ConStyX: Content Style Augmentation for Generalizable Medical Image Segmentation

Jinzhu Yang; Osmar R. Zaiane; Peng Cao; Xi Chen; Zhiqiang Shen

REVIEW 4 major objections 5 minor 26 references

ConStyX: Content Style Augmentation for Generalizable Medical Image Segmentation

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read ConStyX claims that moving a deep feature along directions captured by intra-class covariance and the smallest segmentation-loss gradient components produces new features that behave like augmented same-class images, and that reweighting…

desk verdict A plausible feature-space augmentation recipe with a modest benchmark gain, but the core mechanism is unverified and the results lack error bars. read the letter →

arxiv 2506.10675 v1 pith:VRFUUAMF submitted 2025-06-12 eess.IV cs.CV

Xi Chen , Zhiqiang Shen , Peng Cao , Jinzhu Yang , Osmar R. Zaiane This is my paper · ORCID

classification eess.IVcs.CV

keywords domaingeneralizationmedicalimagesegmentationfundusimagesfeatureaugmentationcontentstyledeepinterpolationopticdiscrandomization

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 claims that perturbing a deep feature along two guided directions—the intra-class variation captured by a class-conditional covariance, and the k smallest loss-gradient feature components—yields new features that correspond to plausible same-class images with changed content and style. It proposes ConStyX, which uses these perturbed features as extra training data and reweights them according to cosine similarity to the original feature and prediction confidence, so that over-augmented features do not harm training. On five fundus datasets under the single-domain generalization setting, the method reports an average Dice of 78.88%, beating eight domain-generalization baselines and the second-best method by 2.09 points. If the assumption holds, the work offers a controllable feature-level augmentation recipe that goes beyond style-only perturbation.

What carries the argument

The machinery is a two-part feature-space augmentation scheme. DFA constructs two movement distributions: a multivariate normal $\mathcal{N}(0, \lambda_1 \Sigma_c)$ driven by an online-estimated class-conditional covariance $\Sigma_c$, and a uniform distribution $\mathcal{U}(0, \lambda_2 v_i^j)$ over a binary mask selecting the $k$ channels with the smallest loss-gradient magnitudes. AFU combines a cosine-similarity map $S_i$ and a prediction-confidence map $F_i$ into a pixel-wise training weight $W_i$, which is applied to the segmentation loss. These two components together define how features are moved and how the moved features are used during training.

What would settle it

Decode or project the augmented features back to image space and inspect the resulting images: if the moved features do not reconstruct into plausible same-class images with intact structures, the augmentation is noise. Alternatively, replace the guided movement directions with random directions of the same magnitude; if the reported DSC gain disappears, the direction selection is not the cause of the improvement.

Watch

Extended reading notes

Core claim

The central discovery is that deep feature movement, guided by intra-class variation and feature-gradient masks, provides a controllable form of content and style augmentation for domain generalization. The Deep Feature Augmentation algorithm (DFA) samples movement vectors from a zero-mean Gaussian with class-conditional covariance, pushing features along the maximum intra-class variation direction, and from a uniform distribution over the k channels with the smallest segmentation-loss partial derivatives, pushing features toward cross-domain directions while preserving semantics. The Augmented Feature Utilization strategy (AFU) then weights each augmented feature by its cosine similarity to the original feature and its prediction confidence, assigning lower weights to over-augmented features and higher weights to well-augmented ones. On joint optic disc and optic cup segmentation across five fundus datasets, ConStyX reports an average DSC of 78.88%, outperforming the second-best method by 2.09 points and the baseline by about 9.8 points.

Load-bearing premise

The method's central premise is that moving a deep feature along the class-covariance direction and along the smallest-loss-gradient channels produces a new feature that still represents the same class, with changed content and style, rather than just adding noise.

Editorial extensions

If this is right

If the augmentation directions are correct, training with content-and-style-augmented features should improve segmentation on unseen domains beyond what style-only augmentation methods achieve.
The AFU weighting scheme implies that automatically distinguishing well-augmented from over-augmented features can prevent performance degradation, making augmentation procedures more robust.
The method is defined at the feature level, so it could be applied to other segmentation backbones and medical image modalities without changing the image-space augmentation pipeline.
The reported gains on five fundus datasets suggest that feature-level content augmentation may generalize to other anatomical structures where intra-class variation is meaningful.

Reading between the lines

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

A direct test of the core assumption would be to decode the augmented feature maps back into image space and check whether they look like realistic same-class images; the paper does not perform this verification.
The claim that Gaussian sampling implements the maximum intra-class variation direction is imprecise, because sampling from $\mathcal{N}(0, \lambda_1 \Sigma_c)$ draws all covariant directions, not only the top eigenvector.
The method could be combined with image-level augmentation, since the two operate at different levels and might provide complementary coverage of unseen domains.
The threshold $\tau$ and the scaled factors $\lambda_1, \lambda_2$ are set by hand; their sensitivity across different datasets or tasks is not explored, so a tuning-free version may need additional validation.

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

Quick take: this is a workmanlike, incremental contribution to single-domain generalization for fundus segmentation. What's new is the combination: class-conditional covariance sampling plus gradient-masked uniform perturbation for feature augmentation, with a similarity/confidence re-weighting scheme. The experiments cover five standard datasets, the code is released, and the reported 2.09 average DSC gain over CCSDG is plausible. Credit goes to a clear experimental structure and honest comparison against eight baselines.

The soft spots are real. First, no error bars or significance tests anywhere; two points on a five-domain average could easily be within run-to-run noise. Second, the position analysis in Table 3 selects the best of three schemes on the very test domains it then reports as the final result—that's post-hoc selection, not a validation. Third, the paper claims sampling from N(0, lambda1 Sigma_c) captures the 'maximum intra-class variation direction,' but a Gaussian draw samples all covariance directions; the wording is inaccurate. Fourth, and most load-bearing, the paper never checks that the moved features actually correspond to realistic same-class images in input space. The stress-test concern about AFU also lands: a feature pushed across a class boundary and predicted with high confidence gets a similarity below tau and confidence high enough to receive weight e^F - 1, so the re-weighting can upweight exactly the over-augmented features it claims to suppress.

None of this sinks the paper. The gains may come from the perturbation acting as a generic regularizer, which is a valid though less exciting story, but the paper asserts a richer mechanism it hasn't verified. Since the code is available, the checks are straightforward: show a handful of decoded augmented features or measure same-class accuracy of the augmented features, and add confidence intervals from repeated runs.

For peer review: I'd send it out. The idea is clear, the experiments are reproducible, and the missing verification is addressable. A competent referee can get this to an acceptable state. Who this is for: people working on domain generalization for medical images, especially fundus. It's not a breakthrough, but it's a genuine data point.

Referee Report

4 major / 5 minor

Summary. The paper proposes ConStyX, a domain-randomization method for single-domain generalization in medical image segmentation. It augments deep features by adding two perturbation vectors: a sample from a class-conditional Gaussian distribution (intra-class variation) and a uniform sample along the k minimum-loss-gradient channels (cross-domain variation), as specified in Eq. (1). The augmented features are then reweighted pixel-wise by cosine similarity to the original features and by prediction confidence, as specified in Eq. (4). Experiments on five fundus datasets for joint optic disc and optic cup segmentation report an average Dice of 78.88% against 76.19% for the strongest baseline, CCSDG, with additional ablations in Tables 2 and 3.

Significance. If the underlying mechanism is valid, ConStyX would be a useful contribution: it extends feature-space augmentation beyond style-only perturbation and explicitly attempts to down-weight over-augmented features, which is a practical concern in medical-image DG. The paper is evaluated on five public fundus datasets with a consistent U-Net backbone, compares against eight recent baselines, and releases code, all of which are strengths. However, the significance currently hinges on two unverified premises: that the feature movements in Eq. (1) produce realistic same-class content/style variation, and that the AFU reweighting in Eq. (4) suppresses rather than amplifies over-augmented features. The reported performance advantage is also not quantified with variance or significance tests. These issues are addressable within the paper's scope, so the work is promising but needs substantive revision.

major comments (4)

[Section 3, Tables 1-3] The central empirical claim that ConStyX 'consistently outperforms' eight baselines rests on single DSC values with no standard deviations, no multiple-seed runs, and no significance tests. The reported 2.09-point average advantage over CCSDG (78.88% vs. 76.19%) is therefore unquantified, and the ablation improvements in Table 2 could reflect run-to-run variance. Please report mean±std over at least three seeds and a paired significance test for the main comparison.
[Section 2.1, Eq. (1)] The paper's core assumption—that adding α_ic ∼ N(0, λ1 Σ_c) and α_cd ∼ U(0, λ2 v_i^j) to z_i^j produces a feature that decodes to another same-class image with different content and style—is never verified in image space. Moreover, the text says the Gaussian samples 'the maximum intra-class variation direction,' but sampling from the full class-conditional covariance draws all covariant directions, and a sufficiently large displacement can cross a class boundary. Please provide direct evidence (e.g., decoding moved features back to images, or measuring class-prediction preservation before and after augmentation) that the moved features remain semantically valid; otherwise the observed gains may be attributable to generic feature-space noise rather than content/style augmentation.
[Section 2.2, Eq. (4)] The AFU reweighting assigns weight e^{F_i^j} - 1 > 0 to all augmented pixels whose cosine similarity to the original feature is below τ, whenever prediction confidence is high. Because prediction confidence is not a measure of semantic correctness, a feature that has been moved into a wrong class and is confidently predicted will be upweighted, which is exactly the over-augmentation the paper says it suppresses. The authors should either replace the confidence term with a semantic-preservation measure or empirically demonstrate that high-confidence, low-similarity augmented features are in fact well-augmented.
[Section 3.2, Table 3] The choice of perturbing the minimum-k gradient positions is compared with random and maximum-k positions using the same five test domains that produce the headline result, and the best configuration is then used in the final model. This is a post-hoc model selection on the test domains; a held-out validation domain, or a sensitivity analysis over k, λ1, λ2, and τ, is needed to support the claim that the chosen direction is the reason for the gains.

minor comments (5)

[Fig. 2] The label 'Confidece Map' should read 'Confidence Map'.
[Section 2.1, Eq. (3)] The definition of pos_j^i says 'the position' but refers to a set of k positions; the notation should be updated accordingly.
[Section 1 (Contributions)] The contributions list mentions 'five state-of-the-art DG methods,' but Table 1 compares against eight methods; this inconsistency should be corrected.
[Section 2.2] The Min-Max normalization in the definition of F_i is not specified as per-image or per-batch; please clarify.
[Section 3 (Implementation Details)] The hyperparameters λ1, λ2, k, and τ are fixed without sensitivity analysis; at least a brief study or discussion would help readers understand robustness to these choices.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported gains are an empirical benchmark outcome, and no load-bearing step reduces to its own inputs by construction.

full rationale

I walked the derivation chain in Sections 2.1 (DFA) and 2.2 (AFU) and found no step in which a prediction is equivalent to a fitted input by construction. Equation (1) builds augmented features from class-conditional covariance and gradient-based masks; these quantities are computed from the source data and segmentation loss, but the reported DSC values in Table 1 are measured on external fundus datasets (BinRushed, Magrabia, REFUGE, ORIGA, Drishti-GS) rather than being recovered from the fitted statistics. The AFU weights in Eq. (4) are derived from cosine similarity and prediction confidence, and the three-way categorization of augmented features is an operational definition, not a self-referential prediction. The paper's core assumption that moving a deep feature produces a same-class image with altered content and style is borrowed from external prior work (references [16] and [18]) and is open to a validity critique: the paper does not verify that the moved features decode to plausible same-class images, and sampling from N(0, λ1 Σ_c) does not literally move only along the maximum-variation direction. These are correctness and precision concerns, not circularity. No load-bearing argument rests on a self-citation; the cited DG baselines and interpolation works are not by the present authors. Hyperparameters such as k=5, λ1=1, λ2=0.5, and τ=0.6 are hand-set training choices, not fitted values renamed as predictions. Under the hard rule that circularity requires a quotable reduction by construction, the honest finding is no significant circularity (score 0).

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

The method rests on two content-style assumptions about the deep feature space and four hand-set hyperparameters. No new physical entities are introduced. The covariance and gradient-mask machinery are standard operations; the load-bearing premise is that feature-space perturbations correspond to plausible image-space augmentations of the same class, which is untested.

free parameters (4)

lambda1 = 1
Scale factor for intra-class Gaussian covariance sampling in Eq. 1; chosen by hand, no sensitivity analysis.
lambda2 = 0.5
Scale factor for gradient-guided uniform sampling in Eq. 1; hand-set hyperparameter.
k = 5
Number of minimum-gradient feature positions masked in Eq. 3; chosen by hand, no sensitivity analysis.
tau = 0.6
Cosine-similarity threshold separating trivial, well, and over-augmented features in Eq. 4; hand-set.

assumptions (3)

domain assumption Moving a deep feature along a direction yields a new feature corresponding to another same-class image with different content and style.
Core premise stated in Section 2 overview and Fig. 1(c); cited to [16,18], not proven for fundus features.
domain assumption Perturbing the k smallest gradient components minimally affects semantics and enables cross-domain variation.
Used to build the feature mask v in Section 2.1; supported only by the position ablation in Table 3.
domain assumption Class-conditional covariance estimated with online weighted averaging captures the intra-class variation direction.
Justifies sampling alpha_ic ~ N(0, lambda1 Sigma_c); the authors describe this as 'maximum intra-class variation', though Gaussian sampling draws all directions.

how reviews work

0 comments

Cite this review

Pith. "Pith review of ConStyX: Content Style Augmentation for Generalizable Medical Image Segmentation." pith.science (2026). https://pith.science/paper/VRFUUAMF

@misc{pith2026250610675,
  author       = {Pith},
  title        = {Pith review of: ConStyX: Content Style Augmentation for Generalizable Medical Image Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VRFUUAMF}},
  note         = {Machine review of arXiv:2506.10675}
}

read the original abstract

Medical images are usually collected from multiple domains, leading to domain shifts that impair the performance of medical image segmentation models. Domain Generalization (DG) aims to address this issue by training a robust model with strong generalizability. Recently, numerous domain randomization-based DG methods have been proposed. However, these methods suffer from the following limitations: 1) constrained efficiency of domain randomization due to their exclusive dependence on image style perturbation, and 2) neglect of the adverse effects of over-augmented images on model training. To address these issues, we propose a novel domain randomization-based DG method, called content style augmentation (ConStyX), for generalizable medical image segmentation. Specifically, ConStyX 1) augments the content and style of training data, allowing the augmented training data to better cover a wider range of data domains, and 2) leverages well-augmented features while mitigating the negative effects of over-augmented features during model training. Extensive experiments across multiple domains demonstrate that our ConStyX achieves superior generalization performance. The code is available at https://github.com/jwxsp1/ConStyX.

Figures

Figures reproduced from arXiv: 2506.10675 by the authors.

**Figure 1.** Visualization of original and augmented images. (a) Original images (left) and style-augmented images (right). (b) Original images (left) and over-augmented images (right). (c) In the deep feature space, when the feature vector of a sample moves along a specific direction (gray dashed arrow), the resulting feature vector corresponds to an augmented image (✓: well-augmented images or ×: over-augmented images). Existi… view at source ↗

**Figure 2.** Overview of the proposed Content Style Augmentation framework (ConStyX). It includes: 1) Deep Feature Augmentation algorithm (DFA) for content-style augmentation under the guidance of intra-class variation and feature gradient and 2) Augmented Feature Utilization strategy (AFU) for augmented feature re-weighting. 2.1 Deep Feature Augmentation (DFA) The fundamental concept of DFA lies in How to appropriately move t… view at source ↗

**Figure 3.** Visualization of OD and OC segmentation results. Blue dashed circles highlight some regions with significant differences in the segmentation results [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

26 extracted references · 18 canonical work pages

[1]

Artificial intelligence review 54, 137–178 (2021)

Asgari Taghanaki, S., Abhishek, K., Cohen, J.P., Cohen-Adad, J., Hamarneh, G.: Deep semantic segmentation of natural and medical images: a review. Artificial intelligence review 54, 137–178 (2021)

work page 2021
[2]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Chen, C., Li, Z., Ouyang, C., Sinclair, M., Bai, W., Rueckert, D.: Maxstyle: Adver- sarial style composition for robust medical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 151–161. Springer (2022)

2022
[3]

In: International Conference on Medical Image Computing and Computer- Assisted Intervention

Chen, Z., Pan, Y., Ye, Y., Cui, H., Xia, Y.: Treasure in distribution: a domain ran- domization based multi-source domain generalization for 2d medical image segmen- tation. In: International Conference on Medical Image Computing and Computer- Assisted Intervention. pp. 89–99. Springer (2023)

work page 2023
[4]

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

Choi, S., Jung, S., Yun, H., Kim, J.T., Kim, S., Choo, J.: Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whiten- ing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11580–11590 (2021)

work page 2021
[5]

IEEE Transactions on Biomedical Engineering69(3), 1173–1185 (2021)

Guan, H., Liu, M.: Domain adaptation for medical image analysis: a survey. IEEE Transactions on Biomedical Engineering69(3), 1173–1185 (2021)

2021
[6]

He,K.,Zhang,X.,Ren,S.,Sun,J.:Deepresiduallearningforimagerecognition.In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770–778 (2016)

2016
[7]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Hu, S., Liao, Z., Xia, Y.: Devil is in channels: Contrastive single domain general- ization for medical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 14–23. Springer (2023) 10 X. Chen et al

work page 2023
[8]

In: European Conference on Computer Vision

Jia, Y., Hoyer, L., Huang, S., Wang, T., Van Gool, L., Schindler, K., Obukhov, A.: Dginstyle: Domain-generalizable semantic segmentation with image diffusion mod- els and stylized semantic control. In: European Conference on Computer Vision. pp. 91–109. Springer (2025)

work page 2025

Show all 26 references

[9]

In: Journal of Physics: Conference Series

Latha, G., Priya, P.A.: Glaucoma retinal image detection and classification using machine learning algorithms. In: Journal of Physics: Conference Series. vol. 2335, p. 012025. IOP Publishing (2022)

2022
[10]

arXiv preprint arXiv:2202.03958 (2022)

Li, X., Dai, Y., Ge, Y., Liu, J., Shan, Y., Duan, L.Y.: Uncertainty modeling for out-of-distribution generalization. arXiv preprint arXiv:2202.03958 (2022)

2022 arXiv
[11]

Medical Im- age Analysis p

Orlando, J.I., Fu, H., Barbosa Breda, J., van Keer, K., Bathula, D.R., Diaz-Pinto, A.,Fang,R.,Heng,P.A.,Kim,J.,Lee,J.,Lee,J.,Li,X.,Liu,P.,Lu,S.,Murugesan, B., Naranjo, V., Phaye, S.S.R., Shankaranarayana, S.M., Sikka, A., Son, J., van den Hengel, A., Wang, S., Wu, J., Wu, Z., ...

2020
[12]

Advances in neural information processing sys- tems 32 (2019)

Paszke, A., Gross, S., Massa, F., Lerer, A., Bradbury, J., Chanan, G., Killeen, T., Lin, Z., Gimelshein, N., Antiga, L., et al.: Pytorch: An imperative style, high- performance deep learning library. Advances in neural information processing sys- tems 32 (2019)

2019
[13]

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

Peng, D., Lei, Y., Hayat, M., Guo, Y., Li, W.: Semantic-aware domain generalized segmentation. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 2594–2605 (2022)

2022
[14]

In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, Oc- tober 5-9, 2015, proceedings, part III 18

Ronneberger, O., Fischer, P., Brox, T.: U-net: Convolutional networks for biomed- ical image segmentation. In: Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international conference, Munich, Germany, Oc- tober 5-9, 2015, proceedings, part III 18....

2015
[15]

In: 2014 IEEE 11th International Symposium on Biomedical Imaging (ISBI)

Sivaswamy, J., Krishnadas, S.R., Datt Joshi, G., Jain, M., Syed Tabish, A.U.: Drishti-gs: Retinal image dataset for optic nerve head(onh) segmentation. In: 2014 IEEE 11th International Symposium on Biomedical Imaging (ISBI). pp. 53–56 (2014). https://doi.org/10.1109/ISBI.2014.6867807

2014
[16]

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

Upchurch, P., Gardner, J., Pleiss, G., Pless, R., Snavely, N., Bala, K., Weinberger, K.: Deep feature interpolation for image content changes. In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 7064–7073 (2017)

2017
[17]

IEEE transactions on knowledge and data engineering35(8), 8052–8072 (2022)

Wang, J., Lan, C., Liu, C., Ouyang, Y., Qin, T., Lu, W., Chen, Y., Zeng, W., Philip, S.Y.: Generalizing to unseen domains: A survey on domain generalization. IEEE transactions on knowledge and data engineering35(8), 8052–8072 (2022)

2022
[18]

IEEE Transactions on Pattern Analysis and Machine Intelligence44(7), 3733–3748 (2021)

Wang, Y., Huang, G., Song, S., Pan, X., Xia, Y., Wu, C.: Regularizing deep net- works with semantic data augmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence44(7), 3733–3748 (2021)

2021
[19]

Medical Image Analysis 69, 101985 (2021)

Xie, X., Niu, J., Liu, X., Chen, Z., Tang, S., Yu, S.: A survey on incorporating domain knowledge into deep learning for medical image analysis. Medical Image Analysis 69, 101985 (2021)

2021
[20]

arXiv preprint arXiv:2007.13003 (2020)

Xu, Z., Liu, D., Yang, J., Raffel, C., Niethammer, M.: Robust and general- izable visual representation learning via random convolutions. arXiv preprint arXiv:2007.13003 (2020)

2020 arXiv
[21]

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

Zhang, Y., Li, M., Li, R., Jia, K., Zhang, L.: Exact feature distribution match- ing for arbitrary style transfer and domain generalization. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 8035–8045 (2022) Content Style Augmentation 11

2022
[22]

In: Proceedings of the 2010 Annual International Conference of the IEEE Engineering in Medicine and Biology

Zhang, Z., Yin, F., Liu, J., Wong, W., Tan, N., Lee, B., Cheng, J., Wong, T.: Origa: An online retinal fundus image database for glaucoma analysis and research. In: Proceedings of the 2010 Annual International Conference of the IEEE Engineering in Medicine and Biology. pp. 3065–3068

2010
[23]

In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

Zhang, Z., Wang, B., Jha, D., Demir, U., Bagci, U.: Domain generalization with correlated style uncertainty. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2000–2009 (2024)

2024
[24]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Zhao, H., Dong, W., Yu, R., Zhao, Z., Du, B., Xu, Y.: Morestyle: relax low- frequency constraint of fourier-based image reconstruction in generalizable medical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 4...

2024
[25]

IEEE Transactions on Pattern Analysis and Machine Intelligence45(4), 4396–4415 (2022)

Zhou, K., Liu, Z., Qiao, Y., Xiang, T., Loy, C.C.: Domain generalization: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence45(4), 4396–4415 (2022)

2022
[26]

In: ICLR (2021)

Zhou, K., Yang, Y., Qiao, Y., Xiang, T.: Domain generalization with mixstyle. In: ICLR (2021)

2021

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Artificial intelligence review 54, 137–178 (2021)

Asgari Taghanaki, S., Abhishek, K., Cohen, J.P., Cohen-Adad, J., Hamarneh, G.: Deep semantic segmentation of natural and medical images: a review. Artificial intelligence review 54, 137–178 (2021)

work page 2021

[2] [2]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Chen, C., Li, Z., Ouyang, C., Sinclair, M., Bai, W., Rueckert, D.: Maxstyle: Adver- sarial style composition for robust medical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 151–161. Springer (2022)

2022

[3] [3]

In: International Conference on Medical Image Computing and Computer- Assisted Intervention

Chen, Z., Pan, Y., Ye, Y., Cui, H., Xia, Y.: Treasure in distribution: a domain ran- domization based multi-source domain generalization for 2d medical image segmen- tation. In: International Conference on Medical Image Computing and Computer- Assisted Intervention. pp. 89–99. Springer (2023)

work page 2023

[4] [4]

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

Choi, S., Jung, S., Yun, H., Kim, J.T., Kim, S., Choo, J.: Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whiten- ing. In: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. pp. 11580–11590 (2021)

work page 2021

[5] [5]

IEEE Transactions on Biomedical Engineering69(3), 1173–1185 (2021)

Guan, H., Liu, M.: Domain adaptation for medical image analysis: a survey. IEEE Transactions on Biomedical Engineering69(3), 1173–1185 (2021)

2021

[6] [6]

He,K.,Zhang,X.,Ren,S.,Sun,J.:Deepresiduallearningforimagerecognition.In: Proceedings of the IEEE conference on computer vision and pattern recognition. pp. 770–778 (2016)

2016

[7] [7]

In: International Conference on Medical Image Computing and Computer-Assisted Intervention

Hu, S., Liao, Z., Xia, Y.: Devil is in channels: Contrastive single domain general- ization for medical image segmentation. In: International Conference on Medical Image Computing and Computer-Assisted Intervention. pp. 14–23. Springer (2023) 10 X. Chen et al

work page 2023

[8] [8]

In: European Conference on Computer Vision

Jia, Y., Hoyer, L., Huang, S., Wang, T., Van Gool, L., Schindler, K., Obukhov, A.: Dginstyle: Domain-generalizable semantic segmentation with image diffusion mod- els and stylized semantic control. In: European Conference on Computer Vision. pp. 91–109. Springer (2025)

work page 2025