IntraStyler: Intra-Domain Style Synthesis for Cross-Modality MRI Domain Adaptation

Benoit M. Dawant; Daniel Moyer; Dewei Hu; Han Liu; Hao Li; Ipek Oguz; Yubo Fan; Zhoubing Xu

REVIEW 4 major objections 5 minor 24 references

IntraStyler: Intra-Domain Style Synthesis for Cross-Modality MRI Domain Adaptation

T0 review · 4 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read IntraStyler claims that a single unpaired translation network can synthesize target-MRI styles from any exemplar image, without pre-specified sub-domains, improving downstream segmentation robustness.

desk verdict Solid incremental extension of the authors' CrossMoDA winner with a genuinely new contrastive style encoder, but the 'no prior knowledge' claim outruns the hand-crafted perturbation prior and the evaluation lacks exact numbers. read the letter →

arxiv 2601.00212 v2 pith:IDSAVTTY submitted 2026-01-01 cs.CV

Han Liu , Yubo Fan , Hao Li , Dewei Hu , Daniel Moyer , Zhoubing Xu , Benoit M. Dawant , Ipek Oguz This is my paper

classification cs.CV

keywords domainadaptationimage-to-imagetranslationcontrastivelearningstyledisentanglementintra-domainvariabilityMRIsegmentationvestibularschwannomasynthesis

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper is trying to establish that intra-domain variability—not just cross-domain shift—is a major bottleneck for cross-modality medical image adaptation, and that it can be handled without hand-defined sub-domains. IntraStyler extracts a style-only vector from any T2 MRI exemplar using a contrastively trained 3D encoder, then injects that vector into the generator via dynamic instance normalization so the translated T1 image takes on the exemplar's appearance while keeping the input anatomy. Because exemplars are drawn freely from the unlabeled target pool, a single network can produce many target styles. The authors report that this diverse synthetic data yields more robust segmentation of vestibular schwannoma and cochlea than prior methods that require scanner or sub-dataset labels.

What carries the argument

The central object is the contrastively trained style encoder E_S producing K=256-dimensional unit-normalized style vectors, together with the dynamic instance normalization (DIN) layers that inject them into a contrastive-unpaired-translation generator. The encoder is trained with an (N+1)-way contrastive classification: query and positive are two patches from the same exemplar; negatives are positive patches under five kinds of intensity perturbation (contrast, smoothing, noise, bias field, or mixture). DIN replaces the last two instance-normalization layers in the decoder: it channel-normalizes the feature, uses the style vector through a 1x1x1 convolutional mapping to produce per-exempla

What would settle it

Hold out T2 images from a scanner or protocol whose appearance is dominated by an artifact the perturbation set cannot mimic (for example, strong Gibbs ringing or motion ghosting). If the style encoder neither places those images near their visual peers in embedding space nor preserves segmentation Dice when they are used as exemplars, the core claim that styles are captured without prior knowledge is falsified.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that 'style' can be learned as a disentangled, unit-norm vector purely by contrastive learning against intensity perturbations. A patch from an exemplar image serves as the query; another patch from the same image is the positive; negatives are copies of that patch with random contrast, smoothing, noise, or bias-field changes. Matching the query style forces the encoder to ignore anatomy and retain appearance. The resulting style vector, fed through dynamic instance normalization, makes the translation output reproduce the exemplar's style; a style consistency loss reinforces this. The paper's experiments show the learned embedding clusters unlabele

Load-bearing premise

The load-bearing premise is that the hand-crafted intensity perturbation set (contrast, smoothing, noise, bias field, and their mixtures) spans the real variation among target-domain MRI images; if it does not, the 'style' the encoder learns will not cover the actual intra-domain diversity, and the downstream gains will not transfer to unseen scanners.

Editorial extensions

If this is right

A single trained network can synthesize as many target-domain styles as there are exemplar images, eliminating per-sub-domain networks or sub-domain labels.
Synthetic data diversity directly affects downstream segmentation: the paper's results show larger Dice improvements and fewer failure cases on intra/extra-meatal VS and cochlea compared with non-diverse or sub-domain-conditioned synthesis.
Because style and anatomy are separated, users can choose an exemplar to control appearance while the input image controls structure—useful for test-time augmentation or targeted harmonization.
SLERP interpolation yields intermediate styles between two exemplars, enabling augmentation along style continua rather than only at observed exemplars.
The learned style embedding can group unlabeled images by visual style, which the paper suggests could support unsupervised scanner or site classification.

Reading between the lines

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

The method inherits its definition of style from the hand-picked perturbation set; a straightforward stress test is to add scanner-specific artifacts not in that set (e.g., Gibbs ringing, motion, wrap-around) and see whether the encoder and downstream segmentation still transfer.
Because the style vector is globally pooled, local appearance control (e.g., tumor texture) is out of reach; a patch-wise or hierarchical style representation would be a natural extension and would test whether local style is separable.
The same exemplar-guided recipe should transfer to other unpaired translation tasks—histology stain variation, CT-to-MRI, multi-contrast MRI—where target style can be represented by exemplars; if it fails there, the dependence on modality-specific perturbation choices is the likely cause.
The style embedding could double as a harmonization tool: selecting an exemplar from a desired scanner cluster and translating source images into that cluster's style may produce more controllable cross-site harmonization than current methods.

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 worth a serious look. The new piece is a contrastive style encoder trained with intensity-perturbed negatives, coupled with exemplar-guided DIN injection into a CUT-based generator. That combination is not in the cited subdomain-based ILDA literature, and it does remove the need to pre-specify subdomains. The downstream segmentation comparison is fair: same nnU-Net framework, same leaderboard, and the boxplots show a robustness advantage, especially for VS and boundary ASSD. That is real evidence, and it comes from an external benchmark, so the main claim is not circular in the way the style-embedding evaluation is.

The soft spots are real but mostly addressable. The biggest one is the "without any prior knowledge" framing. The style space is defined by a hand-picked set of intensity perturbations: contrast, smoothing, noise, bias field, mixtures. That is a prior, and the paper never checks whether that perturbation set spans the actual T2 scanner/site variation in the target data. The authors do flag this as a limitation in the discussion, which earns credit, but the abstract's claim is still overstated. Second, the style-consistency loss uses the same encoder as both guide and evaluator; that is a common trick and not fatal, but it does mean the style-space quality is partly self-confirming. Third, there is no ablation that isolates the exemplar/style mechanism from simply applying the same perturbation set as data augmentation on a single-style translation. The NoDiverse baseline is a single style, not that ablation, so the causal role of the learned style space is not fully pinned down. Fourth, the paper gives no exact numbers, no error bars, no significance tests; for a methods paper that wants to claim robustness, that is a real omission. Minor: the abstract and full text list different code URLs.

All of these are fixable. The method itself is coherent, the extension over prior work is clear, and the external evaluation is a genuine strength. The paper deserves peer review; it should be conditionally accepted with requests for exact numbers, an augmentation-style ablation, and a more honest framing of the perturbation prior. I would cite the contrastive style-encoder idea in the next year, and I might bring it to reading group if a revision lands.

Referee Report

4 major / 5 minor

Summary. The paper proposes IntraStyler, an exemplar-based unpaired image translation method for cross-modality MRI domain adaptation. It extends CUT by adding a 3D style encoder trained with a contrastive objective in which negatives are generated by hand-crafted intensity perturbations; the resulting style vector is injected into a CUT-based generator through dynamic instance normalization, with an additional style consistency loss. SLERP interpolation between style vectors is also introduced. Experiments on CrossMoDA 2023 show clustering of target T2 styles and downstream nnU-Net segmentation comparisons on the post-challenge leaderboard, aimed at demonstrating that diverse synthetic T2 images improve VS and cochlea segmentation compared to prior ILDA methods that rely on pre-specified sub-domains.

Significance. If the claims hold, the paper addresses an important and under-explored issue in image-level domain adaptation: intra-domain variability of the target domain. The exemplar-based, sub-domain-free formulation is attractive, and the external CrossMoDA leaderboard evaluation with a fixed nnU-Net framework is a real strength, as is the plan to release code. The style interpolation via SLERP is a useful practical addition. However, the evidence currently presented is incomplete and partly circular: the style prior is the hand-crafted perturbation set, the style consistency loss uses the same encoder for guidance and evaluation, and the segmentation results are reported only as boxplots without exact values or significance tests. The contribution is therefore promising but not yet fully supported.

major comments (4)

[Section 2.2, Section 2.5, Eq. (1)] The claim that IntraStyler 'can capture diverse intra-domain styles without any prior knowledge' is contradicted by the construction. The only supervision for the style encoder is the hand-crafted set of intensity perturbations (contrast adjustment, Gaussian smoothing, Gaussian noise, bias field, mixtures) used to form negatives in Eq. (1). The learned style space is therefore, by construction, sensitive mainly to transformations in that set; real scanner/site variability outside this set cannot be discovered. The authors themselves acknowledge in Section 4 that perturbation selection may be task-specific. This is not fatal, but the 'no prior knowledge' wording should be removed or qualified, and an experiment should check whether the learned style distribution actually covers real target variation, e.g., by comparing style embeddings of synthetic and real T2 images or by evaluating per-
[Section 3.2, Fig. 5] The downstream segmentation comparison is the central evidence for the robustness claim, yet no exact Dice or ASSD values are reported in the text or in a table. The paper relies entirely on boxplots, 'smaller whiskers', and a hard-to-read red failure-case box, with no indication of the number of cases underlying each failure count and no statistical test. Since the claims are comparative ('more robust than MultiNets/Unified'), exact mean/standard deviation and paired tests across the validation cases are required. The CrossMoDA leaderboard returns per-case metrics; these should be used to provide a table and paired comparisons (e.g., Wilcoxon signed-rank or permutation test).
[Section 3.2 (NoDiverse vs IntraStyler)] The experimental design does not isolate the contribution of the learned style encoder/exemplar mechanism from the intensity perturbation set that defines the style space. A natural control is a baseline that translates source images to a single T2 style and then applies the same intensity perturbations as on-the-fly augmentation while training the segmentation model. If such a control matches IntraStyler's segmentation gains, the benefit is generic augmentation rather than style-space discovery. A second useful control is to inject random style vectors in place of the learned exemplar vectors while keeping the same exemplar-sampling protocol; this separates stochasticity from the learned representation. These ablations are needed before the improvement can be attributed to the proposed style encoder.
[Section 3.1, Eq. (2)] The synthesis evaluation is partially circular. The style embedding is validated by K-means clusters computed in the same embedding, and the style consistency loss in Eq. (2) uses the same encoder E_S both as generator guide and as evaluator. This means the encoder is being used to validate a property of its own embedding. An independent check is needed: e.g., a reader study rating whether the output styles match the exemplar styles, an image-space distribution distance between synthetic outputs and real T2 images (e.g., FID), or downstream segmentation stratified by target style clusters. Without such an external check, the claim of 'style-only' disentanglement is not established.

minor comments (5)

[Abstract / Section 1] The two code URLs differ: the abstract gives github.com/MedICL-VU/IntraStyler, while the footnote in Section 1 gives github.com/han-liu/IntraStyler. Please unify.
[Implementation Details] The text repeatedly defers to 'the Appendix' for preprocessing, network architecture, and hyperparameters, but the version under review contains no appendix. Please ensure the appendix is included in the submission.
[Section 2.4, Eq. (3)] SLERP divides by sin(θ), which is numerically unstable when θ is close to 0 or π. The implementation should use a fallback to linear interpolation in those cases.
[Fig. 5] The figure axes, labels, and failure-case annotation are hard to read in the typeset version. Please enlarge the figure and explicitly report the number of scans with zero Dice per structure in the red boxes.
[Section 3.2] The NoDiverse baseline is not fully specified. Please state which generator/loss it uses (e.g., CUT without style conditioning) and how its single output style is chosen.

Circularity Check

1 steps flagged · score 3.0 of 10

Mild self-referential style-space loop; downstream segmentation benchmark keeps the main claim independent.

self definitional [Section 2.3 Eq. (2); Section 3.1 Fig. 4]
"To further encourage this consistency, we reuse the style encoder ES to obtain the style vector of the generated image and introduce a style consistency loss: Lcon = −sim(ES(y), ES(G(x))) (2) ... Since we do not have the ground truth of style similarity, we use K-means to find the clusters with the most similar styles. ... the samples from the same region have consistent styles while having different anatomies, demonstrating the effectiveness of this well-disentangled style latent space."

The style encoder ES is both the training signal for what counts as a style and the evaluator of whether the generated image matches the exemplar. Lcon encourages ES(G(x)) ≈ ES(y), but ES is trained jointly with G (and with Lstyle, Eq. 1) to define style via hand-crafted perturbations. Agreement in this learned space is therefore not independent evidence that G(x) matches the real target-style distribution if ES is insensitive to unmodeled variations; it only shows agreement with the encoder's own perturbation-defined notion of style. The clustering evaluation in Sec. 3.1 also clusters in this same embedding and then uses the clusters as evidence of disentanglement, so the qualitative validation cannot independently confirm that the extracted 'styles' correspond to true target sub-populati

full rationale

The main downstream claim is grounded in an external benchmark: segmentation performance is obtained on the CrossMoDA 2023 validation leaderboard using nnU-Net trained on synthetic T2 images, so the central result is not reducible to the paper's own fitted values. The self-citations to the authors' prior 1st-place CrossMoDA solution [15] are contextual and not load-bearing: no uniqueness theorem or unverified prior result is imported to force the method. The identified mild circularity is internal to the style space: the style consistency loss (Eq. 2) uses the same encoder that defines style, and the qualitative evaluation (Fig. 4) clusters in that same embedding to demonstrate disentanglement. Additionally, the paper's 'without any prior knowledge' claim is weakened by its own admitted hand-crafted perturbation set and by the Section 4 limitation that 'the selection of intensity perturbation functions may be task-specific'; however, this is an external-validity concern about coverage of real T2 variability, not a self-referential derivation. Overall, no step reduces a prediction to an input by construction, so the circularity score is low.

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

The method introduces no new physical entities. The central load-bearing choices are the hand-crafted intensity perturbation set, the use of global channel statistics to carry style, and the contrastive definition of style as same-image-patch similarity under perturbation. These are explicit and partly acknowledged in the paper's limitations.

free parameters (6)

style vector dimension K = 256
Set empirically (Section 2.5).
contrastive temperature tau = 0.01
Set empirically in Eq. (1).
style consistency weight lambda_con = 5
Set empirically (Section 2.5).
style contrastive weight lambda_style = 5
Set empirically (Section 2.5).
number of negatives N = 8
Set empirically (Section 2.5).
intensity perturbation set = contrast adjustment, Gaussian smooth, Gaussian noise, bias field, mixture
Hand-crafted style prior; defines what the encoder treats as style (Section 2.2).

assumptions (4)

domain assumption Hand-crafted intensity perturbations are a sufficient proxy for real scanner/site variability in the target domain.
The contrastive loss defines style as whatever changes under these perturbations; if true T2 variations are orthogonal, the encoder misses them (Section 2.2; Section 4 limitation).
domain assumption Global channel statistics, modulated by dynamic instance normalization, are sufficient to control the styles that matter for downstream segmentation.
Style is injected only through the last two DIN layers; local textures (e.g., tumor internals) are outside the model's control (Section 2.3; Section 4).
ad hoc to paper Different 3D patches from the same exemplar share style while differing in anatomy.
This is built into the contrastive setup: query and positive are patches from the same image, so the encoder is trained to ignore anatomy (Section 2.2).
domain assumption Unpaired ceT1-to-T2 translation preserves label-relevant anatomy.
The DA pipeline assumes that labels on source ceT1 images transfer to synthetic T2 images, which is the standard ILDA premise (Section 2.1).

how reviews work

0 comments

Cite this review

Pith. "Pith review of IntraStyler: Intra-Domain Style Synthesis for Cross-Modality MRI Domain Adaptation." pith.science (2026). https://pith.science/paper/IDSAVTTY

@misc{pith2026260100212,
  author       = {Pith},
  title        = {Pith review of: IntraStyler: Intra-Domain Style Synthesis for Cross-Modality MRI Domain Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IDSAVTTY}},
  note         = {Machine review of arXiv:2601.00212}
}

read the original abstract

Segmentation of vestibular schwannoma and cochlea from T2 MRI is clinically important yet annotation-intensive. Domain adaptation (DA) has been widely adopted to bridge the gap between labeled contrast-enhanced T1 and unlabeled T2 datasets. While existing methods focus on cross-domain alignment, intra-domain variability within the target domain remains largely overlooked. Images from the same domain may vary substantially due to different scanners, field strengths, and acquisition protocols. Ignoring this variability produces homogeneous synthetic images that limit the generalizability of downstream segmentation models. To address this, we propose IntraStyler, a 3D unpaired image translation method that automatically discovers fine-grained intra-domain styles without any predefined sub-domains, and synthesizes diverse target domain images using per-image style references. To this end, we design a 3D style encoder trained with a novel contrastive learning objective to extract style-only embeddings disentangled from anatomy. IntraStyler is built upon the 1st place CrossMoDA challenge solution and further advances it, generating more diverse synthetic data and achieving more reliable downstream segmentation. Code is available at https://github.com/MedICL-VU/IntraStyler.

Figures

Figures reproduced from arXiv: 2601.00212 by the authors.

**Figure 1.** Left: an illustration of intra-domain variability in UDA tasks. Many subdomains (scanner type) exist in the target domain (T2 MRI). Right: a comparison of different image translation strategies to generate diverse output styles. unpaired and unlabeled. The goal is to leverage these data so that a target domain model can be obtained without any target domain annotations. Image-level domain alignment (ILDA) has been … view at source ↗

**Figure 2.** Illustration of IntraStyler (left) and our contrastive learning setup (right). specified sub-domains. We evaluate our method on both synthesis and segmentation tasks from the largest public dataset for cross-modality domain adaptation, CrossMoDA 2023. Our contributions are summarized as follows: • We propose an exemplar-based unpaired image translation method for controllable style synthesis, which does not requir… view at source ↗

**Figure 3.** As the SLERP interpolated style (orange arrow) rotates clockwise, the style of the interpolated image also transitions between two exemplar images from left to right. Lcon = −sim(ES(y), ES(G(x))) (2) where x ∈ X is the source domain input image and G is the synthesis network. Since the style vectors are unit vectors, sim(·) is simply the dot product. Final Objective. of IntraStyler is expressed as: LCUT +λstyleLstyl… view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Top left: the style embedding space of all unlabeled target domain images. Clustering is done by K-means. Top right: representative MR images selected from two local regions in the style embedding space. It can be seen that the MR images from the same region have consi…

**Figure 5.** Figure 5: The segmentation performance on the CrossMoDA 2023 validation leaderboard. The results were obtained by training nnU-Net on the synthetic T2 images generated by different synthesis strategies. For ASSD plots, the number of failure cases (i.e., no segmentation for the t…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 6 linked inside Pith

[1]

arXiv preprint arXiv:2109.14219 (2021)

Dong, H., Yu, F., Zhao, J., Dong, B., Zhang, L.: Unsupervised domain adaptation in semantic segmentation based on pixel alignment and self-training. arXiv preprint arXiv:2109.14219 (2021)

arXiv 2021
[2]

Medical Image Analysis83, 102628 (2023)

Dorent, R., Kujawa, A., Ivory, M., Bakas, S., Rieke, N., Joutard, S., Glocker, B., Cardoso, J., Modat, M., Batmanghelich, K., et al.: Crossmoda 2021 challenge: Benchmark of cross-modality domain adaptation techniques for vestibular schwan- noma and cochlea segmentation. Medical Image Analysis83, 102628 (2023)

2021
[3]

In: International Conference on Learning Representations (2017),https: //openreview.net/forum?id=BJO-BuT1g

Dumoulin, V., Shlens, J., Kudlur, M.: A learned representation for artistic style. In: International Conference on Learning Representations (2017),https: //openreview.net/forum?id=BJO-BuT1g

2017
[4]

arXiv preprint arXiv:2202.02794 (2022)

Hacohen, G., Dekel, A., Weinshall, D.: Active learning on a budget: Opposite strategies suit high and low budgets. arXiv preprint arXiv:2202.02794 (2022)

arXiv 2022
[5]

arXiv preprint arXiv:2311.15090 (2023)

Han, L., Tan, T., Mann, R.: Fine-grained unsupervised cross-modality do- main adaptation for vestibular schwannoma segmentation. arXiv preprint arXiv:2311.15090 (2023)

arXiv 2023
[6]

In: Proceedings of the IEEE international conference on computer vision

Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: Proceedings of the IEEE international conference on computer vision. pp. 1501–1510 (2017)

2017
[7]

In: Proceedings of the European conference on computer vision (ECCV)

Huang, X., Liu, M.Y., Belongie, S., Kautz, J.: Multimodal unsupervised image-to- image translation. In: Proceedings of the European conference on computer vision (ECCV). pp. 172–189 (2018) 10 H. Liu et al

2018
[8]

Nature methods18(2), 203–211 (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 methods18(2), 203–211 (2021)

2021

Show all 24 references

[9]

General Scientific Researches2(1), 13–17 (2014)

Jafari, M., Molaei, H.: Spherical linear interpolation and b´ ezier curves. General Scientific Researches2(1), 13–17 (2014)

2014
[10]

In: Proceedings of the AAAI conference on artificial intelligence

Jing, Y., Liu, X., Ding, Y., Wang, X., Ding, E., Song, M., Wen, S.: Dynamic instance normalization for arbitrary style transfer. In: Proceedings of the AAAI conference on artificial intelligence. vol. 34, pp. 4369–4376 (2020)

2020
[11]

In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition

Karras, T., Laine, S., Aila, T.: A style-based generator architecture for generative adversarial networks. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 4401–4410 (2019)

2019
[12]

medRxiv pp

Kujawa, A., Dorent, R., Connor, S., Thomson, S., Ivory, M., Vahedi, A., Guil- hem, E., Bradford, R., Kitchen, N., Bisdas, S., et al.: Deep learning for automatic segmentation of vestibular schwannoma: A retrospective study from multi-centre routine mri. medRxiv pp. 2022–08 (2022)

2022
[13]

Radiology: Artificial Intelligence6(5), e230521 (2024)

Li, H., Liu, H., von Busch, H., Grimm, R., Huisman, H., Tong, A., Winkel, D., Penzkofer, T., Shabunin, I., Choi, M.H., et al.: Deep learning–based unsupervised domain adaptation via a unified model for prostate lesion detection using multisite biparametric mri datasets. Radiol...

2024
[14]

In: International MICCAI Brainlesion Workshop

Liu, H., Fan, Y., Oguz, I., Dawant, B.M.: Enhancing data diversity for self-training based unsupervised cross-modality vestibular schwannoma and cochlea segmen- tation. In: International MICCAI Brainlesion Workshop. pp. 109–118. Springer (2022)

2022
[15]

In: Interna- tional Challenge on Cross-Modality Domain Adaptation for Medical Image Seg- mentation, pp

Liu, H., Fan, Y., Xu, Z., Dawant, B.M., Oguz, I.: Learning site-specific styles for multi-institutional unsupervised cross-modality domain adaptation. In: Interna- tional Challenge on Cross-Modality Domain Adaptation for Medical Image Seg- mentation, pp. 372–385. Springer (2023)

2023
[16]

Advances in neural information processing systems30(2017)

Liu, M.Y., Breuel, T., Kautz, J.: Unsupervised image-to-image translation net- works. Advances in neural information processing systems30(2017)

2017
[17]

In: Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part IX 16

Park, T., Efros, A.A., Zhang, R., Zhu, J.Y.: Contrastive learning for unpaired image-to-image translation. In: Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part IX 16. pp. 319–

2020
[18]

arXiv preprint arXiv:2203.16557 (2022)

Shin, H., Kim, H., Kim, S., Jun, Y., Eo, T., Hwang, D.: Cosmos: Cross- modality unsupervised domain adaptation for 3d medical image segmentation based on target-aware domain translation and iterative self-training. arXiv preprint arXiv:2203.16557 (2022)

2022 arXiv
[19]

arXiv preprint arXiv:2506.12006 (2025)

Wijethilake, N., Dorent, R., Ivory, M., Kujawa, A., Cornelissen, S., Langenhuizen, P., Okasha, M., Oviedova, A., Dong, H., Kang, B., et al.: crossmoda challenge: Evo- lution of cross-modality domain adaptation techniques for vestibular schwannoma and cochlea segmentation from ...

2021 arXiv
[20]

In: International Workshop on Machine Learning in Clin- ical Neuroimaging

Wijethilake, N., Kujawa, A., Dorent, R., Asad, M., Oviedova, A., Vercauteren, T., Shapey, J.: Boundary distance loss for intra-/extra-meatal segmentation of vestibular schwannoma. In: International Workshop on Machine Learning in Clin- ical Neuroimaging. pp. 73–82. Springer (2022)

2022
[21]

In: Proceedings of the IEEE interna- tional conference on computer vision

Zhu, J.Y., Park, T., Isola, P., Efros, A.A.: Unpaired image-to-image translation using cycle-consistent adversarial networks. In: Proceedings of the IEEE interna- tional conference on computer vision. pp. 2223–2232 (2017) IntraStyler 11

2017
[22]

Advances in neural information processing systems30(2017)

Zhu, J.Y., Zhang, R., Pathak, D., Darrell, T., Efros, A.A., Wang, O., Shechtman, E.: Toward multimodal image-to-image translation. Advances in neural information processing systems30(2017)

2017
[23]

arXiv preprint arXiv:2311.11578 (2023)

Zhuang, Y.: A 3d multi-style cross-modality segmentation framework for segment- ing vestibular schwannoma and cochlea. arXiv preprint arXiv:2311.11578 (2023)

2023 arXiv
[24]

In: MICCAI Workshop on Data Augmen- tation, Labelling, and Imperfections

Zuo, L., Liu, Y., Xue, Y., Han, S., Bilgel, M., Resnick, S.M., Prince, J.L., Carass, A.: Disentangling a single mr modality. In: MICCAI Workshop on Data Augmen- tation, Labelling, and Imperfections. pp. 54–63. Springer (2022)

2022

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.

[1] [1]

arXiv preprint arXiv:2109.14219 (2021)

Dong, H., Yu, F., Zhao, J., Dong, B., Zhang, L.: Unsupervised domain adaptation in semantic segmentation based on pixel alignment and self-training. arXiv preprint arXiv:2109.14219 (2021)

arXiv 2021

[2] [2]

Medical Image Analysis83, 102628 (2023)

Dorent, R., Kujawa, A., Ivory, M., Bakas, S., Rieke, N., Joutard, S., Glocker, B., Cardoso, J., Modat, M., Batmanghelich, K., et al.: Crossmoda 2021 challenge: Benchmark of cross-modality domain adaptation techniques for vestibular schwan- noma and cochlea segmentation. Medical Image Analysis83, 102628 (2023)

2021

[3] [3]

In: International Conference on Learning Representations (2017),https: //openreview.net/forum?id=BJO-BuT1g

Dumoulin, V., Shlens, J., Kudlur, M.: A learned representation for artistic style. In: International Conference on Learning Representations (2017),https: //openreview.net/forum?id=BJO-BuT1g

2017

[4] [4]

arXiv preprint arXiv:2202.02794 (2022)

Hacohen, G., Dekel, A., Weinshall, D.: Active learning on a budget: Opposite strategies suit high and low budgets. arXiv preprint arXiv:2202.02794 (2022)

arXiv 2022

[5] [5]

arXiv preprint arXiv:2311.15090 (2023)

Han, L., Tan, T., Mann, R.: Fine-grained unsupervised cross-modality do- main adaptation for vestibular schwannoma segmentation. arXiv preprint arXiv:2311.15090 (2023)

arXiv 2023

[6] [6]

In: Proceedings of the IEEE international conference on computer vision

Huang, X., Belongie, S.: Arbitrary style transfer in real-time with adaptive instance normalization. In: Proceedings of the IEEE international conference on computer vision. pp. 1501–1510 (2017)

2017

[7] [7]

In: Proceedings of the European conference on computer vision (ECCV)

Huang, X., Liu, M.Y., Belongie, S., Kautz, J.: Multimodal unsupervised image-to- image translation. In: Proceedings of the European conference on computer vision (ECCV). pp. 172–189 (2018) 10 H. Liu et al

2018

[8] [8]

Nature methods18(2), 203–211 (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 methods18(2), 203–211 (2021)

2021