Pith. sign in

REVIEW 4 major objections 6 minor 32 references

Learning Semantic Directions for Feature Augmentation in Domain-Generalized Medical Segmentation

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Learning which feature channels to perturb, and how strongly, improves medical segmentation on unseen domains.

desk verdict Solid feature-augmentation DG paper with a genuinely interesting selective consistency loss, but the core SDS mechanism is underspecified to the point that the main claim is not reproducible as written. read the letter →

arxiv 2507.23326 v1 pith:6JCNGFYU submitted 2025-07-31 cs.CV

Yingkai Wang , Yaoyao Zhu , Xiuding Cai , Yuhao Xiao , Haotian Wu , Yu Yao This is my paper
classification cs.CV
keywords domaingeneralizationmedicalimagesegmentationfeatureaugmentationsemanticdirectioncovariance-basedsamplingselectiveconsistencylossmulti-centerbenchmarks
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 claims that domain shift in medical image segmentation can be countered by perturbing deep features along learned semantic directions rather than by injecting the same noise into every channel. Its proposed framework augments an encoded feature $z$ as $\tilde z = z + d \odot s$, where a semantic direction selector learns which channels change and a covariance-based intensity sampler sets how much they change using inter-domain statistics. A selective consistency loss penalizes the augmented path only when it degrades segmentation, which the paper argues stabilizes the selector. On leave-one-domain-out evaluations over multi-center prostate MRI and fundus benchmarks, the paper reports the best average Dice among the compared domain generalization methods, and the visualizations indicate that perturbations concentrate in contextual regions rather than target anatomy. If correct, this offers a path to models that transfer across hospitals and scanners without target-domain data.

What carries the argument

The load-bearing object is the semantic direction mask $d \in \{0,1\}^C$, a per-channel on/off choice that turns blind feature noise into selective perturbation through $\tilde z_i = z_i + d \odot s$. The selector $f_{\mathrm{cs}}$ is described as two convolutional layers, global average pooling, and a fully connected layer, and the intensity $s$ is sampled as $\mu + \sigma \odot \xi$ with $\xi \sim N(0,\Sigma_k)$ where $\Sigma_k$ is the inter-domain covariance of a randomly chosen domain pair. The selective consistency loss is the stabilizer: by penalizing only augmentations that increase the segmentation loss, it gives the selector a training signal that favors channels whose perturbation is safe. The paper does not specify how the binary mask in Equation (2) is produced from continuous network outputs, so the learnability of the selector rests on an unstated binarization step.

What would settle it

Train the same framework with the learned direction selector replaced by a random binary mask of matched channel count; if the prostate average Dice stays at 86.34, the learned direction is not carrying the gain.

Watch

Extended reading notes

Core claim

The central claim is that channel-selective, statistics-guided feature perturbation generalizes better to unseen clinical domains than image-level augmentation or uniform feature-level augmentation. The paper defines the augmented feature as $\tilde z_i = z_i + d \odot s$, with $d \in \{0,1\}^C$ the output of a learnable semantic direction selector built from convolutions, global average pooling, and a fully connected layer, and $s$ an instance-adaptive intensity produced by shifting and scaling noise drawn from inter-domain covariance matrices. The supervised loss supervises segmentation from both the original and augmented features, and the selective consistency loss $\mathcal{L}_{\mathrm{cons}}$ applies only when the augmented feature gives a higher segmentation loss than the original, so harmful perturbations are penalized while helpful ones are not. On the three benchmark tasks the paper reports average Dice of 86.34 over 85.53 on prostate, 79.62 over 77.99 on optic cup, and 92.00 over 91.35 on optic disc.

Load-bearing premise

The method depends on the semantic direction selector being able to learn a useful binary channel mask by gradient descent, but the paper never specifies how the binary output in Equation (2) is obtained from continuous network outputs or how gradients flow through it.

Editorial extensions

If this is right

  • If the reported results are correct, SDFA becomes the top-scoring method on all three leave-one-domain-out tasks among the compared approaches, with the largest margin on optic cup segmentation.
  • The ablation results imply that the direction selector and intensity sampler each contribute independently: SDS alone improves average Dice to 85.63 and SIS alone to 85.51 from 85.29, and together they reach 86.34.
  • The selective consistency loss stabilizes selection: with it, the average Dice rises to 86.34 from 85.94 and the number of selected channels fluctuates less during training.
  • Reconstruction visualizations suggest the augmentation edits contextual regions while sparing target anatomy, which would be why segmentation Dice can improve without distorting the region of interest.
  • Because the method tunes feature noise by domain statistics rather than target data, it applies directly to the standard multi-source DG setting where only source centers are available at training time.

Reading between the lines

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

  • Editorial extension: if the gain comes from channel selectivity, then the number of selected channels per sample should correlate with domain gap; one could test this by checking whether low-selection samples come from source-like domains within the training set.
  • Editorial extension: the covariance-sampled intensity could be recycled for test-time adaptation by re-estimating $\Sigma_k$ on a small unlabeled target batch, a use the paper does not consider.
  • Editorial extension: the selective consistency loss is a 'do no harm' regularizer; replacing the hard indicator in Eq. (9) with a margin-based soft penalty could smooth optimization and may improve stability on smaller datasets, but this is not tested in the paper.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes a domain generalization framework (SDFA) for medical image segmentation, built around three components: a Semantic Direction Selector (SDS) that learns which feature channels to perturb, a covariance-based Semantic Intensity Sampler (SIS) that controls perturbation magnitudes via inter-domain statistics, and a Selective Consistency Loss (SCL) that only penalizes augmentations that degrade segmentation. The method is evaluated on the Prostate and Fundus benchmarks under a leave-one-domain-out protocol, reporting the best average DSC among compared methods in Tables II-IV.

Significance. If fully specified and reproducible, the method is a plausible structured feature-augmentation approach for domain generalization, and the use of public multi-center benchmarks with a fixed leave-one-domain-out protocol is a strength. However, the core technical details needed to implement and train the proposed SDS and SIS are missing, and the empirical claim of 'consistently outperforming' is not supported by the per-domain results. The paper would be a useful contribution only after filling these gaps and providing statistical evidence for the reported improvements.

major comments (4)
  1. [III-B1, Eq. (2)] Equation (2) defines the semantic direction mask as d = f_cs(z) ∈ {0,1}^C, but f_cs is specified only as 'two convolutional layers, followed by a global average pooling and a fully connected layer,' which outputs continuous logits. The paper never states how the binary mask is produced or how gradients flow through it during training. If a hard threshold is applied, the gradient of the total loss with respect to the SDS parameters is zero almost everywhere, so the selector cannot be learned by standard backpropagation; if a soft relaxation or straight-through estimator is used, Eq. (2) is inaccurate and the actual training procedure is unspecified. Since SDS is the core novel component, this omission makes the method irreproducible and undermines the empirical results in Tables II-IV.
  2. [III-B2, Eqs. (3)-(4)] The construction of the inter-domain covariance matrices Σ_k is not operationally defined. The text says 'we compute a set of covariance matrices {Σ_k}_{k=1}^N' where each Σ_k captures 'the inter-domain covariance between a pair of domains (D_si, D_sj)', but it never states which feature maps enter the covariance (per-sample features, per-batch means, or per-domain statistics), how the domain pair is chosen for a given feature z, or how the index k is enumerated (with K=4 or 6 source domains, the number of pairs is not N as in a batch). Also, because z = Φ(x_i) depends on the input, it is unclear whether Σ_k is a fixed precomputed statistic or a running batch estimate; the two interpretations lead to different training dynamics and different gradients. This needs to be specified precisely for reproduction.
  3. [III-C, Eq. (9)] The Selective Consistency Loss uses an indicator function I(Laug > Lori), which is non-differentiable with respect to the SDS parameters. The paper claims SCL 'stabilizes the learning of the SDS,' but without a differentiable surrogate or a gradient approximation, the loss as written cannot provide learning signal to the SDS. The absolute value in Eq. (9) is also non-smooth, though that is secondary. Please specify how gradients are computed through the indicator, or reformulate SCL as a smooth loss.
  4. [IV-C, Tables II-IV] The abstract claims the framework 'consistently outperforms existing domain generalization approaches,' but the per-domain results do not support 'consistently': in Table III, SDFA is below CSU on Domain-1 (80.40 vs 80.48) and below TriD on Domain-4 (81.94 vs 84.21); in Table IV, SDFA is below EFDMixStyle on Domain-3 (92.16 vs 93.19). In addition, no standard deviations or significance tests are reported, and the average DSC gaps are small (0.81, 1.63, and 0.65 points). Please report multiple seeds with variance and statistical tests, or revise the claim to 'best average performance in most settings.'
minor comments (6)
  1. [Table II caption] The caption says 'Average indicates the mean DSC and ASD across all four domains,' but the Prostate benchmark has six domains (Domain-1 through Domain-6); this should read 'six domains.'
  2. [IV-A] There is a stray subsection header '1) Quantitative Results on the Fundus Dataset:' immediately after the evaluation metrics paragraph, with no content following it; this appears to be a leftover from an earlier draft and should be removed.
  3. [Figure 3] The caption reads 'Segmentation results with and without consistency loss,' but the surrounding text describes the number of selected channels over training steps; the caption should be updated to match the actual content of the figure.
  4. [Tables VI and VII] The word 'Efficay' in the captions 'Efficay of Each Module' and 'Efficay of SCL' is misspelled; it should be 'Efficacy.'
  5. [Table V and surrounding text] The text states 'the DSC increases as λ grows,' but the reported average DSC at λ=0.2 is 85.51 and at λ=0.4 is 85.49, so the trend is not monotonic; the claim should be stated as 'DSC tends to increase with λ, with the best value at λ=1.0.'
  6. [Notation, Eqs. (3) and (9)] The symbol N is overloaded: in Eq. (3) it denotes the number of covariance matrices/domain pairs, while in Eq. (9) it denotes the batch size; please use distinct symbols such as N_pairs and N_batch.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: SDFA's claim rests on held-out-domain evaluation; the one self-citation is non-load-bearing.

full rationale

The derivation chain is self-contained. The augmented feature z̃_i = z_i + d⊙s (Eq. 1) is produced by the learnable selector f_cs (Eq. 2) and the covariance-based sampler (Eqs. 3-5), and the segmentation losses (Eqs. 6-8) plus the selective consistency term (Eq. 9) are all defined on source-domain training examples. No parameter is fitted to a target domain, and the leave-one-domain-out protocol in Sec. IV-A tests on held-out centers, so the average DSC gains in Tables II-IV are not forced by construction. The only self-citation, Ref. [25], is used to motivate a conceptual decomposition of semantic augmentation into direction and intensity; the paper adds its own learnable components and the empirical claim is evaluated independently. The unspecified binarization of d in Eq. (2) and the non-differentiable indicator in Eq. (9) are reproducibility/correctness concerns, but they do not make any prediction equivalent to an input. Hence no circular step is present.

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

The central claim rests on standard deep-learning fitting plus three domain assumptions: anatomical consistency, label preservation under feature-space perturbation, and the validity of covariance-sampled noise as a model of domain variation. The free parameters are the learnable intensity shift and scale, the loss weight lambda, and the network weights. No new physical entities are introduced.

free parameters (3)
  • mu (semantic intensity shift) = learned, in R^C, no final values reported
    Introduced in Eq. (5) to shift the sampled semantic intensity; fit on source-domain training data.
  • sigma (semantic intensity scale) = learned, in R^C, no final values reported
    Introduced in Eq. (5) to scale the covariance-sampled noise; fit on source-domain training data.
  • lambda (augmented-loss weight) = 1.0 (best in Table V)
    Balances original and augmented supervised losses in Eq. (6); tuned on the validation split over values from 0.2 to 1.0.
assumptions (4)
  • domain assumption Medical images have consistent anatomy, and domain shift mainly changes imaging conditions.
    Section I uses this to justify perturbing domain-variant features while preserving anatomical semantics; if false, feature perturbation could alter structures that matter for segmentation.
  • domain assumption Semantic augmentation in feature space, z~ = z + d ⊙ s, preserves segmentation labels when d is correctly selected.
    Eq. (1) and Section III-A invoke the linearization and semantic augmentation results of Refs. 23 to 25; the safety of the method rests on this.
  • domain assumption Covariance-sampled noise xi ~ N(0, Sigma_k) from paired domains is a valid model of domain-variant feature directions.
    Eq. (4) uses inter-domain covariance to generate perturbation intensity, but the estimator for Sigma_k is not defined and the assumption that such noise is semantically safe is not quantitatively tested.
  • ad hoc to paper The binary mask d in Eq. (2) can be optimized by gradient descent despite unspecified binarization.
    The paper never specifies a relaxation or straight-through estimator; trainability of the SDS is an implicit assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning Semantic Directions for Feature Augmentation in Domain-Generalized Medical Segmentation." pith.science (2026). https://pith.science/paper/6JCNGFYU

@misc{pith2026250723326,
  author       = {Pith},
  title        = {Pith review of: Learning Semantic Directions for Feature Augmentation in Domain-Generalized Medical Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6JCNGFYU}},
  note         = {Machine review of arXiv:2507.23326}
}
read the original abstract

Medical image segmentation plays a crucial role in clinical workflows, but domain shift often leads to performance degradation when models are applied to unseen clinical domains. This challenge arises due to variations in imaging conditions, scanner types, and acquisition protocols, limiting the practical deployment of segmentation models. Unlike natural images, medical images typically exhibit consistent anatomical structures across patients, with domain-specific variations mainly caused by imaging conditions. This unique characteristic makes medical image segmentation particularly challenging. To address this challenge, we propose a domain generalization framework tailored for medical image segmentation. Our approach improves robustness to domain-specific variations by introducing implicit feature perturbations guided by domain statistics. Specifically, we employ a learnable semantic direction selector and a covariance-based semantic intensity sampler to modulate domain-variant features while preserving task-relevant anatomical consistency. Furthermore, we design an adaptive consistency constraint that is selectively applied only when feature adjustment leads to degraded segmentation performance. This constraint encourages the adjusted features to align with the original predictions, thereby stabilizing feature selection and improving the reliability of the segmentation. Extensive experiments on two public multi-center benchmarks show that our framework consistently outperforms existing domain generalization approaches, achieving robust and generalizable segmentation performance across diverse clinical domains.

Figures

Figures reproduced from arXiv: 2507.23326 by the authors.

Figure 1
Figure 1. Overview of the proposed Semantic Direction Feature Augmentation (SDFA) framework. The model is trained on [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Qualitative results on the Prostate and Fundus Datasets [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Segmentation results with and without consistency loss. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visualization results of semantically augmented features. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 24 canonical work pages

  1. [1]

    A survey on deep learning in medical image analysis,

    G. Litjens, T. Kooi, B. E. Bejnordi, A. A. A. Setio, F. Ciompi, M. Ghafoorian, J. A. Van Der Laak, B. Van Ginneken, and C. I. S´anchez, “A survey on deep learning in medical image analysis,” Medical image analysis, vol. 42, pp. 60–88, 2017

    2017

  2. [2]

    U-net: Convolutional networks for biomedical image segmentation,

    O. Ronneberger, P. Fischer, and T. Brox, “U-net: Convolutional networks for biomedical image segmentation,” in Medical image computing and computer-assisted intervention–MICCAI 2015: 18th international con- ference, Munich, Germany, October 5-9, 2015, proceedings, part III 18 . Springer, 2015, pp. 234–241

    2015

  3. [3]

    nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,

    F. Isensee, P. F. Jaeger, S. A. Kohl, J. Petersen, and K. H. Maier-Hein, “nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,” Nature methods, vol. 18, no. 2, pp. 203–211, 2021

    2021

  4. [4]

    Unetr: Transformers for 3d medical image segmentation,

    A. Hatamizadeh, Y . Tang, V . Nath, D. Yang, A. Myronenko, B. Land- man, H. R. Roth, and D. Xu, “Unetr: Transformers for 3d medical image segmentation,” in Proceedings of the IEEE/CVF winter conference on applications of computer vision , 2022, pp. 574–584

    work page 2022

  5. [5]

    Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images,

    A. Hatamizadeh, V . Nath, Y . Tang, D. Yang, H. R. Roth, and D. Xu, “Swin unetr: Swin transformers for semantic segmentation of brain tumors in mri images,” in International MICCAI brainlesion workshop . Springer, 2021, pp. 272–284

    2021

  6. [6]

    Machine learning with multi-site imaging data: An empirical study on the impact of scanner effects,

    B. Glocker, R. Robinson, D. C. Castro, Q. Dou, and E. Konukoglu, “Machine learning with multi-site imaging data: An empirical study on the impact of scanner effects,” arXiv preprint arXiv:1910.04597 , 2019

    arXiv 1910

  7. [7]

    Domain generalization for medical image analysis: A review,

    J. S. Yoon, K. Oh, Y . Shin, M. A. Mazurowski, and H.-I. Suk, “Domain generalization for medical image analysis: A review,” Proceedings of the IEEE, 2024

    2024

  8. [8]

    Proto- typical pseudo label denoising and target structure learning for domain adaptive semantic segmentation,

    P. Zhang, B. Zhang, T. Zhang, D. Chen, Y . Wang, and F. Wen, “Proto- typical pseudo label denoising and target structure learning for domain adaptive semantic segmentation,” in Proceedings of the IEEE/CVF con- ference on computer vision and pattern recognition , 2021, pp. 12 414– 12 424

    work page 2021

Show all 32 references
  1. [9]

    Self-attentive spatial adaptive normalization for cross-modality domain adaptation,

    D. Tomar, M. Lortkipanidze, G. Vray, B. Bozorgtabar, and J.-P. Thiran, “Self-attentive spatial adaptive normalization for cross-modality domain adaptation,” IEEE transactions on medical imaging , vol. 40, no. 10, pp. 2926–2938, 2021

    2021

  2. [10]

    Where and how to transfer: Knowledge aggregation-induced transferability perception for unsupervised domain adaptation,

    J. Dong, Y . Cong, G. Sun, Z. Fang, and Z. Ding, “Where and how to transfer: Knowledge aggregation-induced transferability perception for unsupervised domain adaptation,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 3, pp. 1664–1681, 2021

    2021

  3. [11]

    Shape-aware meta-learning for generalizing prostate mri segmentation to unseen domains,

    Q. Liu, Q. Dou, and P.-A. Heng, “Shape-aware meta-learning for generalizing prostate mri segmentation to unseen domains,” in Interna- tional conference on medical image computing and computer-assisted intervention. Springer, 2020, pp. 475–485

    2020

  4. [12]

    Causality-inspired single-source domain generalization for medical image segmentation,

    C. Ouyang, C. Chen, S. Li, Z. Li, C. Qin, W. Bai, and D. Rueckert, “Causality-inspired single-source domain generalization for medical image segmentation,” IEEE Transactions on Medical Imaging , vol. 42, no. 4, pp. 1095–1106, 2022

    2022

  5. [13]

    Bayeseg: Bayesian modeling for medical image segmentation with interpretable generalizability,

    S. Gao, H. Zhou, Y . Gao, and X. Zhuang, “Bayeseg: Bayesian modeling for medical image segmentation with interpretable generalizability,” Medical Image Analysis , vol. 89, p. 102889, 2023

    2023

  6. [14]

    Improving cross- domain generalizability of medical image segmentation using uncer- tainty and shape-aware continual test-time domain adaptation,

    J. Zhu, B. Bolsterlee, Y . Song, and E. Meijering, “Improving cross- domain generalizability of medical image segmentation using uncer- tainty and shape-aware continual test-time domain adaptation,” Medical Image Analysis, vol. 101, p. 103422, 2025

    2025

  7. [15]

    Synthseg: Segmentation of brain mri scans of any contrast and resolution without retraining,

    B. Billot, D. N. Greve, O. Puonti, A. Thielscher, K. Van Leemput, B. Fischl, A. V . Dalca, J. E. Iglesias et al. , “Synthseg: Segmentation of brain mri scans of any contrast and resolution without retraining,” Medical image analysis , vol. 86, p. 102789, 2023

    2023

  8. [16]

    Randstainna: Learning stain- agnostic features from histology slides by bridging stain augmentation and normalization,

    Y . Shen, Y . Luo, D. Shen, and J. Ke, “Randstainna: Learning stain- agnostic features from histology slides by bridging stain augmentation and normalization,” in International Conference on Medical Image Computing and Computer-Assisted Intervention . Springer, 2022, pp. 212–221

    2022

  9. [17]

    Test-time image- to-image translation ensembling improves out-of-distribution generaliza- tion in histopathology,

    M. Scalbert, M. Vakalopoulou, and F. Couzini ´e-Devy, “Test-time image- to-image translation ensembling improves out-of-distribution generaliza- tion in histopathology,” in International Conference on Medical Image Computing and Computer-Assisted Intervention . Springer, 2022,...

    2022

  10. [18]

    Mutual information-based disentangled neural net- works for classifying unseen categories in different domains: Application to fetal ultrasound imaging,

    Q. Meng, J. Matthew, V . A. Zimmer, A. Gomez, D. F. Lloyd, D. Rueck- ert, and B. Kainz, “Mutual information-based disentangled neural net- works for classifying unseen categories in different domains: Application to fetal ultrasound imaging,” IEEE transactions on medical imagi...

    2020

  11. [19]

    Mi- segnet: Mutual information-based us segmentation for unseen domain generalization,

    Y . Bi, Z. Jiang, R. Clarenbach, R. Ghotbi, A. Karlas, and N. Navab, “Mi- segnet: Mutual information-based us segmentation for unseen domain generalization,” in International Conference on Medical Image Comput- ing and Computer-Assisted Intervention . Springer, 2023, pp. 130–140

    2023

  12. [20]

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

    Z. Zhou, L. Qi, X. Yang, D. Ni, and Y . Shi, “Generalizable cross- modality medical image segmentation via style augmentation and dual normalization,” in Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition , 2022, pp. 20 856–20 865

    2022

  13. [21]

    Treasure in distribution: A domain randomization based multi-source domain generalization for 2d medical image segmentation,

    Z. Chen, Y . Pan, Y . Ye, H. Cui, and Y . Xia, “Treasure in distribution: A domain randomization based multi-source domain generalization for 2d medical image segmentation,” in International Conference on Medical Image Computing and Computer-Assisted Intervention. Springer, 20...

    2023

  14. [22]

    Domain generalization with correlated style uncertainty,

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

    2024

  15. [23]

    Deep feature interpolation for image content changes,

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

    2017

  16. [24]

    Regularizing deep networks with semantic data augmentation,

    Y . Wang, G. Huang, S. Song, X. Pan, Y . Xia, and C. Wu, “Regularizing deep networks with semantic data augmentation,” IEEE Transactions on Pattern Analysis and Machine Intelligence , vol. 44, no. 7, pp. 3733– 3748, 2021

    2021

  17. [25]

    Bsda: Bayesian random semantic data augmentation for medical image classification,

    Y . Zhu, X. Cai, X. Wang, X. Chen, Z. Fu, and Y . Yao, “Bsda: Bayesian random semantic data augmentation for medical image classification,” Sensors, vol. 24, no. 23, p. 7511, 2024

    2024

  18. [26]

    Dofe: Domain-oriented feature embedding for generalizable fundus image seg- mentation on unseen datasets,

    S. Wang, L. Yu, K. Li, X. Yang, C.-W. Fu, and P.-A. Heng, “Dofe: Domain-oriented feature embedding for generalizable fundus image seg- mentation on unseen datasets,” IEEE Transactions on Medical Imaging , vol. 39, no. 12, pp. 4237–4248, 2020

    2020

  19. [27]

    Learning generalized medical image segmentation from decoupled feature queries,

    Q. Bi, J. Yi, H. Zheng, W. Ji, Y . Huang, Y . Li, and Y . Zheng, “Learning generalized medical image segmentation from decoupled feature queries,” in Proceedings of the AAAI conference on artificial intelligence, vol. 38, no. 2, 2024, pp. 810–818

    2024

  20. [28]

    A comprehensive retinal image dataset for the assessment of glaucoma from the optic nerve head analysis,

    J. Sivaswamy, S. Krishnadas, A. Chakravarty, G. Joshi, A. S. Tabish et al. , “A comprehensive retinal image dataset for the assessment of glaucoma from the optic nerve head analysis,” JSM Biomedical Imaging Data Papers, vol. 2, no. 1, p. 1004, 2015

    2015

  21. [29]

    Rim-one: An open retinal image database for optic nerve evaluation,

    F. Fumero, S. Alay ´on, J. L. Sanchez, J. Sigut, and M. Gonzalez- Hernandez, “Rim-one: An open retinal image database for optic nerve evaluation,” in 2011 24th international symposium on computer-based medical systems (CBMS) . IEEE, 2011, pp. 1–6

    2011

  22. [30]

    Refuge challenge: A unified framework for evaluating automated methods for glaucoma assessment from fundus photographs,

    J. I. Orlando, H. Fu, J. B. Breda, K. Van Keer, D. R. Bathula, A. Diaz- Pinto, R. Fang, P.-A. Heng, J. Kim, J. Lee et al. , “Refuge challenge: A unified framework for evaluating automated methods for glaucoma assessment from fundus photographs,” Medical image analysis, vol. 59...

    2020

  23. [31]

    Domain generalization with mixstyle,

    K. Zhou, Y . Yang, Y . Qiao, and T. Xiang, “Domain generalization with mixstyle,” arXiv preprint arXiv:2104.02008 , 2021

    2021 arXiv

  24. [32]

    Exact feature distribution matching for arbitrary style transfer and domain generalization,

    Y . Zhang, M. Li, R. Li, K. Jia, and L. Zhang, “Exact feature distribution matching for arbitrary style transfer and domain generalization,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 8035–8045

    2022

Pith tools

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