Abstract

In medical image analysis, significant challenges arise from domain shifts. Models trained on one dataset often struggle to generalize to unseen domains, limiting their clinical utility. To overcome this challenge, recent advancements have tried to increase the diversity of training data with data augmentation, in which the augmentation rules are pre-set before training commences and remain unchanged throughout the training process. Previous methods do not augment according to the unique characteristics of individual samples. As a result, they fail to cover the full diversity of unseen domains. To tackle this problem, we propose a learnable framework, the Adaptive Augmentation Framework (ADA), which can adaptively augment data catering to each individual sample. It has three operators for different purposes: 1) the Learnable Bezier Remap operator dynamically adjusts parameters to do the augmentation according to its content features. 2) the Channel Shift Control operator dynamically tunes shift and scale parameters for each color channel. By capturing fine-grained variations and improving spectral detail representation. 3) The Gradient-guided Feature Weaken operator dynamically reduces the influence of high-impact features to improve the model’s ability to generalize. Extensive experiments conducted on seven medical segmentation datasets demonstrate that adaptive augmentation is more likely to cover large diversity in the unseen domain.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/0315_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

N/A

BibTex

@InProceedings{HuaRun_ADA_MICCAI2025,
        author = { Huang, Runlin and Cai, Hongmin and Zhuo, Weipeng and Cai, Shangyan and Lin, Haowei and Fan, Wentao and Su, Weifeng},
        title = { { ADA: An Adaptive Augmentation Framework for Single-Source Domain Generalization in Medical Image Segmentation } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15969},
        month = {September},
        page = {43 -- 52}
}


Reviews

Review #1

  • Please describe the contribution of the paper

    The paper proposed a new method called the Adaptive Augmentation Framework, which augments training data based on the unique characteristics of individual samples to improve the model’s generalization. The authors conducted extensive experiments to validate the effectiveness of their method.

  • 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.

    They propose a novel augmentation method which can be adaptive to individual samples. The method enhance generalization by controlling image augmentation, such as color transformation, with learnable parameters, thus improving the ability of domain transfer

  • 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, too few references 2, confusing notations, e.g., in section 2.1 and 2.2, b refers to batch index or bias? 3, incomplete paper 4, no visualization results

  • Please rate the clarity and organization of this paper

    Poor

  • 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 does not provide sufficient information for reproducibility.

  • 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/2025/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.

    (1) Strong Reject — must be rejected due to major flaws

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The paper appears incomplete, with missing explanations or sections that are underdeveloped, affecting the overall clarity 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.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A



Review #2

  • Please describe the contribution of the paper

    In this paper, the authors propose an Adaptive Augmentation Framework (ADA) for medical image segmentation, which consists of three key components designed for dynamic image augmentation: (1) the Learnable Bézier Remap operator for adaptive intensity transformation, (2) the Channel Shift Control operator for per-channel scaling and shifting, and (3) the Gradient-guided Feature Weaken operator for regularizing feature learning. Extensive experimental results demonstrate that the proposed framework consistently outperforms existing state-of-the-art methods on various medical image segmentation tasks.

  • 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.

    The proposed three operators are well-motivated and grounded in theory. These operators are designed to be model-agnostic and can be seamlessly integrated with various backbone architectures. The experimental results further validate the effectiveness of these operators, demonstrating consistent performance improvements across different medical image segmentation tasks.

  • 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.

    While the proposed method is well-designed, there are several weaknesses that could be improved:

    1. The paper lacks a detailed explanation of the overall training pipeline, particularly how the learnable parameters within each proposed operator are jointly optimized during training.

    2. The experiments are limited to segmentation tasks and primarily rely on DeepLabv3+ as the backbone. Evaluating the proposed framework with other popular architectures, such as Swin Transformer-based models, would strengthen the generality and applicability of the method.

    3. The ablation study shows that the Gradient-guided Feature Weaken operator has relatively minor impact when combined with only one of the other two operators, but achieves good performance when all three operators are used together. This observation is interesting, but the paper lacks sufficient analysis or discussion to explain this phenomenon.

    4. The complexity of the Bézier curve is determined empirically based on experimental results from specific datasets. This heuristic choice may limit the generalizability of the framework across different tasks or 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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/2025/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 — could be accepted, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    The proposed three operators are well-motivated and grounded in theory. These operators are designed to be model-agnostic and can be seamlessly integrated with various backbone architectures. The experimental results, although only on segmentation, but shows superiority performance over existing techniques. The author need to update the manuscript to include the whole training procedure.

  • 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.

    The author’s feedback addressed and resolved my concerns.



Review #3

  • Please describe the contribution of the paper

    The main contribution of this paper is the proposal of ADA (Adaptive Augmentation), a method designed to achieve robust segmentation performance on medical images, even when applied to previously unseen datasets from different domains. The paper also presents the implementation of a dedicated operator to realize ADA, along with comprehensive evaluations across a wide range of 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.

    One major strength of the paper is the introduction of a dynamic augmentation strategy, which addresses the limitations of traditional static augmentation methods used in medical image segmentation. This dynamic approach adaptively adjusts augmentation parameters based on the training process, allowing the model to generalize better to unseen domains. Furthermore, the paper presents a mathematically grounded implementation of a dedicated operator to realize this adaptive mechanism, adding both theoretical rigor and practical value to the proposed method.

  • 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.

    One potential weakness of the paper is that the performance improvement achieved by the proposed method is relatively modest—around 1% to 2% in most cases. While this gain is consistent and meaningful in the context of domain generalization for medical image segmentation, some may argue that it does not represent a dramatic improvement in accuracy.

  • 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 authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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/2025/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.

    (6) Strong Accept — must be accepted due to excellence

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?

    I recommend a strong accept because the paper presents a completely novel approach and successfully implements it based on a solid mathematical formulation.

  • Reviewer confidence

    Confident but not absolutely certain (3)

  • [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.

    N/A

  • [Post rebuttal] Please justify your final decision from above.

    N/A




Author Feedback

We sincerely thank all reviewers for their valuable feedback and recognition of our contributions, including: “the method is well-motivated and well-designed” (Reviewer #2), “the paper presents a completely novel approach” (Reviewer #3), and “propose a novel augmentation method” (Reviewer #1).

@Reviewer #1 Regarding Comments 7.3 and 7.4: “no visualization results” and “incomplete paper” We respectfully clarify that our manuscript contains extensive visualizations. For instance, Fig. 3 compares predicted masks of ADA and competing methods across three datasets. Fig. 4 shows t-SNE plots of target pixel features, illustrating that ADA better separates background, optic disc, and optic cup features, compared to other methods.
The clarity and organization of this paper were also positively noted by Reviewer #2 and Reviewer #3.

Regarding Comment 7.1: “too few references” Owing to the strict page limitations of the conference, we prioritized citing the most relevant and foundational works. We fully acknowledge the importance of comprehensive citation and will consider including additional related works in the extended version.

Regarding Comment 7.2: “notational confusion” We apologize for any confusion. The subscript ‘b’ denotes batch size, while the other symbols refer to bias terms. We will clarify this notation in the revised manuscript to avoid ambiguity.

@Reviewer #2 Thank you for your insightful and encouraging feedback. Regarding Comment 7.1: “how the learnable parameters within each proposed operator are jointly optimized during training” During training, gradients from the downstream task are back-propagated to both the Learnable Bézier Remap Operator and the Channel Shift Control Operator. The input image is first remapped in each channel by the Bézier operator, and the resulting distribution is further adjusted by the shift control operator. This entire process is end-to-end trainable, with both operators learning jointly via backpropagation to adaptively optimize their parameters.

Regarding Comment 7.2:“backbone constraints” Our method is designed to be plug-and-play and agnostic to the backbone architecture. While we used DeepLabv3+ for consistency and fair comparison with prior works, our operators can be seamlessly integrated with other backbones, such as Swin Transformer, Resnet and etc.

Regarding Comment 7.3: “ablation study” This is indeed an important observation. When used independently, the Bézier and Channel Shift Control operators tend to reach a performance plateau due to overfitting. To mitigate this, we introduce the Gradient-Guided Feature Weaken Operator, which serves to inject regularization noise and improve generalization. As such, its main contribution is not to improve accuracy directly, but to prevent overfitting when it is used in conjunction with the other two operators.

Regarding Comment 7.4:“the complexity of the Bézier curve” We appreciate this valuable insight. In the context of medical imaging (including RGB and grayscale images), we found that cubic Bézier curves strike a good balance between flexibility and stability. However, we agree that in other domains, exploring higher-order or alternative parametric curves may yield further benefits—a promising direction for future research.

@Reviewer #3 We sincerely thank you for your positive comments and recognition. Regarding Comment 7.1: We thank the reviewer for the comment. While the gains on certain datasets may appear modest, they are meaningful given the challenges of medical image segmentation under domain generalization, where cross-domain variations (e.g., scanners, patient populations, anatomical regions) are substantial. Our method consistently outperforms strong baselines across seven datasets and achieves robust results on class generalization, few-shot, and organ generalization tasks, demonstrating its practical value and strong generalization ability in real-world clinical scenarios.




Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • If your recommendation is “Provisional Reject”, then summarize the factors that went into this decision. In case you deviate from the reviewers’ recommendations, explain in detail the reasons why. You do not need to provide a justification for a recommendation of “Provisional Accept” or “Invite for Rebuttal”.

    N/A

  • 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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    N/A



Meta-review #2

  • After you have reviewed the rebuttal and updated reviews, please provide your recommendation based on all reviews and the authors’ rebuttal.

    Reject

  • Please justify your recommendation. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’
    1. Insufficient comparison with the SOTA single-domain generalization methods. Only papers accepted in 2020 and 2023 were used for comparison.
    2. Missing the deep analysis of the dynamic data augmentation.
    3. Missing statistically significant analysis.
    4. It is important to report the performance when using more datasets for training as the upper bound.



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. You may optionally write justifications for ‘accepts’, but are expected to write a justification for ‘rejects’

    Only one reviewer gave a strong rejection, while the other two reviewers recommended acceptance (one strong accept, one weak accept). Upon review, the justification for the rejection appears to be too general and lacks specific, detailed critiques or constructive suggestions. In contrast, the positive reviews highlight clear strengths in the paper. Given the overall balance of reviewer opinions and the limited substantiation behind the negative assessment, I recommend acceptance.



back to top