List of Papers Browse by Subject Areas Author List
Abstract
Reliable clinical deployment of deep medical image models is hindered by distribution shifts across scanners, sites, and acquisition protocols. Existing domain generalization (DG) methods often focus on style or intensity diversification, but they can still leave networks dependent on domain-specific texture correlations. Inspired by evidence that Fourier phase encodes semantic structure, we introduce \textsc{PhaseAT}, a phase-aware adversarial training framework for medical DG. \textsc{PhaseAT} forms phase-perturbed training views in the Fourier domain by iteratively updating a bounded phase perturbation while keeping the amplitude spectrum unchanged, thereby stressing spatial organization under matched appearance statistics. Perturbations are applied only to the luminance channel in YCbCr color space to avoid chromatic artifacts. Additionally, a simple phase-saliency mask concentrates updates on the most influential frequencies. The model is trained with a weighted combination of losses on clean and phase-perturbed samples, supporting both single-source and multi-source DG. We validate our method on two challenging medical datasets and demonstrate that \textsc{PhaseAT} achieves over 20\% improvement in single-source domain generalization, outperforming several state-of-the-art DG methods. The code implementation is available at: \url{https://github.com/ahmed-sharshar/PhaseAT}
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3917_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/ahmed-sharshar/PhaseAT
Link to the Dataset(s)
N/A
BibTex
@InProceedings{ShaAhm_PhaseAT_MICCAI2026,
author = { Sharshar, Ahmed AND Hanif, Asif AND Kummari, Naveen Kumar AND Yaqub, Mohammad AND Guizani, Mohsen},
title = { { PhaseAT: Fourier Phase Adversarial Training for Medical Image Domain Generalization } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16881},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper introduces PhaseAT, a phase-aware adversarial training framework that operates in the Fourier domain to enhance the model’s generalization ablities. The core idea is to adversarially perturb the Fourier phase of the luminance channel while most previous works are primarily focused on the intensity distribution. Evaluation on two medical datasets shows that PhaseAT can achieve over 20% improvement under single source DG settings.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.Novel adversarial perturb on the fourier phase domain: This is one relatively novel method considering most current works are mainly focused on the intensity distribution or style diversification. Adversarial attack on the phase domain can provide one novel insight for enhancing the model’s robustness. 2.Strong single-source DG results with large margins: Table 1 show substantial and consistent improvements across all source domains on both datasets. 3.Comprehensive ablation study: Table 3(b) systematically isolates each design choice — phase-only vs. amplitude-only vs. both, with/without saliency mask, RGB vs. YCbCr vs. luminance-only, adversarial vs. random perturbation.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.Incomplete baseline comparison: Notable missing baselines include: (a) standard Fourier-based augmentation methods like FedDG (b) Traditional augmentation based methods like DeepAug 2.Evaluation limited to classification only — no dense prediction tasks: The paper claims PhaseAT is a general medical DG framework, yet only evaluates on patch-level classification while the absence of any segmentation benchmark significantly limits the demonstrated generality.
Minors: Figure 1: it seems the raw image and image after pertubation remain the same. Table 1: the author need to report one average performance on all domains.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
Overall this paper proposed one interesting novel adversarial based DG method for medical image analysis.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
The rebuttal partially addressed my concerns but I would prefer the author can still includes more comparision like suggested from R4.Overall I would like to hold my Weak Accept decision.
Review #2
- Please describe the contribution of the paper
The authors propose PhaseAT, an adversarial training scheme designed to improve robustness to domain shift through the generation of adversarial image examples during training.
To generate the adversarial samples, PhaseAT decomposes the image into its Fourier phase and amplitude components, and applies perturbations to the Fourier phase only. The adversarial example generated from this process is input into the model alongside the original image, and the model is trained to minimize the cross entropy loss on both images.
The authors evaluate their method in two image domains. The histopathology dataset Camelyon17-WILDS already has data from multiple acquisition sites. The authors also compile a diabetic retinopathy dataset from four existing datasets, each acting as a unique “site”.
PhaseAT is evaluated in single domain (the model is trained on one domain and tested on all others) and in multi-domain (model is tested on all domains except one and the left-out domain is used for testing) settings and compared with several recent examples from literature. Consistently, PhaseAT demonstrates strong performance compared to baselines.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
Clearly motivated and well-structured: The authors clearly motivate their paper with a detailed discussion of the background literature and related works. The paper is structured sensibly in a way that is easy to follow. Figures are all clear and easy to understand.
Commitment to reproducibility: the authors have included an anonymized link to their source code and extensive details regarding their implementation and experimental set up, as well as full details of their algorithm in pseudocode.
Robust experimental validation: the authors evaluate their method across both single-domain and multi-source domain generalization tasks, following similar processes to other literature. They also include robust validations investigating sensitivity across various hyper parameters, generalization across architectures, and the importance of the various components of their proposed methodology. Such exhaustive experimentation is encouraging to see.
Ablations provide clarity on the importance of the individual decisions made: Table 3 (b), in particular, investigates the contribution of each component of the proposed methodology, as well as a discussion as to why removing/changing each component leads to a drop in efficacy. This kind of interrogation of design choices in the method is really nice to see and provides a clear justification for all choices.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
Incorrect “best model” highlighted in Tables: Tables 1 and 2 both highlight the best performing model in bold text. Across all test cases, PhaseAT is highlighted as being the best, even when it is not. This is likely just a LaTEX formatting mistake. For example, in Table 1 (a) under source domain C4, PhaseAT isn’t the best performing model in 2 of the 4 test cases (C1 and C3) nor on average, yet it is highlighted in bold.
Missing statistical significance: Tables 1 and 2 report whether the best performing model shows a statistically significant performance improvement over other methods. Including statistical significance of the results would strengthen the claims around performance improvements as it would show that these are not simply due to noise in the results.
No examples of adversarial images: The paper doesn’t show an example of an adversarial image generated by PhaseAT. I appreciate that page limits are tight, and the perturbations are learned during training, but the inclusion of an example would help to illustrate the image features that the Fourier phase perturbation changes in an image.
Causes of differences in source domain performance are not discussed: Table 1, the chosen source domain has a significant impact on the model’s generalization (for example, with source domain C4 in the Camelyon1-WILDS dataset, the model struggles to generalize to all other domains). Do the authors have a theory for the reason behind this? Is this domain comprised of more challenging examples? Or does it represent a more extreme domain shift from the other sources?
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
In Table 3 (b) it would be nice to see a “baseline” row to compare to, to avoid comparing with models in other tables. This would help to make the value of each individual component even clearer.
For future work, it would be interesting to see how this approach might apply in contexts of shortcut learning, for example. This is mentioned in the introduction, and is out of the scope of this work, but I would be very interested to see how this approach might help address such problems.
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(4) Weak Accept — marginally above the acceptance threshold, but would not mind if rejected, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The authors have proposed an interesting and well motivated method for adversarial model training to improve generalization to new distributions. The paper is clearly motivated and well-written. I would have liked to see measures of statistical significance over other methods that the authors compare to, particularly in cases where the performance gains are only modest, as well an example of a perturbed image feature that highlights specifically what image features the proposed method alters.
- Reviewer confidence
Confident but not absolutely certain (3)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
After reading the authors’ rebuttal, I am pleased to see their commitment to addressing the issues raised by the other reviewers and me.
I am particularly pleased to see the reviewers confirm that they will include statistical significance tests in their results tables. As they mention in their review, the results demonstrate low variability across random seeds, but the gap between best and second-best is, on occasion, small. For example, the authors give the example in their rebuttal: “in Table 1(a) (Source C0), Phase-AT achieves 71.30±1.3 vs. 69.06±1.9 for RAS⁴DG”; these results are within a few percent of one another and show overlap in their deviations. The inclusion of, for example, t-tests can help show whether this improvement is significant.
Considering the author’s commitment to addressing the concerns and the relatively small scale of these corrections, I am happy to recommend acceptance.
However, I would like to comment on one point in the rebuttal. The authors claim: “The ‘Ours’ row was bolded for formatting consistency rather than to indicate per-cell superiority”. This is in direct contrast to the table captions in the paper, which explicitly state that the bold text highlights the best-performing model. Additionally, in Table 2, the authors appear to underline the worst-performing model (except in cases where they have incorrectly applied the bold text to their own model to indicate best performance). I would suggest adding text to the table captions to make this clear, and adding it to Table 1 as well for consistency.
Review #3
- Please describe the contribution of the paper
This paper proposes a phase-perturbed training method for medical domain generalization (DG). The approach aims to improve out-of-distribution generalization by introducing adversarial perturbations during training. Experiments are conducted on two medical image classification datasets.
- Please list the major strengths of the paper: you should highlight a novel formulation, an original way to use data, demonstration of clinical feasibility, a novel application, a particularly strong evaluation, or anything else that is a strong aspect of this work. Please provide details, for instance, if a method is novel, explain what aspect is novel and why this is interesting.
1.The overall narrative and logical flow of the paper are clear. 2.Research in medical domain generalization (DG) is of high practical importance.
- Please list the major weaknesses of the paper. Please provide details: for instance, if you state that a formulation, way of using data, demonstration of clinical feasibility, or application is not novel, then you must provide specific references to prior work.
1.The proposed phase perturbation as a data augmentation strategy contradicts prior insights and is internally inconsistent with the authors’ own claims. The paper states: “Since standard CNNs are biased toward texture cues [9], they frequently overfit to these unstable amplitude-driven correlations…”. If overfitting to amplitude-driven information is the issue, a more reasonable approach would be to introduce amplitude perturbations to improve robustness to cross-domain variations. Instead, the method perturbs the phase, which directly disrupts semantic structure. This may be unreliable, especially for medical tasks that heavily rely on structural information, such as segmentation and detection. 2.The novelty of the method is limited. Essentially, the approach applies Fourier-based data augmentation, which has already been widely explored in domain generalization literature. The only difference is that the perturbation is applied to the phase component rather than amplitude [1][2][3]. 3.The experimental scope is too limited. Medical image analysis typically involves more segmentation and detection tasks, and evaluating only classification datasets is insufficient to validate the effectiveness of the method. Broader experimental validation is needed. 4.Regarding the claim: “This not only improves computational efficiency but also produces more structured perturbations with fewer color artifacts,” the improvement in computational efficiency is unclear. What is the additional computational cost introduced by the learnable phase perturbation compared to the baseline?
References: [1] Xu Q, Zhang R, Zhang Y, et al. A Fourier-based framework for domain generalization. CVPR 2021.[2] Xu Q, Zhang R, Fan Z, et al. Fourier-based augmentation with applications to domain generalization. Pattern Recognition, 2023.[3] Wang X, Zhao Q, Zhang C, et al. Enhanced dynamic feature representation learning framework by Fourier transform for domain generalization. Information Sciences, 2023.
- Please rate the clarity and organization of this paper
Good
- Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.
The submission has provided an anonymized link to the source code, dataset, or any other dependencies.
- Based on your review and your understanding of the MICCAI Scientific Code of Ethics, do you believe this submission may involve a potential ethics concern or violation?
N/A
- Optional: If you have any additional comments to share with the authors, please provide them here. Please also refer to our Reviewer’s guide on what makes a good review and pay specific attention to the different assessment criteria for the different paper categories: https://conferences.miccai.org/2026/en/REVIEWER-GUIDELINES.html
N/A
- Rate the paper on a scale of 1-6, 6 being the strongest (6-4: accept; 3-1: reject). Please use the entire range of the distribution. Spreading the score helps create a distribution for decision-making.
(3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
The overall score is mainly influenced by the limited novelty of the proposed method and the insufficient experimental validation. The approach is closely related to existing Fourier-based augmentation techniques, with only a minor variation in perturbing the phase component. In addition, the theoretical motivation is not fully convincing, and there are inconsistencies in the reasoning behind the method design. The experimental evaluation is also limited in scope, lacking validation on more diverse medical tasks such as segmentation or detection.
These factors together weaken the overall contribution and impact of the work.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Accept
- [Post rebuttal] Please justify your final decision from above.
N/A
Author Feedback
We thank the reviewers and meta-reviewer. The reviews confirm PhaseAT’s key advantages: a novel, well-motivated DG formulation that moves beyond style perturbations to adversarial Fourier-phase perturbations; strong and consistent gains, especially in single-source DG; and robust validation via multi-domain experiments, ablations, clear presentation, and reproducibility. We address all concerns below:
Statistical Significance [R2, Meta] Tables 1–2 report mean ± std over 5 seeds, showing low variance and consistent performance across runs. The improvements over the strongest baselines are stable; e.g., in Table 1(a) (Source C0), Phase-AT achieves 71.30±1.3 vs. 69.06±1.9 for RAS⁴DG, and in Table 1(b) (Source Aptos), 44.99±1.9 vs. 41.50±2.3.We will report detailed paired t-test across the same 5 seeds and include 95% confidence intervals of the performance difference in the revision.
Computational Cost [R4, Meta] PhaseAT, like other adversarial training methods, uses K inner maximization steps and is therefore more expensive than ERM; our efficiency claim is only relative to adversarial variants perturbing all RGB or all YCbCr channels. By restricting perturbations to luminance (Y), PhaseAT reduces the inner-maximization dimensionality by 3×. We acknowledge this does not directly translate to wall-clock savings and add training time and memory in the revision.
Baselines (FedDG, DeepAug) [R1, Meta] Our evaluation covers three SOTA directions: style/intensity (AdverIN), disentanglement (ConDiSR), and Fourier-based methods (RAS⁴DG). FedDG and DeepAug are complementary but not directly aligned: FedDG focuses on cross-client amplitude mixing, while DeepAug uses pixel-space augmentations that emphasize texture cues. In contrast, PhaseAT operates in the Fourier domain with fixed amplitude and adversarial phase perturbations.
Dense Prediction [R1, R4, Meta] We thank the reviewers for the suggestion. PhaseAT is currently scoped to medical classification DG. Due to space constraints, dense prediction tasks are left as future work.
Source Domain Variability [R2] We agree that source-domain choice strongly affects SDG performance, as Camelyon17 centers differ in staining, scanners, and preparation, inducing different shortcut correlations. PhaseAT consistently improves over ERM across all sources, though gains vary (e.g., large for C2: 52.08→74.31, smaller for C4: 55.64→58.64), suggesting it is most effective when ERM relies more on domain-specific shortcuts. We will add this discussion and include a quantitative domain-distance analysis to better explain this variability.
Bold Formatting [R2] The “Ours” row was bolded for formatting consistency rather than to indicate per-cell superiority. In the revision, we will highlight only the best value in each column.
Adversarial Examples [R2] We appreciate the reviewers’ recognition of space constraints. We will make visualizations of adversarial examples available in the accompanying GitHub repository.
Phase vs Amplitude Perturbation [R4] PhaseAT complements, rather than replaces, amplitude-based DG by enforcing robustness under fixed amplitude (appearance) via small, bounded phase perturbations. To preserve semantics, perturbations are constrained, saliency-masked, and applied only to luminance. Table 3(b) shows amplitude-only (77.89%) and phase+amplitude (78.50%) underperform PhaseAT (85.7%), indicating constrained, gradient-guided phase perturbation is more effective in our setting.
Novelty [R4] PhaseAT’s novelty is not generic Fourier augmentation, but a constrained phase-adversarial objective: (i) inner maximization instead of stochastic mixing, (ii) phase-only perturbation with fixed amplitude, (iii) Wirtinger-style phase gradients, and (iv) saliency-masked, luminance-only perturbations for stability. Ablations show each is critical: random phase is weaker (81.90 vs. 85.7), removing the mask drops to 56.47, and RGB/YCbCr perturbations underperform.
Meta-Review
Meta-review #1
- Your recommendation
Invite for Rebuttal
- Please justify your decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. In case of an invitation for rebuttal, clarify which points are important to address in the rebuttal.
Reviewers (scores 4, 4, 3) overall acknowledge that the paper has merit but have several concerns that the authors are encouraged to address in a rebuttal where possible. In particular:
- statistical testing
- details on computational complexity / runtime
- general clarifications and a discussion of differences in source domain performances (R2)
Those seem addressable in a rebuttal.
Concerns centered around the selection of baseline methods and performance on dense prediction tasks are further shared concerns among reviewers, that however, are likely challenging to address without additional experiments.
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
All three reviewers are positive post-rebuttal suggesting to Accept this paper.
Reviewers appreciate the clarity and motivation of this work, as well as a comprehensive evaluation including ablation studies. Overall results and method are convincing. The authors submitted an effective rebuttal. Overall, I support the reviewers’ recommendation to accept this work.
Meta-review #2
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
PhaseAT introduces a novel adversarial training approach in the Fourier phase domain, moving beyond conventional amplitude based or style based domain generalization methods. The method is well validated through extensive ablations and multi domain experiments, showing consistent and substantial improvements. Reviewer #1 turned to Accept after rebuttal, and the theoretical concerns raised by Reviewer #4 were adequately addressed by ablation evidence (amplitude only 77.89% vs. PhaseAT 85.7%).
Meta-review #3
- After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.
Accept
- Please justify your recommendation.
The paper proposes PhaseAT, a phase-aware adversarial training framework for medical domain generalization. All three active reviewers converged on acceptance after the rebuttal, consistently recognizing the novelty and performance gains. The authors should prepare the final version to incorporate the promised significance tests, formatting corrections, and discussion on Fourier-based DG works.