Pith. sign in

REVIEW 5 major objections 6 minor 60 references

Multi-Channel Uncertainty-Weighted Score Matching for Conditional Diffusion in Medical UDA

T0 review · 5 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The paper claims that conditioning a target-domain diffusion model on multiple ranked pseudo-label maps, aggregated via pixel-wise confidence weighting, makes score matching robust to noisy pseudo-labels and improves medical image segmentat

desk verdict A genuinely useful empirical UDA package with a promising uncertainty-weighted diffusion idea, but the promised theory is missing and the training objective as written doesn't match inference; worth refereeing, not worth endorsing yet. read the letter →

arxiv 2509.22476 v2 pith:TLDCS4WY submitted 2025-09-26 cs.CV

Chen Li , Meilong Xu , Xiaoling Hu , Weimin Lyu , Chao Chen This is my paper
classification cs.CV
keywords unsuperviseddomainadaptationconditionaldiffusionmodelmedicalimagesegmentationpseudo-labelnoiseuncertainty-guidedscorematchingBéziercurvestyletransfercross-modalityscore-basedgenerativemodels
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 establish that a conditional diffusion model can be trained in an unlabeled target medical domain using only pseudo-labels from an imperfect source-trained segmenter, provided the score-matching objective aggregates score estimates from several ranked pseudo-label maps (most likely class, second most likely, etc.) weighted pixel-wise by their confidence. The authors argue that in uncertain regions the second-most-likely label is often correct, so committing to Arg-Max alone corrupts generation; their uncertainty-reweighted score is meant to fix that. Combined with a low-degree-of-freedom Bézier-curve intensity adaptation that reduces the appearance gap, the framework generates labeled target-style images that augment training and, in experiments on BraTS, MM-WHS, and abdominal CT/MRI datasets, improve Dice and Hausdorff metrics over several UDA baselines. A sympathetic reader would care because this points to a way of producing additional labeled target-domain samples without manual annotation.

What carries the argument

The uncertainty-reweighted score network ŝθ(x_t, c~, t) = Σ_{k=1}^{|C|} c^(k) ⊙ sθ(x_t, y~(k), t), where ⊙ is the Hadamard product, c^(k) is the k-th ranked softmax confidence map, y~(k) is the k-th ranked pseudo-label map, and |C| is the number of classes. This object carries the argument: it converts one noisy conditional input into many soft conditions and aggregates their score estimates pointwise. A supporting component is the confidence-thresholding rule: where the Arg-Max confidence exceeds δ, the alternatives are replaced by the Arg-Max label with uniform weight 1/|C|, forcing the model to trust high-confidence pixels completely. The Bézier adaptation (a cubic Bézier curve with four

What would settle it

Train the same CDM on the same target images under three conditioning schemes: (a) the uncertainty-reweighted multi-map score, (b) a single Arg-Max map, and (c) a single map drawn per pixel from the softmax distribution. If scheme (c) matches or beats (a) on downstream segmentation Dice, the advantage is not specific to confidence-weighted convex combination. Alternatively, replace c^(k) with uniform weights over the same top-k maps; if uniform weighting performs equally well, the confidence weighting itself is not the operative mechanism.

Watch

Extended reading notes

Core claim

The central claim is stated in Section 3.2: given an imperfect segmentation model, the conditional score can be estimated by evaluating the score network on multiple ranked pseudo-label maps (Arg-Max, Arg-2nd, etc.) and combining the outputs with pixel-wise confidence maps. The reweighted score is then used in a standard denoising score-matching loss, and the authors report that this training, together with the Bézier adaptation, yields target-domain synthetic images that are both realistic and aligned with their conditioning masks, and that augmenting existing UDA pipelines with these images improves segmentation performance. The abstract promises a minimum-MSE convex-combination justificat

Load-bearing premise

The load-bearing premise is that the pixel-wise confidence-weighted sum of score-network outputs, each conditioned on a different ranked pseudo-label, is a valid estimate of the true conditional score under the segmenter-induced surrogate label distribution; the abstract promises a minimum-MSE convex-combination justification, but Section 3.2 of the body contains no theorem, lemma, or proof, so if this estimator does not approximate the true conditional score, the CDM trainin

Editorial extensions

If this is right

  • If the central claim holds, target-domain conditional diffusion training no longer requires clean labels: an imperfect source-trained segmenter suffices, because multiple ranked pseudo-labels are obtained from the same softmax output.
  • The generated target-style images can be added to existing UDA pipelines (ADVENT, GenericSSL, pseudo-labeling) as extra labeled data, which the experiments report improves their Dice and 95% Hausdorff performance.
  • The ablation with k=2 (Arg-Max plus Arg-2nd) outperforming both k=1 and k=3 implies that, for the tested tasks, the second-ranked label carries most of the uncertainty information; higher ranks add compute without benefit.
  • The method is plug-and-play in the sense that the CDM architecture at inference time still takes only a single segmentation mask as its condition, so existing augmentation workflows can adopt it without architectural changes.

Reading between the lines

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

  • Editorial extension: if the reweighted score is truly a minimum-MSE estimator of the conditional score, then the procedure is equivalent to marginalizing over a discrete surrogate label posterior at each pixel; a natural testable extension is to replace ranked Arg-k maps with Monte Carlo samples from the softmax distribution, which should give the same or better robustness.
  • Editorial extension: the absence of a proof means the empirical gains could come from the confidence-thresholding rule or from Bézier adaptation rather than from the convex-combination principle itself; an ablation that keeps thresholding fixed and varies only the weight construction would separate these effects.
  • Editorial extension: the authors themselves note (Section H) that uncertainty-guided training increases GPU memory and compute relative to standard conditional diffusion training; for practical deployment, the open question is whether the reported segmentation gain justifies that cost.
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

5 major / 6 minor

Summary. The paper proposes a two-stage unsupervised domain adaptation (UDA) framework for medical image segmentation. In the first stage, a low-degree-of-freedom Bézier-curve intensity transformation is optimized by feature-space matching (Section 3.1) to reduce the source-target modality gap and to produce a better initial segmentation model. In the second stage, a conditional diffusion model (CDM) is trained on target-domain images paired with pseudo-labels generated by that imperfect segmenter. To mitigate pseudo-label noise, the authors propose an uncertainty-guided score-matching objective (Section 3.2) that aggregates multiple ranked pseudo-label maps (Arg-Max, Arg-2nd, ...) through pixel-wise confidence weights. At augmentation time, the CDM is conditioned on a single segmentation mask to synthesize labeled target-style images, which are then added to the UDA training set. Experiments on BraTS, MM-WHS, and Abdominal Multi-Organ datasets compare the method against GAN-based, diffusion-based, and self-training baselines, reporting improved Dice and 95% Hausdorff distance in most settings. The paper also includes ablations on the confidence threshold, the number of ranked maps, the style-transfer module, and the pseudo-labeling hyperparameters.

Significance. If the central claims are established, the contribution is practically valuable: it offers a generative augmentation route for medical UDA that does not require target annotations, and the uncertainty-weighted CDM training is an intuitive way to handle noisy pseudo-labels. The Bézier adaptation component is simple, low-parameter, and appears to improve both accuracy and calibration (Table 5 and Table 6). The empirical gains over several strong baselines are substantial in a number of settings, and the paper makes a good-faith effort to ablate the key components. However, the paper's main methodological novelty—the uncertainty-weighted score combination—is currently not rigorously formulated or theoretically justified despite an explicit promise in the abstract. The training objective as written contains undefined quantities and notational inconsistencies, and the relationship between the training-time weighted ensemble and the inference-time single-mask condition is not established. These issues are load-bearing for the paper's central claim and must be addressed before the work can be judged sound.

major comments (5)
  1. [Section 3.2, displayed objective] The uncertainty-guided objective is not well-formed. The text defines ŝθ(xt, c~, t) = Σ_k c^(k) · sθ(xt, y~(k), t), but the loss is written with ŝθ(xt, y~(k), t), and the text says “⊙” is the Hadamard product while the formula uses scalar multiplication. More importantly, the target gt(x0, xt, y) = ∇_xt log p_{t|0}(xt | x0, y) contains a condition y that is never defined. If y is the true label, it is unavailable for target-domain training; if y is a pseudo-label, the paper must specify which one (Arg-Max, the full set, or a draw from the confidence-weighted distribution). This ambiguity prevents the objective from being reproducible and makes it impossible to verify that the correct score-matching target is used.
  2. [Abstract and Section 3.2] The abstract promises a theoretical justification that confidence-weighted aggregation follows a minimum-MSE convex-combination principle under the segmenter-induced surrogate label distribution. The body contains no theorem, lemma, proof, or formal statement of this principle. The only support is the verbal assertion that high-confidence predictions contribute sufficient semantic information. This is a load-bearing omission: the entire uncertainty-weighting mechanism rests on the claim that Σ_k c^(k) ⊙ sθ(xt, y~(k), t) is a valid estimator of the conditional score ∇_xt log p_t(xt | Y). The authors should either provide the promised derivation or remove the claim and rephrase the contribution as an empirical heuristic.
  3. [Section 3.2 vs. Section 1 and Fig. 1(b)] There is a train/inference mismatch. Training constrains only the confidence-weighted sum of score-network outputs over all ranked heads. At the application stage, however, the CDM is conditioned on a single segmentation mask (Section 1, Fig. 1(b)), i.e., the score network is evaluated at only the Arg-Max mask. Nothing in the training objective isolates or supervises this individual head, so the exact function used at inference is never directly trained. In low-confidence regions the weighting deliberately down-weights the Arg-Max head, so the model receives the least supervision exactly where the single-mask condition is most uncertain. The paper should address this gap, e.g., by adding a per-head score-matching term or by using the same confidence-weighted ensemble at inference.
  4. [Section 4.2, Table 4 and Table 7] The ablation evidence for the uncertainty-weighting mechanism is suggestive but not conclusive. Table 4 compares “w/o uncertainty” (k=1, Arg-Max only) with Ours-PL (k=2), and Table 7 shows k=2 outperforming k=1. However, this comparison confounds the use of multiple pseudo-label maps with the specific confidence-weighted aggregation; a simple average of heads or an oracle-selected head is not compared. In addition, no standard deviations, error bars, or multiple-seed results are reported for any table, and the k=2 advantage over k=1 is reported for a single target domain (T1CE). Given that the paper's main novelty is precisely this weighting scheme, the authors should provide a more controlled ablation and statistical evidence.
  5. [Section 3.2, confidence-thresholding strategy] The thresholding rule as stated is ambiguous. It says that for k>1, if c^(1)(i,j) > δ, then replace all pseudo-labels at that pixel with the Arg-Max prediction and set c^(k)(i,j) = 1/|C| for all k in C. The text says “all the pseudo-labels at this pixel”, which appears to include k=1, but the sentence begins “for k>1”. If c^(1) is also reset to 1/|C|, then in high-confidence regions the Arg-Max head is down-weighted to the same level as all other heads, which seems contrary to the stated intent of “trusting the Arg-Max prediction completely”. The authors should clarify the exact update for k=1 and justify its effect.
minor comments (6)
  1. [Title and metadata] The arXiv metadata title is “Multi-Channel Uncertainty-Weighted Score Matching for Conditional Diffusion in Medical UDA”, while the full text title is “Bézier Meets Diffusion: Robust Generation Across Domains for Medical Image Segmentation”. The two abstracts also differ, with the metadata abstract promising a theoretical justification that the full-text abstract does not. This discrepancy should be resolved before submission.
  2. [Section 3.1, Eq. (1)] The Bézier curve formula writes B(t) = Σ_{i=1}^n ... P_i; the standard Bernstein form starts at i=0. This is likely a typographical slip, but it makes the control-point indexing inconsistent with the later discussion of P_0...P_3.
  3. [Algorithm 1, line 10] The pseudocode's gradient update line is incomplete: it writes ∇_θ || Σ_k c^(k) sθ − g_t(x0, xt, y) ||²_2 without specifying the arguments of sθ or the definition of y. This should match the corrected main-text objective.
  4. [Tables 1–3] The paper reports a single number per metric per method. Given the variability of medical UDA benchmarks and the small test sets (e.g., MM-WHS with 20 volumes), reporting mean ± std over at least three runs would materially strengthen the comparisons.
  5. [Section H, Limitations] The limitations section only mentions GPU memory and computational cost. Given that the method is a self-training pipeline whose pseudo-labels and confidence maps come from the same initial segmenter, the authors should also discuss the risk of error propagation and the conditions under which the uncertainty-weighted CDM could amplify systematic segmenter errors.
  6. [Reproducibility] The metadata abstract states that code is available at a GitHub URL, but the full text does not include this statement or a link. Please include the code/data availability information consistently in the camera-ready version.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central result is an empirical UDA improvement, and the missing minimum-MSE proof is a grounding gap, not a circular reduction.

full rationale

The paper's central claim is an empirical UDA improvement (Tables 1–3, 8) measured against external baselines; it is not derived from its own conclusion. The pipeline is self-training-like, but the circularity concern is mitigated because the CDM is conditioned, at generation time, on source ground-truth segmentation masks, not on target pseudo-labels: Appendix C states 'we generate 30514 target domain synthetic images using segmentation masks from the source domain.' Thus the synthetic images added to the segmentation training set carry true source labels; target pseudo-labels are used only to train the CDM. The uncertainty-weighted score estimator in Section 3.2 is an architectural/training choice, not a fitted parameter renamed as a prediction. The abstract advertises a 'theoretical justification showing that confidence-weighted aggregation follows a minimum-MSE convex-combination principle,' but Section 3.2 contains no theorem or proof and only asserts that 'predictions with high confidence contribute sufficient semantic information.' This is an omitted-proof / grounding gap, not a circular reduction: no displayed equation makes the claimed improvement equal to its inputs by construction. The displayed objective is indeed ill-formed (it depends on an undefined y and trains only the weighted sum while inference uses a single mask), but that is a rigor issue rather than circularity. The self-citations (Li et al. 2023, 2024; Xu et al. 2025) appear in related-work and metric references and are not load-bearing. No uniqueness theorem or external result is imported to force the method. The score is therefore 0.

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

The central claim depends on several unproved domain assumptions. The most critical: the score combination is assumed valid without the promised derivation. The method also relies on pseudo-label self-training, which is standard but means the final numbers are not independent of the pipeline's own pseudo-labels. No genuinely new physical entities are introduced.

free parameters (4)
  • Confidence threshold delta = 0.8 (ablated over 0.7-0.9)
    Chosen by hand for the confidence-thresholding strategy in Algorithm 1; Table 10 shows sensitivity.
  • Bezier control points P1, P2 (per cluster) = optimized per dataset via Nelder-Mead
    Fitted to unlabeled target images by minimizing feature MSE (Algorithm 2); four scalars per curve.
  • Pseudo-label thresholds tau_l, tau_u and loss weight lambda = 0.5, 0.7, 0.7
    Hand-tuned hyperparameters of the mean-teacher pseudo-labeling framework (Table 10). They affect final segmentation but not the core CDM objective.
  • Number of ranked channels k = 2
    Selected by ablation (Table 7); k=2 best, k=3 worse.
assumptions (4)
  • domain assumption Label distribution is shared across source and target domains
    Invoked to justify conditioning the CDM on source segmentation masks and applying it to target images (Section 3, Fig. 1b).
  • domain assumption Feature-space similarity from a Stable Diffusion autoencoder pretrained on natural images aligns medical modalities structurally
    Bezier control points are optimized by matching features of a natural-image autoencoder (Section 3.1, Algorithm 2, Section G). If false, the style transfer may distort anatomy.
  • domain assumption Softmax confidences of the source-trained segmenter are meaningful uncertainty estimates on the target domain
    The confidence maps c^(k) are used as weights in the score combination (Section 3.2). The paper provides calibration experiments (Table 5) as indirect support.
  • ad hoc to paper The confidence-weighted sum of per-channel score estimates approximates the true conditional score
    The central assumption of the uncertainty-guided score matching; the abstract promises a proof that is absent from the body.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-Channel Uncertainty-Weighted Score Matching for Conditional Diffusion in Medical UDA." pith.science (2026). https://pith.science/paper/TLDCS4WY

@misc{pith2026250922476,
  author       = {Pith},
  title        = {Pith review of: Multi-Channel Uncertainty-Weighted Score Matching for Conditional Diffusion in Medical UDA},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TLDCS4WY}},
  note         = {Machine review of arXiv:2509.22476}
}
read the original abstract

Robust medical image segmentation across modalities remains challenging due to severe domain shifts and the lack of target-domain labels. While diffusion models have been explored for cross-domain generation and augmentation, target-domain conditional diffusion training typically relies on highly noisy pseudo masks; naively conditioning on a single Arg-Max pseudo-label can corrupt diffusion training and downstream segmentation. We propose UPDiff-UDA, a unified UDA framework whose core is an uncertainty-guided training objective for target-domain conditional diffusion. Given an imperfect source-trained segmenter, we use its per-pixel softmax distribution to form ranked pseudo-label maps (Arg-Max, Arg-2nd, Arg-3rd, ...). Each map yields a conditional score estimate, and we aggregate them via pixel-wise confidence weighting to obtain an uncertainty-reweighted score for score matching, improving robustness to pseudo-label noise while leveraging alternative plausible labels in uncertain regions. We further provide a theoretical justification showing that confidence-weighted aggregation follows a minimum-MSE convex-combination principle under the segmenter-induced surrogate label distribution. To improve pseudo-condition quality, we also introduce a feature-guided, low-degree-of-freedom B\'ezier curve adaptation to reduce appearance gaps. Experiments on multiple public datasets and modality shifts show that UPDiff-UDA generates high-fidelity labeled target-style samples for augmentation and consistently outperforms strong UDA baselines. The code for this project is available at: https://github.com/superlc1995/Multi-Channel-Uncertainty-Diffusion-UDA

Figures

Figures reproduced from arXiv: 2509.22476 by the authors.

Figure 1
Figure 1. The comparison between (a) GAN-based style transfer methods and (b) our CDM-based [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The overview of our contributions: Bézier adaptation and uncertainty guided CDM. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The overview of Bézier-curve-based style transfer. (a) illustrates the impact of varying [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: An overview of the proposed uncertainty-guided CDM training framework. [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Segmentation results of Ours-PL and other baselines. From top to bottom: (1) BraTS [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Qualitative results for generative models. We also show the segmentation results of the [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: Segmentation results of ours (Ours-PL) and other baselines. [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

60 extracted references · 1 linked inside Pith

  1. [1]

    Segdiff: Image segmentation with diffusion probabilistic models

    Tomer Amit, Tal Shaharbany, Eliya Nachmani, and Lior Wolf. Segdiff: Image segmentation with diffusion probabilistic models. arXiv preprint arXiv:2112.00390, 2021

    arXiv 2021

  2. [2]

    Wasserstein generative adversarial networks

    Martin Arjovsky, Soumith Chintala, and L \'e on Bottou. Wasserstein generative adversarial networks. In ICML, 2017

    2017

  3. [3]

    Verification of forecasts expressed in terms of probability

    Glenn W Brier. Verification of forecasts expressed in terms of probability. Monthly weather review, 1950

    1950

  4. [4]

    u ggemann, Christos Sakaridis, Tim Br \

    David Br \"u ggemann, Christos Sakaridis, Tim Br \"o dermann, and Luc Van Gool. Contrastive model adaptation for cross-condition robustness in semantic segmentation. In ICCV, 2023

    2023

  5. [5]

    Synergistic image and feature adaptation: Towards cross-modality domain adaptation for medical image segmentation

    Cheng Chen, Qi Dou, Hao Chen, Jing Qin, and Pheng-Ann Heng. Synergistic image and feature adaptation: Towards cross-modality domain adaptation for medical image segmentation. In AAAI, 2019

    2019

  6. [6]

    Unsupervised bidirectional cross-modality adaptation via deeply synergistic image and feature alignment for medical image segmentation

    Cheng Chen, Qi Dou, Hao Chen, Jing Qin, and Pheng Ann Heng. Unsupervised bidirectional cross-modality adaptation via deeply synergistic image and feature alignment for medical image segmentation. TMI, 2020

    2020

  7. [7]

    A generalist framework for panoptic segmentation of images and videos

    Ting Chen, Lala Li, Saurabh Saxena, Geoffrey Hinton, and David J Fleet. A generalist framework for panoptic segmentation of images and videos. In ICCV, 2023

    2023

  8. [8]

    Exact histogram specification

    Dinu Coltuc, Philippe Bolon, and J-M Chassery. Exact histogram specification. TIP, 2006

    2006

Show all 60 references
  1. [9]

    Diffusion models beat gans on image synthesis

    Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021

    2021

  2. [10]

    Scribble-based domain adaptation via co-segmentation

    Reuben Dorent, Samuel Joutard, Jonathan Shapey, Sotirios Bisdas, Neil Kitchen, Robert Bradford, Shakeel Saeed, Marc Modat, S \'e bastien Ourselin, and Tom Vercauteren. Scribble-based domain adaptation via co-segmentation. In MICCAI, 2020

    2020

  3. [11]

    Curves and surfaces for CAGD: a practical guide

    Gerald E Farin. Curves and surfaces for CAGD: a practical guide. Morgan Kaufmann, 2002

    2002

  4. [12]

    Digital image processing

    Rafael C Gonzales and Paul Wintz. Digital image processing. Addison-Wesley Longman Publishing Co., Inc., 1987

    1987

  5. [13]

    Generative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014

    2014

  6. [14]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020

    2020

  7. [15]

    Cycada: Cycle-consistent adversarial domain adaptation

    Judy Hoffman, Eric Tzeng, Taesung Park, Jun-Yan Zhu, Phillip Isola, Kate Saenko, Alexei Efros, and Trevor Darrell. Cycada: Cycle-consistent adversarial domain adaptation. In ICML, 2018

    2018

  8. [16]

    Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation

    Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In CVPR, 2022

    2022

  9. [17]

    Synseg-net: Synthetic segmentation without target modality ground truth

    Yuankai Huo, Zhoubing Xu, Hyeonsoo Moon, Shunxing Bao, Albert Assad, Tamara K Moyo, Michael R Savona, Richard G Abramson, and Bennett A Landman. Synseg-net: Synthetic segmentation without target modality ground truth. TMI, 2018

    2018

  10. [18]

    Ddp: Diffusion model for dense visual prediction

    Yuanfeng Ji, Zhe Chen, Enze Xie, Lanqing Hong, Xihui Liu, Zhaoqiang Liu, Tong Lu, Zhenguo Li, and Ping Luo. Ddp: Diffusion model for dense visual prediction. In ICCV, 2023

    2023

  11. [19]

    Psigan: Joint probabilistic segmentation and image distribution matching for unpaired cross-modality adaptation-based mri segmentation

    Jue Jiang, Yu-Chi Hu, Neelam Tyagi, Andreas Rimner, Nancy Lee, Joseph O Deasy, Sean Berry, and Harini Veeraraghavan. Psigan: Joint probabilistic segmentation and image distribution matching for unpaired cross-modality adaptation-based mri segmentation. TMI, 2020

    2020

  12. [20]

    A style-based generator architecture for generative adversarial networks

    Tero Karras, Samuli Laine, and Timo Aila. A style-based generator architecture for generative adversarial networks. In CVPR, 2019

    2019

  13. [21]

    Chaos challenge-combined (ct-mr) healthy abdominal organ segmentation

    A Emre Kavur, N Sinem Gezer, Mustafa Bar s , Sinem Aslan, Pierre-Henri Conze, Vladimir Groza, Duc Duy Pham, Soumick Chatterjee, Philipp Ernst, Sava s \"O zkan, et al. Chaos challenge-combined (ct-mr) healthy abdominal organ segmentation. MedIA, 2021

    2021

  14. [22]

    The brain tumor segmentation (brats) challenge 2023: focus on pediatrics (cbtn-connect-dipgr-asnr-miccai brats-peds)

    Anahita Fathi Kazerooni, Nastaran Khalili, Xinyang Liu, Debanjan Haldar, Zhifan Jiang, Syed Muhammed Anwar, Jake Albrecht, Maruf Adewole, Udunna Anazodo, Hannah Anderson, et al. The brain tumor segmentation (brats) challenge 2023: focus on pediatrics (cbtn-connect-dipgr-asnr-m...

    2023

  15. [23]

    Learning texture invariant representation for domain adaptation of semantic segmentation

    Myeongjin Kim and Hyeran Byun. Learning texture invariant representation for domain adaptation of semantic segmentation. In CVPR, 2020

    2020

  16. [24]

    Miccai multi-atlas labeling beyond the cranial vault--workshop and challenge

    Bennett Landman, Zhoubing Xu, J Igelsias, Martin Styner, Thomas Langerak, and Arno Klein. Miccai multi-atlas labeling beyond the cranial vault--workshop and challenge. In Proc. MICCAI Multi-Atlas Labeling Beyond Cranial Vault—Workshop Challenge, 2015

    2015

  17. [25]

    Confidence estimation using unlabeled data

    Chen Li, Xiaoling Hu, and Chao Chen. Confidence estimation using unlabeled data. In ICLR, 2023

    2023

  18. [26]

    Spatial diffusion for cell layout generation

    Chen Li, Xiaoling Hu, Shahira Abousamra, Meilong Xu, and Chao Chen. Spatial diffusion for cell layout generation. In MICCAI, 2024

    2024

  19. [27]

    Bidirectional learning for domain adaptation of semantic segmentation

    Yunsheng Li, Lu Yuan, and Nuno Vasconcelos. Bidirectional learning for domain adaptation of semantic segmentation. In CVPR, 2019

    2019

  20. [28]

    Susan: segment unannotated image structure using adversarial network

    Fang Liu. Susan: segment unannotated image structure using adversarial network. Magnetic resonance in medicine, 2019

    2019

  21. [29]

    Unsupervised image-to-image translation networks

    Ming-Yu Liu, Thomas Breuel, and Jan Kautz. Unsupervised image-to-image translation networks. In NeurIPS, 2017

    2017

  22. [30]

    Optimized bézier curve based intensity mapping scheme for low light image enhancement

    R Magudeeswaran, A K Bhandari, and B Subramani. Optimized bézier curve based intensity mapping scheme for low light image enhancement. Multimedia Tools and Applications, 2021

    2021

  23. [31]

    Least squares generative adversarial networks

    Xudong Mao, Qing Li, Haoran Xie, Raymond YK Lau, Zhen Wang, and Stephen Paul Smolley. Least squares generative adversarial networks. In ICCV, 2017

    2017

  24. [32]

    Confidence-aware learning for deep neural networks

    Jooyoung Moon, Jihyo Kim, Younghak Shin, and Sangheum Hwang. Confidence-aware learning for deep neural networks. In ICML, 2020

    2020

  25. [33]

    Obtaining well calibrated probabilities using bayesian binning

    Mahdi Pakdaman Naeini, Gregory Cooper, and Milos Hauskrecht. Obtaining well calibrated probabilities using bayesian binning. In AAAI, 2015

    2015

  26. [34]

    Improved denoising diffusion probabilistic models

    Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021

    2021

  27. [35]

    Diffusion-based image translation with label guidance for domain adaptive semantic segmentation

    Duo Peng, Ping Hu, Qiuhong Ke, and Jun Liu. Diffusion-based image translation with label guidance for domain adaptive semantic segmentation. In ICCV, 2023

    2023

  28. [36]

    Loopda: Constructing self-loops to adapt nighttime semantic segmentation

    Fengyi Shen, Zador Pataki, Akhil Gurram, Ziyuan Liu, He Wang, and Alois Knoll. Loopda: Constructing self-loops to adapt nighttime semantic segmentation. In WACV, 2023

    2023

  29. [37]

    Denoising diffusion implicit models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021

    2021

  30. [38]

    Learning to adapt structured output space for semantic segmentation

    Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kihyuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker. Learning to adapt structured output space for semantic segmentation. In CVPR, 2018

    2018

  31. [39]

    Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation

    Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick P \'e rez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In CVPR, 2019

    2019

  32. [40]

    Towards generic semi-supervised framework for volumetric medical image segmentation

    Haonan Wang and Xiaomeng Li. Towards generic semi-supervised framework for volumetric medical image segmentation. In NeurIPS, 2023

    2023

  33. [41]

    Cyclesgan: A cycle-consistent and semantics-preserving generative adversarial network for unpaired mr-to-ct image synthesis

    Runze Wang, Alexander F Heimann, Moritz Tannast, and Guoyan Zheng. Cyclesgan: A cycle-consistent and semantics-preserving generative adversarial network for unpaired mr-to-ct image synthesis. Computerized Medical Imaging and Graphics, 2024

    2024

  34. [42]

    Diffusion models for implicit image segmentation ensembles

    Julia Wolleb, Robin Sandk \"u hler, Florentin Bieder, Philippe Valmaggia, and Philippe C Cattin. Diffusion models for implicit image segmentation ensembles. In MIDL, 2022

    2022

  35. [43]

    Fpl+: Filtered pseudo label-based unsupervised cross-modality adaptation for 3d medical image segmentation

    Jianghao Wu, Dong Guo, Guotai Wang, Qiang Yue, Huijun Yu, Kang Li, and Shaoting Zhang. Fpl+: Filtered pseudo label-based unsupervised cross-modality adaptation for 3d medical image segmentation. TMI, 2024 a

    2024

  36. [44]

    Medsegdiff: Medical image segmentation with diffusion probabilistic model

    Junde Wu, Rao Fu, Huihui Fang, Yu Zhang, Yehui Yang, Haoyi Xiong, Huiying Liu, and Yanwu Xu. Medsegdiff: Medical image segmentation with diffusion probabilistic model. In MIDL, 2024 b

    2024

  37. [45]

    A one-stage domain adaptation network with image alignment for unsupervised nighttime semantic segmentation

    Xinyi Wu, Zhenyao Wu, Lili Ju, and Song Wang. A one-stage domain adaptation network with image alignment for unsupervised nighttime semantic segmentation. PAMI, 2021

    2021

  38. [46]

    Sepico: Semantic-guided pixel contrast for domain adaptive semantic segmentation

    Binhui Xie, Shuang Li, Mingjia Li, Chi Harold Liu, Gao Huang, and Guoren Wang. Sepico: Semantic-guided pixel contrast for domain adaptive semantic segmentation. PAMI, 2023

    2023

  39. [47]

    Topocellgen: Generating histopathology cell topology with a diffusion model

    Meilong Xu, Saumya Gupta, Xiaoling Hu, Chen Li, Shahira Abousamra, Dimitris Samaras, Prateek Prasanna, and Chao Chen. Topocellgen: Generating histopathology cell topology with a diffusion model. In CVPR, 2025

    2025

  40. [48]

    Fda: Fourier domain adaptation for semantic segmentation

    Yanchao Yang and Stefano Soatto. Fda: Fourier domain adaptation for semantic segmentation. In CVPR, 2020

    2020

  41. [49]

    Prototypical pseudo label denoising and target structure learning for domain adaptive semantic segmentation

    Pan Zhang, Bo Zhang, Ting Zhang, Dong Chen, Yong Wang, and Fang Wen. Prototypical pseudo label denoising and target structure learning for domain adaptive semantic segmentation. In CVPR, 2021

    2021

  42. [50]

    Zero-reference deep curve estimation for low-light image enhancement

    Y Zhang, J Zhang, X Guo, and et al. Zero-reference deep curve estimation for low-light image enhancement. Sensors, 2023

    2023

  43. [51]

    Translating and segmenting multimodal medical volumes with cycle-and shape-consistency generative adversarial network

    Zizhao Zhang, Lin Yang, and Yefeng Zheng. Translating and segmenting multimodal medical volumes with cycle-and shape-consistency generative adversarial network. In CVPR, 2018

    2018

  44. [52]

    Unsupervised domain adaptation for semantic segmentation with pseudo label self-refinement

    Xingchen Zhao, Niluthpol Chowdhury Mithun, Abhinav Rajvanshi, Han-Pang Chiu, and Supun Samarasekera. Unsupervised domain adaptation for semantic segmentation with pseudo label self-refinement. In WACV, 2024

    2024

  45. [53]

    Generalizable cross-modality medical image segmentation via style augmentation and dual normalization

    Ziqi Zhou, Lei Qi, Xin Yang, Dong Ni, and Yinghuan Shi. Generalizable cross-modality medical image segmentation via style augmentation and dual normalization. In CVPR, 2022

    2022

  46. [54]

    Models genesis: Generic autodidactic models for 3d medical image analysis

    Zongwei Zhou, Vatsal Sodha, Md Mahfuzur Rahman Siddiquee, Ruibin Feng, Nima Tajbakhsh, Michael B Gotway, and Jianming Liang. Models genesis: Generic autodidactic models for 3d medical image analysis. In MICCAI, 2019

    2019

  47. [55]

    Unpaired image-to-image translation using cycle-consistent adversarial networks

    Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In ICCV, 2017

    2017

  48. [56]

    Multi-scale patch and multi-modality atlases for whole heart segmentation of mri

    Xiahai Zhuang and Juan Shen. Multi-scale patch and multi-modality atlases for whole heart segmentation of mri. MedIA, 2016

    2016

  49. [57]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

  50. [58]

    @esa (Ref

    \@ifxundefined[1] #1\@undefined \@firstoftwo \@secondoftwo \@ifnum[1] #1 \@firstoftwo \@secondoftwo \@ifx[1] #1 \@firstoftwo \@secondoftwo [2] @ #1 \@temptokena #2 #1 @ \@temptokena \@ifclassloaded agu2001 natbib The agu2001 class already includes natbib coding, so you should ...

  51. [59]

    \@lbibitem[] @bibitem@first@sw\@secondoftwo \@lbibitem[#1]#2 \@extra@b@citeb \@ifundefined br@#2\@extra@b@citeb \@namedef br@#2 \@nameuse br@#2\@extra@b@citeb \@ifundefined b@#2\@extra@b@citeb @num @parse #2 @tmp #1 NAT@b@open@#2 NAT@b@shut@#2 \@ifnum @merge>\@ne @bibitem@firs...

  52. [60]

    E Z "E U

    @open @close @open @close and [1] URL: #1 \@ifundefined chapter * \@mkboth \@ifxundefined @sectionbib * \@mkboth * \@mkboth\@gobbletwo \@ifclassloaded amsart * \@ifclassloaded amsbook * \@ifxundefined @heading @heading NAT@ctr thebibliography [1] @ \@biblabel @NAT@ctr \@bibset...

    1955

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.