Paper Info Reviews Author Feedback Meta-Review Back to top
List of Papers Browse by Subject Areas Author List

Abstract

Extending deep learning models to out-of-distribution (o.o.d.) data remains a persistent challenge, especially in domains like medical imaging with restricted data availability and limited data sharing. This challenge is particularly evident in pulmonary nodule detection, as the model struggles to distinguish nodules from the surrounding normal tissues across different data distributions. To address this issue, we propose a Distributionally Regularized Mamba Network (DRMNet). Inspired by Mamba, we propose a Feature-Augmented State-Space module that unifies pulmonary nodule features to effectively distinguish nodules from surrounding confounding tissues. Furthermore, a Region-Aware Distribution Alignment module is elaborately introduced to reduce disparities in feature distributions between domains. We construct a pulmonary nodule detection dataset, named Generalization for Pulmonary Nodule Detection (GPND), comprising diverse domains, including private and well-known public datasets. Extensive experiments conducted on GPND demonstrate that DRMNet outperforms state-of-the-art domain generalization methods. The code is available at \url{https://github.com/TzhongBoyyy97/DRMNet}.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1719_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/TzhongBoyyy97/DRMNet

Link to the Dataset(s)

https://github.com/TzhongBoyyy97/DRMNet

BibTex

@InProceedings{LanTia_Domain_MICCAI2025,
        author = { Lan, Tianzhong AND Chen, Nan AND Yi, Zhang AND Xu, Xiuyuan AND Zhu, Min},
        title = { { Domain Generalization for Pulmonary Nodule Detection via Distributionally-Regularized Mamba } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15965},
        month = {September},
        page = {152 -- 162}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper proposed DRMNet as a novel framework for domain generalization (DG) in pulmonary nodule detection. The authors address the challenge of out-of-distribution (OOD) generalization by integrating two key components: FASS Module and RADA Module.

  • 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 integration of Mamba for long-range context modeling is novel in pulmonary nodule detection, effectively distinguishing nodules from vascular structures. RADA’s dual mechanism (KL divergence + low-rank regularization) mitigates feature distribution shifts without overfitting.

  • 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 definitions of loss functions in Eq.1 are unclear.
    2. The detailed demographic of experimental datasets is missing.
    3. The justification for KL divergence as the optimal alignment metric is underdeveloped.
  • Please rate the clarity and organization of this paper

    Satisfactory

  • 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.

    (3) Weak Reject — could be rejected, dependent on rebuttal

  • Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
    1. The authors are recommended to compare KL divergence with alternative distribution metrics to validate design choices.
    2. While PONSD and GGO are described as “desensitized,” nodule annotation protocols and inter-rater reliability metrics are omitted.
    3. The definitions of loss functions in Eq.1 are unclear.
  • 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

    This paper proposes a U-Net based network, for lung nodule detection, performance on 4 datasets achieves SOTA.

  • 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. FASS module was proposed to incorporate with U-Net, to extract global features.
    2. KL divergence was integrated in to extract more features.
    3. SOTA performance.
  • 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 author seems to be exaggerating the contribution of the RADA module. The core idea seems to be the use of KL divergence. Please elaborate on this point.
    2. In the visualization results, only solid nodules and ground glass nodules are shown. This method does not seem to be suitable for higher-risk nodules such as cavitary nodules and semi-solid nodules. Although the overall effect seems good, it may be difficult to promote in practical applications
  • 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.

    (3) Weak Reject — could be rejected, dependent on rebuttal

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

    See strengths and weaknesses.

  • 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 authors have almost addressed my concerns.

    Although there are still some minor flaws, I think the paper should be accepted



Review #3

  • Please describe the contribution of the paper

    1) propose the FASS module to incorporate global features by leveraging the spatial relationship between pulmonary nodules and vascular structures to integrate global features. 2) introduce the RADA module further to align pulmonary nodule features from different source domains, enabling effective generalization to the target domain. 3) construct a domain generalization dataset GPND for pulmonary nodule detection. It contains two private datasets and two public 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) Domain generalization is important for real-world use of CAD systems. This topic is very relevant.

    2) Authors did a good job comparing DRMNet to other DG methods and Mamba-based models. The proposed method had the best performance. There was also an ablation study to demonstrate the importance of individual component in the method.

    3) The experiment setup was thorough to show the importance of DG when training and evaluation data have very different features.

  • 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) Besides DG and other Mamba methods, it would be nice to also compare against results using other AI methods such as transformer-based or CNN-based lung nodule detection. 2) If possible, I would like to see the statistical significance of the improvement, especially when comparing against other Mamba methods since the improvement was relatively minor.

  • 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.

    (5) Accept — should be accepted, independent of rebuttal

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

    Important topic, thorough experiment setups (although I would like to see comparison with CNN and transformer results as well), results showed improvement (would be stronger if statistical significance was shown)

  • 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.

    Authors have properly addressed my concerns.



Author Feedback

We thank the reviewers for their positive and constructive comments. They highlighted “the proposed DRMNet as a novel framework” (R1), “the topic is important and very relevant”, and “the experiment setup was thorough” (R2). Here we address the main points in their reviews.

Response to Reviewer #1: Q1: KL Divergence Justification Our core contribution is the concept of RADA (joint distribution alignment via latent space regularization). Specifically, KL divergence, a widely used domain discrepancy measurement, is selected to achieve this. The key reasons can be summarized as: (1) It theoretically aligns with minimizing domain shifts via an upper bound (Eq. 3); (2) It is computationally efficient; (3) In the revised Table 4, we show KL outperforms the other two metrics [a, b] (+2.43% FROC and +3.01% FROC, respectively).

Q2: Dataset Details Revised Sec.3.1 now includes: (1) Demographic statistics (case counts, nodule types) via charts; (2) Inter-rater metrics [c, d] (PONSD: c=0.87, d=0.90; GGO: c=0.85, d=0.88), these values meet the reliability standards for medical image annotation; (3) Due to space constraints, the complete three-tier annotation workflow is provided in our GitHub.

Q3: Loss Function Clarification Eq.1 is refined in Sec.2.2: L_cls: Focal Loss (γ=2); L_reg: Smooth L1 Loss (Faster R-CNN encoding); L_rank: SVD-based low-rank regularization, as described in original Sec.2.2; KL term: Constrains features to N(0,1), derived in original Sec.2.4 and implemented via torch.nn.KLDivLoss.

We hope these revisions address your concerns and kindly request the opportunity to publish a corrected version.

Response to Reviewer #2: Q1: Expanding Comparisons to Transformer-Based Methods The original Table 1 compared DRMNet with CNN-based methods (e.g., LDDG, DGER). As suggested, we have included two transformer-based baselines, e and f, in the revised Table 1. DRMNet achieves 4.1% and 3.3% higher average FROC than these methods, respectively. This demonstrates the superiority of our Mamba-based architecture over both CNN and transformer paradigms. Q2: Statistical Significance Validation To validate the performance improvements, we add paired t-test results in the revised Table 3. The results confirm that DRMNet gains over all baselines are statistically significant.

Your suggestions have strengthened the completeness of our evaluation. We appreciate your constructive input.

Response to Reviewer #3: Q1: Clarification on RADA Module Contributions The RADA module combines KL divergence (for domain alignment) and low-rank (LR) regularization (for noise suppression in features). As shown in the original Table 2, removing LR leads to a 2.9% drop in average FROC, while removing KL causes a 2.7% drop in average FROC. These results highlight that both components contribute distinctively to improving generalization, with LR mitigating feature redundancy and KL enforcing domain-invariant distributions.

Q2: Visualization Results We appreciate your attention to clinical relevance. The second row in the original Fig. 3 corresponds to the semi-solid nodule you mentioned. Although both the PONSD and PN9 datasets contain cavitary nodules, we did not present them in the original Fig. 3. This is because cavitary nodules are relatively rare in the overall dataset, and there remains controversy regarding whether cavitary nodules should be considered high-risk [g,h]. In the revised Fig. 3, we add a cavitary nodule (PONSD case #34) and another semi-solid nodule (GGO case #26).

We sincerely hope these revisions meet your expectations and kindly request the opportunity to publish this corrected version.

[a] 10.1609/aaai.v32i1.11784 [b] 10.5555/2188385.2188410 [c] 10.1109/FUZZY.2010.5584447 [d] 10.1186/s12874-018-0550-6 [e] 10.48550/arXiv.2504.19574 [f] 10.48550/arXiv.2010.04159 [g] 10.1128/CMR.00060-07 [h] 10.1183/09031936.96.09102017



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.

    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 #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’

    After checking the first reviewers’ concerns and the rebuttal, I agree with the reviewer that there are still some minor flaws, but it may be acceptable.



back to top