (2024)

2024
[14]

Liu,S.,Chen,Z.,Yang,Q.,Yu,W.,Dong,D.,Hu,J.,Yuan,Y.:Polyp-gen:Realistic and diverse polyp image generation for endoscopic dataset expansion (2025)

2025
[15]

In: ICLR (2017) 10 Qiu et al

Loshchilov, I.: Decoupled weight decay regularization. In: ICLR (2017) 10 Qiu et al

2017
[16]

In: CVPR (2023)

Meng, C., Rombach, R., Gao, R., Kingma, D., Ermon, S., Ho, J., Salimans, T.: On distillation of guided diffusion models. In: CVPR (2023)

2023
[17]

In: AAAI (2024)

Mou, C., Wang, X., Xie, L., Wu, Y., Zhang, J., Qi, Z., Shan, Y.: T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. In: AAAI (2024)

2024
[18]

In: CVPR (2025)

Qiu, K., Gao, Z., Zhou, Z., Sun, M., Guo, Y.: Noise-consistent siamese-diffusion for medical image synthesis and segmentation. In: CVPR (2025)

2025
[19]

In: ICASSP (2024)

Qiu, K., Zhou, Z., Guo, Y.: Learn from zoom: Decoupled supervised contrastive learning for wce image classification. In: ICASSP (2024)

2024
[20]

In: CVPR (2022)

Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: CVPR (2022)

2022
[21]

In: CVPR (2023)

Ruiz, N., Li, Y., Jampani, V., Pritch, Y., Rubinstein, M., Aberman, K.: Dream- booth: Fine tuning text-to-image diffusion models for subject-driven generation. In: CVPR (2023)

2023
[22]

IEEE TMI (2025)

Shao, M., Wang, Z., Duan, H., Huang, Y., Zhai, B., Wang, S., Long, Y., Zheng, Y.: Rethinking brain tumor segmentation from the frequency domain perspective. IEEE TMI (2025)

2025
[23]

In: IJCAI Workshop (2024)

Shao, S., Yuan, X., Huang, Z., Qiu, Z., Wang, S., Zhou, K.: Diffuseexpand: Ex- panding dataset for 2d medical image segmentation using diffusion models. In: IJCAI Workshop (2024)

2024
[24]

INT J COMPUT ASS RAD9, 283–293 (2014)

Silva, J., Histace, A., Romain, O., Dray, X., Granado, B.: Toward embedded de- tection of polyps in wce images for early diagnosis of colorectal cancer. INT J COMPUT ASS RAD9, 283–293 (2014)

2014
[25]

In: ICLR (2020)

Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2020)

2020
[26]

In: ICML (2023)

Song, Y., Dhariwal, P., Chen, M., Sutskever, I.: Consistency models. In: ICML (2023)

2023
[27]

In: ICLR (2020)

Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score- based generative modeling through stochastic differential equations. In: ICLR (2020)

2020
[28]

IEEE TMI35(2), 630–644 (2015)

Tajbakhsh, N., Gurudu, S.R., Liang, J.: Automated polyp detection in colonoscopy videos using shape and context information. IEEE TMI35(2), 630–644 (2015)

2015
[29]

PloS one17(5) (2022)

Thambawita, V., Salehi, P., Sheshkal, S.A., Hicks, S.A., Hammer, H.L., Parasa, S., Lange, T.d., Halvorsen, P., Riegler, M.A.: Singan-seg: Synthetic training data generation for medical image segmentation. PloS one17(5) (2022)

2022
[30]

arXiv preprint arXiv:2501.15248 (2025)

Tian, Y., Ucurum, E., Han, X., Young, R., Chatwin, C., Birch, P.: Enhancing fetal plane classification accuracy with data augmentation using diffusion models. arXiv preprint arXiv:2501.15248 (2025)

2025 arXiv
[31]

In: NeurIPS (2017)

Van Den Oord, A., Vinyals, O., et al.: Neural discrete representation learning. In: NeurIPS (2017)

2017
[32]

J HEALTHC ENG2017(1), 4037190 (2017)

Vázquez, D., Bernal, J., Sánchez, F.J., Fernández-Esparrach, G., López, A.M., Romero, A., Drozdzal, M., Courville, A.: A benchmark for endoluminal scene seg- mentation of colonoscopy images. J HEALTHC ENG2017(1), 4037190 (2017)

2017
[33]

In: MICCAI (2021)

Wei, J., Hu, Y., Zhang, R., Li, Z., Zhou, S., Cui, S.: Shallow attention network for polyp segmentation. In: MICCAI (2021)

2021
[34]

In: ICCV (2023)

Wu, W., Zhao, Y., Shou, M.Z., Zhou, H., Shen, C.: Diffumask: Synthesizing images with pixel-level annotations for semantic segmentation using diffusion models. In: ICCV (2023)

2023
[35]

arXiv preprint arXiv:2507.05148 (2025)

Xie, C., Yoshii, Y., Kitahara, I.: Sv-drr: High-fidelity novel view x-ray synthesis using diffusion model. arXiv preprint arXiv:2507.05148 (2025)

2025
[36]

In: ICCV (2023)

Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: ICCV (2023)

2023

Pith tools

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

[1] [1]

saliency maps from physicians

Bernal, J., Sánchez, F.J., Fernández-Esparrach, G., Gil, D., Rodríguez, C., Vilar- iño, F.: Wm-dova maps for accurate polyp highlighting in colonoscopy: Validation vs. saliency maps from physicians. COMPUT MED IMAG GRAP 43, 99–111 (2015)

work page 2015

[2] [2]

Medical Image Analysis (2024)

Chen, J., Mei, J., Li, X., Lu, Y., Yu, Q., Wei, Q., Luo, X., Xie, Y., Adeli, E., Wang, Y., et al.: Transunet: Rethinking the u-net architecture design for medical image segmentation through the lens of transformers. Medical Image Analysis (2024)

work page 2024

[3] [3]

In: NeurIPS (2021)

Dhariwal, P., Nichol, A.: Diffusion models beat gans on image synthesis. In: NeurIPS (2021)

work page 2021

[4] [4]

CAAI AIR2, 9150015 (2021)

Dong, B., Wang, W., Fan, D.P., Li, J., Fu, H., Shao, L.: Polyp-pvt: Polyp segmen- tation with pyramid vision transformers. CAAI AIR2, 9150015 (2021)

work page 2021

[5] [5]

In: MICCAI (2023)

Du, Y., Jiang, Y., Tan, S., Wu, X., Dou, Q., Li, Z., Li, G., Wan, X.: Arsdm: colonoscopy images synthesis with adaptive refinement semantic diffusion models. In: MICCAI (2023)

work page 2023

[6] [6]

arXiv preprint arXiv:1904.00445 (2019)

Heller, N., Sathianathen, N., Kalapara, A., Walczak, E., Moore, K., Kaluzniak, H., Rosenberg, J., Blake, P., Rengel, Z., Oestreich, M., et al.: The kits19 challenge data: 300 kidney tumor cases with clinical context, ct semantic segmentations, and surgical outcomes. arXiv preprint arXiv:1904.00445 (2019)

arXiv 2019

[7] [7]

In: NeurIPS (2017)

Heusel,M.,Ramsauer,H.,Unterthiner,T.,Nessler,B.,Hochreiter,S.:Ganstrained by a two time-scale update rule converge to a local nash equilibrium. In: NeurIPS (2017)

work page 2017

[8] [8]

In: NeurIPS (2020)

Ho, J., Jain, A., Abbeel, P.: Denoising diffusion probabilistic models. In: NeurIPS (2020)

2020