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

Abstract

Regular mammography screening is essential for early breast cancer detection and deep learning-based risk prediction methods have sparked interest to adjust screening intervals for high-risk groups. While early methods focused only on current mammograms, recent approaches leverage the temporal aspect of screenings to track breast tissue changes over time, requiring spatial alignment across different time points. Two main strategies for this have emerged: explicit feature alignment through deformable registration and implicit learned alignment using techniques like transformers, with the former providing more control over the alignment. However, the optimal approach for explicit alignment in mammography remains underexplored. In this study, we provide insights into where explicit alignment should occur (input space vs. representation space) and if alignment and risk prediction should be jointly optimized. We demonstrate that jointly learning explicit alignment in representation space while optimizing risk estimation performance, as done in the current state-of-the-art approach, results in a trade-off between alignment quality and predictive performance and show that image-level alignment is superior to representation-level alignment, leading to better deformation field quality and enhanced risk prediction accuracy. The code is available at https://github.com/sot176/Longitudinal_Mammogram_Alignment.git.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2025/paper/3852_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/sot176/Longitudinal_Mammogram_Alignment

Link to the Dataset(s)

EMBED dataset: https://aws.amazon.com/marketplace/pp/prodview-unw4li5rkivs2#overview CSAW-CC dataset: https://researchdata.se/en/catalogue/dataset/2021-204-1

BibTex

@InProceedings{ThrSol_Reconsidering_MICCAI2025,
        author = { Thrun, Solveig and Hansen, Stine and Sun, Zijun and Blum, Nele and Salahuddin, Suaiba A. and Wickstrøm, Kristoffer and Wetzer, Elisabeth and Jenssen, Robert and Stille, Maik and Kampffmeyer, Michael},
        title = { { Reconsidering Explicit Longitudinal Mammography Alignment for Enhanced Breast Cancer Risk Prediction } },
        booktitle = {proceedings of Medical Image Computing and Computer Assisted Intervention -- MICCAI 2025},
        year = {2025},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 15961},
        month = {September},
        page = {494 -- 504}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    In this study, the investigators explored explicit alignment techniques to register mammography images across different time points (two in this study) in both feature space and image space and assessed their impact on breast cancer risk prediction. They introduced a deep learning-based method, MammoRegNet, to perform image-space alignment of longitudinal mammograms. The results indicated that image-space alignment outperformed feature-space alignment in terms of predictive performance.

    Breast cancer risk prediction is a critical component of women’s healthcare, as accurate and early predictions can significantly improve patient outcomes by enabling timely interventions. Leveraging longitudinal mammograms for risk prediction is particularly valuable, as they can capture subtle temporal changes in imaging biomarkers associated with breast cancer development.

  • 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.

    However, aligning longitudinal mammograms remains a major technical challenge due to the soft-tissue nature of the breast and variability in imaging protocols (e.g., differing compression levels and positioning). This study aimed to address key challenges in this domain and proposed a novel methodology to improve alignment. While the contributions are notable, there are several important limitations and areas for improvement.

  • 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- Lack of Multi-View Integration: Breast cancer risk prediction benefits significantly from using both standard mammographic views—craniocaudal (CC) and mediolateral oblique (MLO). The study does not describe any methodology for integrating both views into the prediction model, which limits the potential effectiveness of the approach.

    2- Lack of Benchmark Comparisons: Several established models for breast cancer risk prediction—such as MIRAI [1], LRP-NET [2], LoMaR [3], and Prime+ [4]—have been validated using large and independent datasets, including EMBED and CSAW (the same datasets used in this study). Some of these models use only single time-point images, while others incorporate longitudinal data. Despite this, the proposed model is not compared against any of these benchmarks, raising concerns about the validity and relative performance of MammoRegNet.

    3- Inappropriate Evaluation Metric for Image Registration: The registration quality of MammoRegNet was assessed using normalized cross-correlation (NCC), which may not be suitable for medical imaging tasks. In clinical settings, small and subtle features—such as early pathological signs—are critical. A high NCC score may indicate structural similarity but could obscure or distort diagnostically relevant information. For instance, in Figure 2, the registered prior image appears overly similar to the current image, suggesting potential loss of important prior characteristics.

    4- Missing Statistical Validation: The study lacks formal statistical tests (e.g., DeLong’s test) to compare AUCs across models or risk prediction time points. P-values are not reported, making it difficult to assess whether performance differences are statistically significant.

    5- Minor Comment – Clarification Needed in Figure 4(d): The figure caption states “AUC by weighting,” but it is unclear which specific model was used to generate AUCs for 1- to 5-year risk predictions.

    References:

    1. Yala, A., et al. (2021). Toward robust mammography-based models for breast cancer risk. Science Translational Medicine, 13(578), eaba4373.

    2. Dadsetan, S., Arefan, D., et al. (2022). Deep learning of longitudinal mammogram examinations for breast cancer risk prediction. Pattern Recognition, 132, 108919.

    3. Karaman, B.K., et al. (2024). Longitudinal mammogram risk prediction. In MICCAI. Springer.

    4. Lee, H., et al. (2023). Enhancing breast cancer risk prediction by incorporating prior images. In MICCAI. Springer.

  • 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

    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?

    Overall, the paper is well-organized and clearly written. The topic is compelling and addresses an important unmet clinical need—specifically, the alignment of mammography images across different time points. However, the lack of benchmark comparisons and the absence of appropriate evaluation metrics for the proposed methodology raise concerns about its practical applicability and generalizability.

  • 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

    The authors propose a method for longitunal mammographies alignment with the goal of risk prediction. The novelty of the method consists in the introduction of the explicit image regiistration using a deep neural network. The authors use two state-of-the-art dataset, perform quantitive assessment of risk prediction and evaluate the quality of the performed registration. The authors compare several methods and demonstrate the advantage of the proposed approach.

  • 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 is well written and the readings flows nicely. The method presentation if well structured and comprehensive.

    The proposed method is appealing as it efficiently intruduces the image-wise registration of 2D mammography with the goal of the future cancer development prediction.

    The experiements plan and the evaluation sound reasonable and quite complete to allow for the understanding of the contribution.

  • 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 proposed method relies a lot on the state-of the art work of Wang et al [24], referring to it multiple times. This creates a diffuculty of not having multuple details: network design, the dataset composition and requiring the reader to read the referenced paper. Moreover, there does not seem to be a direct comparison to the results from [24], creating confusion about the proposesd performances. That is, the performances presented in [24] are substantially higher, leaving the reader puzzled about the paper contribution.

    [24] https://doi.org/10.1007/978-3-031-72378-0_15

  • 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
    • I would like the authors to comment on the comparison of their work to the work of Wang et al. [24]
    • Could the authors precise whether the risk prediction is done per image (CC and MLO view separately) or per breast (CC and MLO views combined)?
    • Could the authors comment more on the registration as it may or may not involve irrealistic structural changes?
    • I’d like the authors to comment on the image size chosen. That is, the risk prediction is performed on a quite low (for mammography) resolution, that may negatively affect the performances.
    • I suggest the authors to revise the presentation of the Figure 1. That is, it might require time to understand that the risk prediction model 1e relates to all but 1a. A more clear separation could help.
    • Finally, on a minor note, I suggest the authors to revise the presentation of references in the numerical order.
  • 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 paper is a worthy conference material, however the lack of fair comparison to the prior work makes the understanding of the contribution more difficult than expected.

  • 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 explores the impact of explicit alignment between prior and current mammograms on breast cancer risk prediction. The authors examine several alignment strategies and find that image-level registration yields the best performance on test sets. Additionally, they adapt the Non-Iterative Coarse-to-Fine Transformer for 2D image alignment, applying it within the risk prediction pipeline

  • 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 paper addresses an understudied but important area, understanding how temporal alignment affects risk prediction performance. This could inform future research and model design.
    2. The authors conduct a rigorous evaluation, reporting both risk prediction metrics and registration performance, which helps contextualize the value of explicit alignment.
    3. The observation of trade-offs between risk prediction accuracy and deformation field quality is especially interesting. However, this phenomenon is only demonstrated on the EMBED dataset. It would strengthen the paper to confirm whether similar trends hold on the CSAW-CC dataset as well.
  • 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 overall risk prediction performance is lower than state-of-the-art methods, such as LoMaR and MIRAI, which leverage implicit alignment strategies. However, since the paper’s main focus is not on developing a new predictive model, but rather on isolating the effect of alignment, this limitation could be understandable.

  • 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

    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 appreciate studies that not only introduce a new or enhanced methodology but also offer guidance to the community on what works, what doesn’t, and why, helping to inform future research. This paper does exactly that. In addition to a well-motivated approach, the insights provided are valuable for the broader field.

  • 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 the reviewers for their insightful and constructive comments and are pleased that the reviewers recognized the novelty and strengths of our work: the comprehensive evaluation of different explicit alignment methods for longitudinal mammograms in breast cancer risk prediction (R1, R2, R3), the introduction of MammoRegNet for image-space alignment and its superior performance (R1, R2, R3), and the rigorous assessment of both risk prediction and registration quality (R3). We also appreciate the recognition of our method’s clinical relevanc (R1, R3).

1) Regarding multi-view integration and image-wise vs. breast-wise prediction (R1, R2): We acknowledge that integrating multiple views has the potential to improve risk prediction performance. At present, our risk prediction is performed per image, with the craniocaudal (CC) and mediolateral oblique (MLO) views analyzed separately. We recognize the advantages of combining both views into a unified prediction model, as this could enhance the accuracy and robustness of risk assessment. We plan to investigate this in future work.

2) Regarding the lack of benchmark comparisons and comparison to the work of Wang et al (R1, R2): Thank you for your thoughtful comment and for emphasizing the importance of benchmark comparisons. We would like to clarify that the primary focus of this paper is not the development of a new state-of-the-art predictive model, but rather to provide guidance to the community on the impact of alignment—what works, what doesn’t, and why. By focusing on alignment, our aim is to offer insights that can inform and strengthen future research. Given this objective, we prioritized evaluating alignment methods for a standard risk-model baseline. Note, that the FeatAlign baseline corresponds to our re-implementation of the Multi-time baseline also used by Wang et al. (see Table 2 in their paper). However, we acknowledge the value of such comparisons and plan to include them in future work to provide a more comprehensive assessment of our approach.

3) Regarding the evaluation metric for image registration (R1): Thank you for your feedback. In this study, we chose to use NCC, as it is one of the most commonly used metrics for evaluating medical image registration quality. However, we agree that exploring additional metrics would be valuable.

4) Regarding the image size (R2): Thank you for highlighting the importance of image resolution in risk prediction. In this study, we used an image size of 512×1024, following the approach used by Wang et al., to maintain consistency with their methodology. However, higher resolutions, such as 1664×2048—used in methods like MIRAI and LoMaR can preserve additional detail and potentially improve performance. Our preliminary results with this increased resolution indicate that while performance increases for all methods, the insights provided in this paper remain the same.

5) Improvements to presentation and clarity (R1, R2): Thank you for your feedback. We will improve the overall presentation of the paper, including refining figures, captions, and references, to enhance clarity and readability in the final version.

Overall, we sincerely appreciate the valuable feedback from the reviewers and will integrate their suggestions into the final version. As highlighted by the reviewers, our work represents a significant step forward in evaluating explicit alignment methods for longitudinal mammograms, introducing innovative techniques like MammoRegNet, and achieving superior performance in both risk prediction and registration quality.



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