List of Papers Browse by Subject Areas Author List
Abstract
Accurate breast cancer risk prediction from screening mammography is critical for enabling personalized screening intervals and early detection. Recent deep learning methods have shown the value of longitudinal data and explicit temporal alignment. However, existing approaches either perform explicit alignment using a single mammographic view or model multiple views without explicit longitudinal alignment, limiting their ability to exploit the complementary spatial–temporal information used in clinical practice.
To address this gap, we propose LMV-Net, a longitudinal multi-view breast cancer risk prediction model that jointly analyzes anatomically complementary CC and MLO views within an explicitly aligned longitudinal framework. We evaluate our approach on the public EMBED and CSAW-CC datasets, comparing it to state-of-the-art breast cancer risk prediction methods. Our model consistently outperforms existing approaches in overall risk prediction performance and across different breast density and cancer subgroups. Importantly, these improvements highlight the potential of longitudinal multi-view modeling to enhance risk stratification, paving the way for future work on personalized screening, earlier identification of high-risk patients, and more efficient screening resource allocation. The code is publicly available \href{https://github.com/sot176/LMV-Net}{here}.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2399_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/LMV-Net
Link to the Dataset(s)
EMBED dataset: https://aws.amazon.com/marketplace/pp/prodview-unw4li5rkivs2#overview%7D
CSAW-CC dataset: https://researchdata.se/en/catalogue/dataset/2021-204-1
BibTex
@InProceedings{ThrSol_Longitudinal_MICCAI2026,
author = { Thrun, Solveig AND Sun, Zijun AND Salahuddin, Suaiba A. AND Wickstrøm, Kristoffer AND Wetzer, Elisabeth AND Hansen, Stine AND Jenssen, Robert AND Kampffmeyer, Michael},
title = { { Longitudinal Multi-view Modeling for Breast Cancer Risk Prediction } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16895},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
The paper proposes LMV-Net, a DL framework for breast cancer risk prediction from longitudinal screening mammography that operates at the individual breast level and integrates both standard views (CC and MLO) across multiple time points. A pretrained frozen MammoRegNet estimates per-view deformation fields that register prior to current exams (applied in feature space to align longitudinal feature maps within each view). Current, aligned-prior, and temporal-subtraction features are concatenated channel-wise per view and refined by a Dual Stream Attention block combining multi-head self-attention within each view with cross-attention between CC and MLO streams. A Mirai-initialised ResNet-18 serves as the shared encoder. Baselines include Mirai, VMRA-MaR, OA-BreaCR, and ImgFeatAlign.
- 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.Evaluation spans two large, independent public datasets (EMBED and CSAW-CC) with different populations, vendors, and density-labelling schemes (BI-RADS vs LIBRA tertiles). 2.Statistical discipline is above average: 95% bootstrap CIs with 1000 resamples, a paired non-parametric bootstrap test against the second-best baseline, patient-level splits with MammoRegNet trained on a disjoint cohort to prevent leakage, and realistic ~5% positive rate rather than artificial balancing. 3.Reporting both frozen and fine-tuned LMV-Net variants allows architectural gains to be separated from encoder-adaptation gains. 4.On EMBED, the gains over strong baselines are large and statistically significant. Fine-tuned LMV-Net achieves C-index 81.4 vs 74.7 for the next-best method. The two fine-tuned baselines (VMRA-MaR, OA-BreaCR) are compared under matched encoder conditions. 5.The seven-variant ablation is informative and cleanly isolates several contributions. 6.The density-stratified analysis is a clinically meaningful robustness check
- 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 novelty claims are overstated against the paper’s own reference list. The abstract describes the work as “the first breast-wise longitudinal multi-view risk prediction framework that explicitly aligns the feature maps of each CC and MLO view across multiple time points”, yet the cited prior work already covers essentially every property in this conjunction. The alignment mechanism itself is explicitly imported from MammoRegNet (cited). The abstract should attribute the alignment to MammoRegNet and remove “the first”. 2.The Dual Stream Attention block, positioned as a novel contribution, is a standard two-stream cross-attention transformer encoder. Its specification is a conventional transformer recipe with well-established precedents. The cross-attention between CC and MLO is also geometrically naive: CC and MLO are non-coplanar projections, but the block treats their feature-map positions as interchangeable tokens with no projection-aware positional encoding. 3.The ablation does not include a single-stream transformer baseline operating on the concatenation of CC and MLO tokens with matched parameter count and layer depth. Variant (iv) (“same view both branches”) removes complementary information entirely and is not an equivalent substitute, because a concatenated-token single-stream transformer has access to both views through self-attention over the joint sequence. Without this row, the contribution of the two-stream design cannot be empirically isolated from extra capacity plus access to the complementary view. 4.The improvements are strongly asymmetric between datasets and this is not discussed. On EMBED the fine-tuned C-index gain is 6.7 points; on CSAW-CC only 1.7 points, and the 5-year gain is 0.7 points and not statistically significant. The single-view ablation costs 10.6 C-index points on EMBED but only 0.9 on CSAW-CC, suggesting the value of multi-view modelling is substantially dataset-dependent. The paper currently presents the gains as uniform; the discussion should report the asymmetry explicitly. 5.ImgFeatAlign, the baseline most conceptually similar to LMV-Net (both use MammoRegNet-derived alignment), is reported only with a frozen encoder. The headline comparison therefore conflates architectural with encoder-adaptation gains. The frozen-vs-frozen comparison permitted by the table shows a 3.6-point C-index gap on EMBED and approximately parity on CSAW-CC. This architectural-only gap should be reported explicitly. 6.Table 1 shows 29 positive breast-level exams at 5 years and 34 at 4 years. Bootstrap CIs on AUC from ~30 positives are wide enough that differences below ~0.05 are not reliably distinguishable from noise. Given the ~5% positive rate, the paper would also benefit from PR-AUC or sensitivity at fixed specificity alongside AUC.
- 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
Ambiguity int the training-schedule description. The text states models were “trained for 30 epochs” and also that “training stopped after 15 epochs, based on validation C-index”. The intended meaning is presumably a 30-epoch maximum with early stopping after 15 stagnant epochs; this should be clarified.
Evaluation constructs “all consecutive time-point pairs” from patients with more than two visits. For the paired bootstrap significance test to be valid, resampling should be clustered at the patient level rather than the pair level, since multiple pairs from the same patient are not independent. Do confirm the clustering level of the bootstrap.
- 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?
This is a competent integration paper with a strong empirical story on EMBED and a more modest story on CSAW-CC, backed by above-average statistical discipline and a thorough ablation. Main reasons for weak accept are the 6.7 C-index gain over fine-tuned baselines on EMBED with paired bootstrap significance across all horizons, the use of two large public datasets with realistic class imbalance, patient-level splits with explicit leakage prevention, the informative seven-variant ablation, and honest reporting of both frozen and fine-tuned LMV-Net variants. The main reasons for not rating higher are that the contribution is fundamentally one of integration rather than methodological innovation and the paper overclaims this in the abstract: the alignment mechanism is explicitly borrowed from cited MammoRegNet, the breast-wise longitudinal multi-view setting is already covered by cited prior work, and the Dual Stream Attention block is a standard two-stream cross-attention transformer encoder with no architecturally novel element.
- 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 addresses the majority of concerns raised. The re-computation using patient-level clustered bootstrap confirms that significance claims are robust to the clustering choice, previously significant results remain significant, which was the most serious methodological concern raised. The clarification on the frozen-encoder comparison is the most consequential new information: frozen LMV-Net outperforms frozen ImgFeatAlign by +3.6 C-index points on EMBED and +3.0 on CSAW-CC, both statistically significant, establishing a real architectural contribution independent of encoder fine-tuning and resolving the concern about conflated gains. The commitment to revising the abstract and removing the “first” framing is noted and sufficient for the camera-ready. The dataset asymmetry explanation (EMBED being multi-population and multi-scanner versus CSAW-CC being single-scanner and more homogeneous) is credible and should be stated explicitly in the final version. The concatenated-token single-stream transformer baseline remains absent and the response does not fully justify why variant (iv) is an adequate substitute; this is a residual weakness. On balance, the rebuttal strengthens confidence in the empirical story sufficiently to move the rating from 4 to 5.
Review #2
- Please describe the contribution of the paper
This paper proposed a breast cancer risk prediction model (LMV-Net) based on longitudinal mammograms of both CC and MLO views. By leveraging an existing registration model, the features of the current mammography, aligned prior features, and the difference features between the current and prior were used for risk prediction. A dual stream attention-based model was used to predict multi-label classification at multiple time points.
- 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 inclusion of features from both current, aligned prior, and differences is novel. 2.The paper is validated on two public datasets. Both datasets show superior performance of the proposed method comparing to state-of-the-art methods. 3.The documentation is pretty comprehensive with detailed description about the methods and model parameters (e.g., model size).
- 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.Could it be possible that LMV-Net predicts the risk at middle years (say year 3) to be higher than the risk at later years (say year 5)? It is unclear how the authors handle the first increasing then decreasing risks. Simply summing them up may not be the best solution. 2.In Figure 2, what is the reason that the LMV-Net, tested on mammograms with higher breast density (density category D or high), shows higher C-Index on the EMBED dataset but lower C-Index on the CSAW-CC dataset?
- 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?
This paper introduces a novel approach to breast cancer risk assessment by integrating longitudinal features.
- 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 authors’ rebuttal has addressed my main concerns. Although the dual-stream attention mechanism itself is not entirely new, its application to longitudinal medical imaging research is novel.
Review #3
- Please describe the contribution of the paper
1.LMV-Net is proposed for multi-view, multi-time-point breast cancer risk prediction. 2.A Dual Stream Attention module is introduced to refine the longitudinal features of each view. 3.The method achieves strong performance on 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.The ablation study is comprehensive, and the comparisons with competing methods is fair. 2.The method section is relatively well organized, and the attention visualization serves as a helpful interpretability component.
- 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 novelty claim appears incremental, and the authors’ statement that they introduce the first breast-wise longitudinal multi-view risk prediction framework seems somewhat overstated, especially in light of prior work such as VMRA-MaR and related studies. 2.Figure 1 is visually overloaded and is difficult to interpret. 3.The paper presents several components using ‘novel’ terminology, but the actual innovation seems limited. In particular, Dual Stream Attention is already a common design in other domains, such as video modeling.
- 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.
- 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.
(2) Reject — should be rejected, independent of rebuttal
- Please justify your recommendation. What were the major factors that led you to your overall score for this paper?
1.The methodological innovation appears incremental. 2.The paper contains several overstated claims, with repeated use of terms such as “the first” and “novel,” which seem insufficiently supported. 3.As Figure 1 is intended to highlight the main contribution of the paper, it should be simplified, as the current version appears overly crowded.
- Reviewer confidence
Very confident (4)
- [Post rebuttal] After reading the authors’ rebuttal, please state your final opinion of the paper.
Reject
- [Post rebuttal] Please justify your final decision from above.
Novelty Concerns:
The main claim of this paper, being the first to jointly model multi-view and explicit longitudinal alignment, feels incremental to me. In essence, the authors take an existing registration-based alignment method (MammoRegNet) and add multi-view CC/MLO fusion on top. Both components exist in prior work; combining them is useful but not a substantial methodological contribution.
I also find the Dual Stream Attention block to be overstated as a contribution. Parallel self-attention and cross-attention across two streams is a well-known design pattern in multi-modal and multi-view learning, it has been used extensively in video understanding and vision-language pre-training. Calling this “novel” is a stretch.
Missing Comparisons:
I am not fully convinced by the experimental results. The paper omits comparisons with important longitudinal methods such as PRIME+ (arXiv 2303.15699, MICCAI 2023) and LoMaR (arXiv 2404.19083, MICCAI 2024), both of which use multiple prior time points in mammogram.
Author Feedback
We thank all reviewers for their constructive feedback. We are glad they appreciate our clinically meaningful and thorough analysis(R1), strong ablations(R1,R3), strong performance across datasets(R2,R3), our design and unified modeling approach(R2), and detailed description, ensuring reproducibility(R2).
(R1,R2,R3) Novelty/distinction from prior work We agree that the current phrasing in the abstract can be slightly misleading and will revise it and the introduction to more precisely highlight that the novelty comes from considering the multi-view information within the context of *explicit alignment. Prior approaches are either restricted to single-view modeling (OA-BreaCR, ImgFeatAlign) or support multi-view modeling without explicit longitudinal alignment, instead integrating temporal information implicitly (VMRA-MaR) or not at all (Mirai). In contrast, our method is the first to jointly model multi-view and explicit longitudinal alignment, resulting in substantial performance improvements over these prior approaches. This design is clinically motivated: radiologists interpret CC and MLO jointly when comparing timepoints, as they provide complementary diagnostic information. No prior method reflects this workflow in a longitudinal setting with explicit alignment.
*(R1) Dual Stream Attention We agree that a concatenated single-stream transformer could be another interesting baseline. However, note that this baseline would result in a single view-agnostic representation and thus also not completely allow DSA disentanglement, as our provided ablation study (vi) shows that multi-level learning (having losses both on the individual and joint risk prediction) is important. Regarding positional encoding, CC and MLO are non-coplanar with strong patient- and acquisition-dependent deformations, making fixed spatial correspondence ill-defined. We therefore avoid token-wise positional alignment and instead model complementary cross-view interactions. Projection-aware encoding remains future work.
*(R1,R2) Asymmetric performances EMBED is more diverse (multi-population, multi-scanner), while CSAW-CC is single-scanner and more homogeneous, explaining larger gains on EMBED due to better generalization. For density, EMBED’s density-D subgroup is smaller, while CSAW-CC densities were derived from LIBRA scores and grouped into equally sized categories. Although EMBED density-D shows higher mean performance, the wider confidence intervals indicate greater uncertainty and variability due to the smaller subgroup size. We will clarify this.
*(R1) Finetuning comparisons We first evaluated all methods using frozen encoders to ensure a fair comparison. Here, LMV-Net outperforms ImgFeatAlign on both datasets (+3.6% on EMBED and +3.0% C-index on CSAW). The C-index improvement is statistically significant for both datasets; the missing significance marker for frozen LMV-Net in Table 2 (CSAW) is a typo. We then evaluated fine-tuning both models which improved LMV-Net but not ImgFeatAlign. We agree that it would be beneficial to report this result.
*(R2) Cumulative risk prediction LMV-Net uses a cumulative probability layer (Eq. 1) enforcing monotonicity, such that risk is non-decreasing across years by design.
*(R1) Early stopping We will reformulate: LMV-Net was trained for up to 30 epochs with early stopping after 15 stagnant epochs.
*(R1) Bootstrap resampling and metrics Following prior SOTA work (Refs. 22, 23, 28), we adopt bootstrapping at the exam-pair level and AUC and C-index as the primary metric, ensuring fair comparison. We have now recomputed results using patient-level clustered bootstrap and see only negligible differences; previously significant differences remain significant. The similarity arises as the number of exams per patient is relatively homogeneous (mode: 3–4 exams), limiting the impact on the estimates. We will note additional metrics (PR-AUC, etc.) as future work.
*(R3) Fig. 1 We will improve readability of Fig. 1.
Meta-Review
Meta-review #1
- Your recommendation
Provisional Reject
- 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.
The paper addresses a clinically relevant problem and provides a reasonably strong empirical evaluation on two public mammography datasets. The overall evidence suggests that the system is potentially useful, especially given the public-data evaluation, realistic class imbalance, leakage-aware design, and ablation analysis. However, the main weakness is consistent across all reviewers: the novelty and framing are not sufficiently convincing. The claimed “first” contribution and repeated use of “novel” appear stronger than the actual technical advance, since key elements rely on known longitudinal alignment and standard two-stream attention designs. In addition, the empirical gain is not uniform across datasets, and the absence of a matched single-stream transformer baseline leaves the value of the dual-stream design insufficiently isolated.
The authors should directly revise or defend the novelty claims, especially the use of “first” and “novel,” and clarify the precise distinction from VMRA-MaR, ImgFeatAlign, MammoRegNet-based alignment, and standard two-stream cross-attention models. They should explain whether a matched single-stream transformer baseline can be added or, if not, why the current ablations are sufficient to isolate the contribution of Dual Stream Attention. They should report or discuss frozen-vs-frozen comparisons against ImgFeatAlign to separate architecture gains from encoder fine-tuning gains. They should explain the asymmetric performance between EMBED and CSAW-CC, including density-stratified differences. They should clarify whether cumulative risk predictions are monotonic across prediction years and how inconsistent year-wise risk outputs are handled. They should also clarify early stopping, bootstrap resampling at patient versus pair level, and consider adding PR-AUC or sensitivity at fixed specificity given the small number of positive cases in some horizons.
- 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.
Although the core architecture of dual-stream attention is well-documented in existing literature, pioneering its use within longitudinal medical imaging research presents a clear and valuable novelty.
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.
The paper addresses a clinically relevant problem and presents a well-evaluated framework for longitudinal breast cancer risk prediction using multi-view mammography. Reviewers generally agreed that the empirical evaluation is strong, with two public datasets, realistic class imbalance, leakage-aware experimental design, and comprehensive ablation studies. The main concerns were related to overstated novelty claims and the incremental nature of some components, particularly the use of explicit longitudinal alignment and the Dual Stream Attention design. The rebuttal appropriately clarified the intended contribution, softened the “first” and “novel” framing, and better distinguished the work from prior longitudinal and multi-view approaches. In addition, the authors clarified the frozen-versus-finetuned comparisons, cumulative risk monotonicity, and patient-level bootstrap analysis, which strengthened confidence in the reported empirical gains. While some limitations remain, particularly the absence of a matched single-stream transformer baseline, I believe the paper provides sufficient methodological and empirical contribution for acceptance.
Meta-review #3
- 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.
The rebuttal has not sufficiently addressed the main concerns raised by Reviewer #3, which is also agreed by the AC, that the paper overclaimed its novelty. The rebuttal mentioned that “Prior approaches are either restricted to single-view modeling (OA-BreaCR, ImgFeatAlign) or support multi-view modeling without explicit longitudinal alignment.” However, LoMaR (arXiv 2404.19083, MICCAI 2024) as pointed out by the reviewer indeed took both the multi-view and the longitudinal information into consideration. The explicit temporal alignment was conducted in step 2, using MammoRegNet, which is an existing work. The methodology contribution of the manuscript is hence confusing. Therefore, the paper in its current version does not meet the standard of MICCAI.
