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

Abstract

Breast cancer remains a leading cause of mortality worldwide and is typically detected via screening programs where healthy people are invited in regular intervals. Automated risk prediction approaches have the potential to improve this process by facilitating dynamically screening of high-risk groups. While most models focus solely on the most recent screening, there is growing interest in exploiting temporal information to capture evolving trends in breast tissue, as inspired by clinical practice. Early methods typically relied on two time steps, and although recent efforts have extended this to multiple time steps using Transformer architectures, challenges remain in fully harnessing the rich temporal dynamics inherent in longitudinal imaging data. In this work, we propose to instead leverage Vision Mamba RNN (VMRNN) with a state-space model (SSM) and LSTM-like memory mechanisms to effectively capture nuanced trends in breast tissue evolution. To further enhance our approach, we incorporate an asymmetry module that utilizes a Spatial Asymmetry Detector (SAD) and Longitudinal Asymmetry Tracker (LAT) to identify clinically relevant bilateral differences. This integrated framework demonstrates notable improvements in predicting cancer onset, especially for the more challenging high-density breast cases and achieves superior performance at extended time points (years four and five), highlighting its potential to advance early breast cancer recognition and enable more personalized screening strategies. Our code is available at https://github.com/Mortal-Suen/VMRA-MaR.git.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/1274_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/Mortal-Suen/VMRA-MaR

Link to the Dataset(s)

CSAW-CC dataset: https://researchdata.se/en/catalogue/dataset/2021-204-1

BibTex

@InProceedings{SunZij_VMRAMaR_MICCAI2025,
        author = { Sun, Zijun and Thrun, Solveig and Kampffmeyer, Michael},
        title = { { VMRA-MaR: An Asymmetry-Aware Temporal Framework for Longitudinal Breast Cancer Risk Prediction } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15974},
        month = {September},
        page = {666 -- 676}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The authors integrated a Vision Mamba Recurrent Neural Network (VMRNN) with a Spatial Asymmetry Detector (SAD) and Longitudinal Asymmetry Tracker (LAT) to predict cancer onset, with a specific focus on patients with three ranges of breast density (low, medium and high). The proposed asymmetry-aware temporal framework, termed VMRA-MaR, demonstrated enhanced predictive capabilities relative to the state-of-the-art Longitudinal Mammogram Risk (LoMaR) prediction model, particularly at extended longitudinal time points (years four and five). Quantitatively, VMRA-MaR achieved a C-index of 0.82 and Receiver Operating Characteristic Area Under the Curve (ROCAUC) scores of 0.84 at both year four and year five. These results suggest the potential of VMRA-MaR to facilitate earlier breast cancer detection, thereby enabling more tailored and individualized screening protocols. The primary objective was to synergistically combine temporal encoding (VMRNN) with the spatial asymmetry dynamics of breast tissue. Although VMRA-MaR exhibited an improvement in cancer onset prediction compared to LoMaR, the statistical significance of this improvement was marginal (p = 0.061). In summary, this study introduces a novel deep learning (DL) methodology for predicting cancer onset using longitudinal mammograms, which holds potential for substantial clinical impact pending further validation of statistical significance.

  • 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.

    Some of the strength of the study include:

    • Utilization of a Recurrent Neural Network (RNN) architecture explicitly designed to address inherent limitations in modeling the dynamic evolution of breast tissue over longitudinal timeframes. This temporal modeling capability allows for a more nuanced understanding of tissue changes preceding cancer onset.
    • Furthermore, the integration of bilateral symmetry analysis, building upon the AsymMirai framework, into the risk prediction process represents a significant advancement. This approach accounts for contralateral asymmetry within an end-to-end learning pipeline, enabling a more comprehensive assessment of spatial risk factors.
    • Finally, the utilization of multi-year sequential mammography data contributes to improved risk prediction accuracy, particularly in patients with high breast density. This approach supports the earlier detection of breast cancer by leveraging longitudinal data, facilitating timely intervention and improved patient outcomes.
  • 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.

    Some of the weaknesses of the study include:

    • The inherent complexity of the VMRNN, a DL model, poses challenges in interpretability, potentially obscuring a comprehensive understanding of its decision-making processes. This lack of transparency can limit the ability to identify specific features or patterns driving predictions.
    • While mammographic asymmetry can serve as a potential indicator, its reliability as a standalone method for breast cancer risk assessment remains limited. Existing evidence suggests a weak association between asymmetry in mammographic density and short-to-medium term breast cancer diagnosis risk. The marginal performance improvement of VMRA-MaR over LoMaR, despite incorporating breast asymmetry dynamics, may stem from this inherent limitation.
    • The dataset utilized in this study, namely the CSAW-CC dataset from Karolinska University Hospital, lacks detailed characterization regarding data heterogeneity, class balance, and pathological profiles. These factors, known to influence the performance of DL frameworks, warrant further clarification to ensure robust model validation and generalizability.
    • The study provides an inadequate explanation of the loss function’s behavior. A comprehensive understanding is essential to determine whether the observed performance decline in low-density cases during early time points is attributable to sample bias. Similarly, a justification is required to elucidate the enhanced performance in high-density cases at later time points, thereby clarifying the factors contributing to improved accuracy in these scenarios.
    • Given the potential for class imbalance within the dataset, the weighting process within the LSTM may be skewed, leading to biased predictions. Justification is needed to demonstrate that the model effectively mitigates the impact of class imbalance on the weighting of different time points, ensuring unbiased and accurate risk assessment.
    • The absence of a comprehensive spatial assessment of the entire breast represents a missed opportunity, particularly in light of growing evidence suggesting that regions beyond the annotated lesion contain predictive information for cancer risk. A more holistic spatial analysis could potentially enhance the model’s sensitivity and specificity.
  • 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.

  • 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

    Although the manuscript exhibits a high degree of overall clarity, opportunities exist for further refinement. Enhanced characterization of the study cohort through the inclusion of more detailed demographic and clinical information would facilitate a more comprehensive understanding of the study population. Furthermore, the incorporation of visual aids, such as diagrams or flowcharts depicting the image aggregation process, could significantly improve comprehension of the methodological approach. Visual representations can provide a more intuitive understanding of complex data processing pipelines. While the authors promote reproducibility by utilizing open-source data and providing code availability on GitHub, the manuscript lacks sufficient detail regarding critical parameters, dataset characteristics, and evaluation metrics. The inclusion of this information is essential for ensuring the reproducibility and validation of the study’s findings by independent researchers. Some of the ways to improve in this paper include providing:

    • Detailed specifications concerning model parameters, dataset composition, and evaluation methodologies are required to ensure transparency and reproducibility. This includes providing specific information on parameter settings, dataset demographics and characteristics, and the metrics used to assess model performance.
    • A clear rationale is needed to justify the exclusive use of lesion annotations, given the availability of complete mammographic images for analysis. The justification should address the potential benefits and drawbacks of focusing solely on lesion annotations compared to incorporating information from the entire mammogram.
    • A comprehensive discussion of the study’s limitations, which is essential for a balanced and rigorous scientific evaluation. This discussion should include, but not be limited to, considerations regarding the applicability of the method to patients with unilateral mastectomy, incomplete imaging views, breast implants, or other factors that may limit the generalizability of the findings. The absence of statistically significant differences, despite improvements in the DL framework, warrants further investigation into potential factors such as sample size limitations or confounding variables, necessitating a more in-depth analysis of the framework’s workflow. This analysis should be complemented by a detailed description of the study cohort, including the distribution of cases across density subgroups, as well as a justification of the loss function’s behavior across different density levels and time points. Without these comprehensive details regarding the framework’s operation and cohort characteristics, it remains challenging to ascertain the true value and generalizability of the findings, particularly given the lack of statistical significance, thereby highlighting the need for further investigation and detailed reporting to address these limitations.
  • 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?

    While the study exhibits several methodological limitations, it represents a significant contribution that could be enhanced through future refinement. The absence of detailed specifications regarding the framework architecture, parameter settings, cohort characteristics, and validation procedures currently obscures the value of the findings, especially given the marginal performance gains relative to existing, less complex state-of-the-art methods. Addressing these weaknesses could potentially yield substantial improvements in the overall quality and impact of the work. Therefore, my recommendation/rating is a 4.

  • 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 VMRA-MaR, a novel architecture for breast cancer risk prediction. The model combines:

    1. Vision Mamba RNN (VMRNN), used to capture longitudinal changes over time
    2. An asymmetry tracking module, which incorporates the known correlation between asymmetry and cancer risk and tracks it longitudinally.
  • 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. While the individual components are not entirely new, the integration of longitudinal modeling with asymmetry detection is novel and compelling.
    2. The use of VMRNN could offer a more efficient alternative to Transformers, with a smaller memory footprint.
    3. The authors include a subgroup analysis by breast density, showing that their model performs particularly well in high-density cases, which is an important and challenging subgroup with both higher cancer risk and lower mammographic readability.
  • 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. While VMRNN is a promising component, the overall architecture comprising an encoder, aggregator, VMRNN, and asymmetry module, may be overly complex.
    2. The comparison with LoMaR shows only marginal improvements. The claim that Mamba is better at capturing longitudinal changes would benefit from a more thorough analysis. Furthermore, the reported p-value of 0.061 is not statistically significant, which weakens their core claim.
    3. I would also like to see how LoMaR would perform if it was extended with the same asymmetry detection and tracking module. A fairer comparison would involve augmenting LoMaR with the same asymmetry mechanism. Especially, considering the modest gain of VMR_MaR without the asymmetry module.
  • 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.

  • 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
    • Minor typo in section 2.2 line 3 (integreat)
    • The future work would benefit from a head-to-head comparison with LoMaR + asymmetry tracking to isolate the contribution of VMRNN.
    • Consider a complexity vs. performance analysis that could strengthen your hypothesis.
    • The authors should also consider providing stronger statistical evidence to support the claim of surpassing prior state-of-the-art methods. That said, even if the overall p-value is not statistically significant, enough proof of the model’s performance gains in certain subgroups and its reduced computational footprint may still offer practical value in specific clinical settings.
  • 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 idea of combining longitudinal analysis and asymmetry makes sense and has been discussed in the literature before. However, I feel the core claim, that the Vision Mamba RNN captures dynamic longitudinal changes better than Transformers and other SOTA models, wasn’t supported strongly enough by the current results. The improvements are not clearly significant, and the comparisons could be more rigorous. That said, the direction is meaningful, and the approach is interesting,

  • 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 #3

  • Please describe the contribution of the paper

    This paper presents VMRA-MaR, a novel framework for breast cancer risk prediction that leverages both longitudinal imaging data and bilateral asymmetry in mammograms. The model uses a recent Vision Mamba-based RNN (VMRNN) to capture long-term temporal changes in breast tissue and integrates an asymmetry-aware module to track persistent left-right differences over time. The proposed approach demonstrates improved performance compared to state-of-the-art models, especially for patients with dense breast tissue. The framework is evaluated on the public CSAW-CC dataset, and the code is publicly available.

  • 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 paper addresses an important clinical challenge, improving long-term breast cancer risk prediction using longitudinal mammograms, which is consistent with real-world screening practices.

    The study employs a recent architecture (Vision Mamba) instead of traditional transformer-based models to capture dynamic tissue changes, showing substantial improvement. Importantly, asymmetry information is considered, which is often ignored in current risk prediction studies.

    The results show that the proposed model achieves strong performance, outperforming current state-of-the-art methods (e.g., Mirai, LoMaR), particularly in dense breast cases. The inclusion of Grad-CAM visualizations help support model interpretability.

  • 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.

    The related work section is lacking, especially in explaining the technical motivation for using the Vision Mamba RNN.

    The paper would benefit from clearer descriptions of experimental settings. For example: How does the model handle patients with fewer than five time points? Is there any analysis showing the influence of the number of time points on model performance? How were breast density levels defined? The paper uses the libra_densearea metric, but why not use percent density, which may better reflect true density?

    In Table 1, it’s unclear if the test sets are the same across methods. Also, the reported C-index for AsymMirai is higher than OncoNet, despite lower AUCs, which seems inconsistent. There may be a formatting or numerical error in Table 1: the 5-year AUC for VMR_MaR* is listed as 0.86 (0.84–0.86), which is a very narrow confidence interval.

  • 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 submission has provided an anonymized link to the source code, dataset, or any other dependencies.

  • 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?

    This paper proposes an innovative approach that combines temporal modeling with asymmetry-aware analysis for breast cancer risk prediction. The improvements in predictive accuracy, especially for high-density breasts and long-term predictions, are promising. Some experimental details and technical explanations could be improved to strengthen the overall clarity and quality of the paper.

  • 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



Author Feedback

We thank all the reviewers for their insightful and valuable comments. We are pleased that the reviewers recognized the novelty and strengths of our work: the innovative integration of longitudinal temporal modeling with bilateral asymmetry analysis (R1, R2, R3), the significant performance in challenging high-density breast cases (R1, R2, R3), and the strong results surpassing current state-of-the-art methods (R3).

  1. Regarding the handling of class imbalance and loss function behavior (R1): Thank you for pointing out that this was not explicitly stated in the submission. We follow LoMaR [1], and leverage a weighted cross-entropy loss to tackle the imbalance problem and ensure accurate risk assessment. The class weights are computed based on the training set distribution. We will clarify this in the final version.

  2. Regarding comparison with an augmented model (R1 & R2): As the proposed asymmetry module is backbone agnostic it would indeed be interesting to see it also being applied to LoMaR. While we do believe that it can improve performance also for the LoMaR backbone, based on our results that demonstrate that our VMR_MaR framework outperforms LoMaR, overall performance is expected to be lower.

  3. Regarding the definition of breast density (R3): We utilize the absolute quantity of dense breast tissue (libra_densearea) in this work as it has been shown that absolute dense tissue tends to provide more relevant information on breast cancer risk compared to percentage density (see for instance [2]).

  4. Regarding the consistency of metrics in Table 1 (R3): The reviewer noted an apparent inconsistency where the reported C-index for AsymMirai is higher than OncoNet, despite AsymMirai having lower AUCs. Note, this occurs because the C-index is a global concordance measure that evaluates concordance over all pairs of patients and all follow-up times, taking censoring into account. In contrast, the time-dependent AUC at a specific time point (e.g., 2 years) only assesses how well the model distinguishes between those who experienced the event by that time and those who were still event-free at that time. Consequently, a model can demonstrate better performance in correctly ordering patients’ risks over the entire study duration (resulting in a higher C-index) even if its discrimination ability at each fixed landmark time is weaker.

Overall, we are grateful for the constructive feedback of all the reviewers and will incorporate these clarifications in the final version. As acknowledged by the reviewers, our approach, which combines temporal modeling with asymmetry-aware analysis, presents a significant advancement in breast cancer risk prediction using longitudinal mammograms, achieving promising results and demonstrating potential for clinical impact.

[1]Longitudinal Mammogram Risk Prediction. MICCAI, 2024. [2]Using mammographic density to predict breast cancer risk: dense area or percentage dense area. Breast Cancer Res. 2010;12(6):R97.



Meta-Review

Meta-review #1

  • Your recommendation

    Provisional Accept

  • 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



back to top