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

Abstract

Breast cancer screening programmes increasingly seek to move from one-size-fits-all intervals to risk-adapted, personalised strategies. Deep learning has enabled image-based risk models that improve short- to medium-term (1–5 year) prediction over traditional risk models, but leading approaches such as Mirai typically rely on convolutional backbones, high-resolution inputs and relatively simple multi-view fusion, with limited explicit modelling of contralateral asymmetry. We hypothesised that combining complementary inductive biases (convolutional and transformer-based) with explicit contralateral modelling could match state-of-the-art 3-year risk prediction even at substantially lower input resolution. We present MamaDino, a Mammography-Aware Multi-view Attentional DINO-based model that fuses frozen self-supervised DINOv3 (ViT-S) features with a trainable CNN encoder at 512×512 resolution and aggregates left–right information via a BilateralMixer to output a 3-year breast cancer risk score. The 512×512 input size was chosen for computational feasibility, not as a preferred clinical resolution. We train on 53,883 women from OPTIMAM and evaluate on matched 3-year case-control cohorts: an in-distribution test set from four UK screening sites and an external out-of-distribution (OOD) set from an unseen site. At breast level, MamaDino was comparable to Mirai on in-distribution and OOD tests while using ~13x fewer input pixels. Adding the BilateralMixer yields an AUC of 0.736 (vs Mirai 0.713) in-distribution and 0.677 (vs 0.666) out-of-distribution with overlapping confidence intervals (CIs), with consistent performance across age, ethnicity, scanner, tumour type and grade. These results indicate that explicit contralateral reasoning plus complementary inductive biases can recover Mirai-level performance despite lower-resolution mammograms.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3605_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

N/A

Link to the Dataset(s)

https://medphys.royalsurrey.nhs.uk/omidb/

BibTex

@InProceedings{SanRug_MamaDino_MICCAI2026,
        author = { Santeramo, Ruggiero AND Zubarev, Igor AND Jug, Florian},
        title = { { MamaDino: Mammography-Aware Multi-view Attentional DINO for 3-Year Breast Cancer Risk Prediction } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16887},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    The paper presents MamaDino, a hybrid model for 3-year breast cancer risk prediction that fuses frozen DINOv3 (ViT-S) features with trainable SE-ResNeXt101 encoders via cross-attention, and aggregates bilateral breast information through a BilateralMixer module. The model operates at 512x512 resolution (~13x fewer input pixels than Mirai’s 1664x2048) and is trained on 53,883 women from OPTIMAM. Evaluation on matched case-control cohorts (in-distribution from 4 UK sites; OOD from an unseen site) shows that MamaDino matches or slightly exceeds Mirai’s AUC (0.736 vs. 0.713 in-distribution; 0.677 vs. 0.666 OOD), with consistent subgroup performance across age, ethnicity, scanner, tumour type and grade.

  • 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.Well-motivated architectural design. The combination of frozen self-supervised DINOv3 features (global semantics) with trainable SE-ResNeXt101 (local texture sensitivity) via cross-attention is principled and well-argued. The rationale that CNN tokens query ViT tokens to integrate globally coherent features into the texture-sensitive stream is architecturally elegant.

    2.BilateralMixer is a genuine contribution. The three-step bilateral fusion — symmetric CLS-based transformer encoding, gated asymmetry weighting via |eL−eR|, and symmetric feature composition (asymmetry + concordance) — explicitly encodes the clinical reasoning process of comparing left and right breasts. This is the most novel component and the ablation (0.736 vs. 0.731 for simple concatenation) confirms its value, albeit with modest gains.

    3.Rigorous evaluation design. Matched case-control test cohorts (in-distribution and OOD from an unseen site) with DeLong’s 95% CIs demonstrate proper evaluation methodology. The OOD cohort from Oxford with different scanner mix (including GE and Hologic) provides genuine external validation.

    4.Comprehensive subgroup analysis. Table 2 stratifies by age, ethnicity, scanner manufacturer, cancer type (DCIS vs. invasive), and tumour grade. This level of subgroup analysis is essential for clinical translation and exceeds what many risk prediction papers provide.

    5.Competitive performance at substantially lower resolution. Matching Mirai at ~13x fewer pixels is practically meaningful for computational deployment, memory constraints, and potential privacy-preserving applications where resolution reduction may be desirable.

    6.Thorough ablation study. Resolution ablations (224, 320, 512), single-stream ablations (DINO-only, SE-ResNeXt-only), and bilateral fusion ablation (concatenation vs. BilateralMixer) systematically isolate the contribution of each design choice.

    7.Honest limitations discussion. The authors transparently note that Mirai was evaluated without retraining (using its released model), UK-only data, under-representation of Siemens and non-White groups, and the inability to fully disentangle resolution vs. architecture effects. This level of honesty is commendable.

  • 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.Confidence intervals overlap on all key comparisons. The headline claims of “matching or exceeding Mirai” must be qualified: on the OOD set, MamaDino achieves 0.677 (0.647–0.707) vs. Mirai 0.666 (0.635–0.696). These CIs substantially overlap. On the in-distribution set, MamaDino’s 0.736 (0.710–0.762) vs. Mirai’s 0.713 (0.686–0.740) also overlap. While the point estimates favor MamaDino, statistical significance is not established. The paper should explicitly state this rather than framing results as “slightly outperforming. “

    2.Mirai comparison is not fully controlled. Mirai is run at its native resolution (1664x2048) using the released model, while MamaDino is trained on OPTIMAM data. The authors acknowledge this but it remains a substantial confounder: Mirai was trained on different data (US-based cohort) and uses clinical covariates that were unavailable here. A fairer comparison would retrain Mirai (or at least its CNN backbone) on the same OPTIMAM data at the same resolution. Without this, we cannot determine whether the gains come from architecture, training data advantage, or resolution sensitivity of Mirai’s specific weights.

    3.Ablation CIs not reported for baselines. Table 1 reports CIs only for Mirai and full MamaDino. The DINO-only, SE-ResNeXt-only, resolution ablations, and concatenation ablation lack CIs, making it impossible to assess whether these differences are meaningful. The BilateralMixer gain (0.736 vs. 0.731) is very small and could easily be within noise without CI.

    4.Limited calibration analysis. For a risk prediction model intended for clinical deployment, discrimination (AUC) alone is insufficient. No calibration analysis (e. g. , expected calibration error, calibration curves) is reported. In risk-adapted screening, the absolute risk probability matters for threshold-based decisions, not just ranking.

    5.Case-control sampling limits generalizability. The matched case-control design, while standard for validation studies, does not permit assessment of the model’s calibration or positive predictive value in a screening population. The 1:2 and 1:4 case-control ratios artificially inflate the case prevalence, and the OOD set composition (ethnicity largely unavailable) limits demographic generalizability assessment.

  • 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 does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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

    The per-channel augmentation strategy (Section 2.3: brightness, contrast, CLAHE as three channels) is creative, but the ablation is only briefly mentioned (“consistent improvements over naive channel replication”) without quantitative results. Including this ablation in a supplementary table would strengthen the claim. The Siemens scanner subgroup (N=27, 9 cases) yields AUC 0.463, which is below chance. While the sample is tiny, this should be explicitly flagged as a concern rather than folded into the general subgroup table without comment. The two-stage training (freeze encoder, train BilateralMixer) is pragmatic but introduces a question: would end-to-end fine-tuning of the entire pipeline improve performance? This could be discussed. For clinical audiences, it would be helpful to report sensitivity at clinically relevant specificity thresholds (e. g. , 90%, 95%) in addition to AUC.

  • 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?

    MamaDino presents a well-designed hybrid architecture with a principled BilateralMixer that explicitly encodes the clinical reasoning process of bilateral breast comparison — a genuine and technically interesting contribution. The evaluation is rigorous: matched case-control cohorts with internal and external validation, DeLong CIs, and comprehensive subgroup analysis across demographics, scanner vendors, and tumour characteristics. Performance matches or slightly exceeds Mirai at ~13x lower resolution, which is practically meaningful.

    The main concerns — overlapping CIs, uncontrolled Mirai comparison, missing calibration analysis — are limitations typical of the field and do not fundamentally undermine the contribution. The architectural design (frozen ViT + trainable CNN + cross-attention + bilateral fusion) is well-ablated and each component demonstrably adds value. The paper advances the state of the art in mammographic risk prediction methodology and has clear translational potential.

    Major factors for my positive recommendation: 1.Architecturally novel bilateral fusion mechanism with clinical motivation. 2.Competitive external validation at substantially lower computational cost. 3.Comprehensive, transparent evaluation with honest limitations discussion. 4.Practical significance for deploying risk models in resource-constrained settings.

  • Reviewer confidence

    Confident but not absolutely certain (3)

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

    N/A

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

    N/A



Review #2

  • Please describe the contribution of the paper

    This paper addresses mammogram-based short-term (3-year) breast cancer risk prediction. It is a clinically relevant problem with potential implications for improving current breast cancer screening strategies. The proposed framework combines frozen DINOv3 features, a trainable CNN-based encoder, and an explicit bilateral fusion module through a multistage training strategy. Using the OPTIMAM dataset, the authors develop and evaluate the model on both internal and external test sets.

  • 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.

    A major strength of the study is the inclusion of an external validation cohort, which increases confidence that the reported performance is not due to overfitting to the development dataset. The paper reports subgroup analyses across several clinically relevant groups, which is useful for understanding model behavior in different populations and settings. In addition, the study somehow considered several ablation-style comparisons, including DINO only, SE ResNeXt only, different input resolutions, and a different patient-level fusion.

  • 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.One major concern is that the paper does not provide a good / resonable motivation for using undersampled mammograms for risk prediction. However, the manuscript repeatedly emphasizes that the proposed model can achieve near SOTA performance despite using lower resolution inputs. I would recommend substantially revising the abstract and introduction to better justify this design choice and to tone down the corresponding claims. From a clinical perspective, higher resolution images are often needed to capture subtle imaging features. This may also help explain why MIRAI appears to perform more consistently in the DCIS subgroup, whereas the proposed method seems somewhat weaker in that setting. 2.In the comparison experiments, the DINO branch uses pretrained frozen weights only. It is not clear why the authors chose not to finetune DINO (the vit) for the breast cancer risk prediction task. The manuscript does not appear to provide a clear justification for this design choice. 3.The manuscript does not clearly describe the distribution of time to cancer in the study population. In addition, the test cohorts are constructed using a matched case-control design, which limits the clinical interpretation of the results. In particular, we don’t know the performance under a realistic screening prevalence setting. Or maybe the models have a higher false positive rate? 4.Although the manuscript provides good subgroup analyses, the overall workload of the experiments is still somewhat limited. The number of comparison methods is relatively small. At least more baselines are always welcome and helpful. For example, in Table 1, confidence intervals and OOD results are only reported for a subset of methods. In contrast, the ablation results are reported less comprehensively, without confidence intervals and without corresponding OOD evaluation. This somewhat reduces the overall quality.

  • 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 does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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.

    (3) Weak Reject — marginally below the acceptance threshold, but would not mind if accepted, dependent on rebuttal

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

    1.The motivation for the proposed problem setup and some of the architectural choices is not yet sufficiently clear, especially regarding the use of low-resolution mammograms and the decision to keep DINO frozen rather than finetune it. 2.Although the reported results are promising, they are not substantially stronger than the main baseline, and the comparative evaluation remains relatively limited.

  • 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.

    I thank the authors for their careful response to my concerns. The rebuttal clarifies several of the key issues I raised, including the motivation for using lower-resolution inputs, the decision to keep DINOv3 frozen, and the limitations of the MIRAI comparison and matched case-control evaluation. I also appreciate that the authors acknowledge these limitations and plan to temper the corresponding claims in the revised manuscript. Given the authors’ clarifications and the overall contribution of the work, I am willing to update my recommendation to Weak Accept.



Review #3

  • Please describe the contribution of the paper

    This paper introduces MamaDino, a multi view, bilateral breast cancer risk prediction model that combines convolutional and transformer based inductive biases to achieve Mirai level 3 year risk prediction at substantially lower input resolution. The work focuses on explicit contralateral modeling and efficient multi view fusion to support personalized, risk adapted screening.

  • 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 proposes a potentially interesting multi view and bilateral architecture, explicitly modeling contralateral asymmetry, a clinically meaningful inductive bias for mammography based risk prediction.
    • The combination of frozen self supervised DINOv3 features with a trainable CNN encoder reflects a thoughtful attempt to leverage complementary inductive biases (global contextual reasoning vs. local texture sensitivity).
    • The model is trained on a large-scale dataset (OPTIMAM, >53k women) and evaluated both in distribution and out of distribution, which strengthens confidence in generalization.
    • Performance is reported across multiple strata (age, ethnicity, scanner, tumor type and grade), which is a strong and commendable aspect rarely included in similar work.
    • Achieving performance comparable to Mirai at significantly reduced resolution is potentially impactful for scalability and deployment in real world screening settings.
  • 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 architectural description lacks clarity. Specifically, the sections “Backbone encoders and cross attention” and “BridgeMixer block”, together with Figures 1a and 1b, should be revised to more clearly illustrate data flow, view specific inputs, parallel processing, and fusion points.
    • The motivation for the cross attention fusion between DINOv3 and SE ResNeXt101 features is not fully justified, and it remains unclear what complementary information each branch uniquely contributes beyond empirical performance.
    • Dataset usage is ambiguous: while 53,883 women are reported for training, the exact composition of training, validation, and test cohorts is unclear. It is not explicitly stated whether in distribution test cases are fully held out from OPTIMAM or included in the reported total, nor how validation data were used. Ethical statment of the additional datasets apart from OPTIMAM are not included.
    • The training strategy is insufficiently described: it is unclear whether the SE ResNeXt101 view specific branches, the DINOv3-SEResNetXt101 cross attention module, and the BridgeMixer remaining blocks are trained jointly end to end or in separate stages.
    • Although the paper claims to assess inductive biases, the experimental design does not include a full resolution comparison analogous to Mirai’s setting, making it difficult to conclude whether inductive bias (rather than resolution reduction alone) drives the results.
    • No interpretability or explanation is provided for the predicted risk scores (e.g., saliency, attention maps, or bilateral importance), which is a notable limitation for a clinical risk prediction model.
    • Relevant prior work, such as AsymMirai and recent surveys on mammography based risk modeling, are not cited, limiting the contextualization and comparison of the contribution. Mendes J, Oliveira B, Araújo C, Galrão J, Mota AM, Garcia NC and Matela N (2025) Deep learning in breast cancer risk prediction: a review of recent applications in full-field digital mammography. Front. Oncol. 15:1656842.doi: 10.3389/fonc.2025.1656842 Donnelly J, Moffett L, Barnett AJ, Trivedi H, Schwartz F, Lo J, Rudin C. AsymMirai: Interpretable Mammography-based Deep Learning Model for 1-5-year Breast Cancer Risk Prediction. Radiology. 2024 Mar;310(3):e232780.doi: 10.1148/radiol.232780.PMID: 38501952.+ While lower resolution is presented as a key advantage, computational cost, memory usage, or inference time savings are not reported, weakening the practical significance of the claim.
    • Overall performance is very close to Mirai, and without a full resolution baseline or cost analysis, it is difficult to demonstrate a clear advantage attributable to inductive bias design beyond architectural variation.
  • 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 does not mention open access to source code or data, but provides a clear and detailed description of the algorithm to ensure reproducibility.

  • 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 presents a sound and clinically motivated contribution to mammography based cancer risk prediction by explicitly integrating complementary inductive biases (CNNs and transformers) with bilateral contralateral modeling, and demonstrates performance comparable to Mirai at a substantially reduced input resolution. The use of a large real world dataset, evaluation on both in distribution and out of distribution cohorts, and stratified analysis across demographic and technical factors strengthen the empirical validity of the approach. However, the contribution is primarily incremental rather than transformative: performance gains over Mirai are modest, the architectural description and training protocol lack clarity, and key claims regarding inductive bias and computational efficiency are not fully substantiated experimentally (e.g., missing full resolution baselines, cost analysis, and interpretability). Despite these limitations, the paper addresses an important problem, introduces a reasonable architectural design with explicit clinical priors, and provides generally solid experimental evidence, making it suitable for acceptance with revisions rather than strong endorsement.

  • 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 the constructive feedback. We group concerns below and point to existing manuscript text where applicable. (I) Resolution (R2,R3). We agree that the abstract/introduction overemphasise low resolution. The 512×512 setting reflects our training compute budget, not a claim that lower resolution is preferable. We do not advocate low-resolution mammography in clinical practice. §4 ¶2 already states this “Lower resolution was … a controlled stress-test of representation and fusion choices, rather than to argue against high-resolution” and the Conclusion echoes the same point, but we agree this framing must move to abstract and introduction. Camera-ready will: (a) frame the contribution as complementary CNN–ViT representations plus explicit bilateral fusion; (b) clarify that 512×512 reflects our compute budget rather than a preference. We also acknowledge R2’s point on DCIS subgroup gap (Tab.2) may be consistent with downsampling attenuating fine-scale signal, and will note this limitation explicitly. (II) MIRAI comparison(R1,R3). As R1 notes, §4 ¶4 (Limitations) addresses this point transparently, but we agree and our camera-ready will point at this more visibly. In applying pretrained MIRAI to unseen cohorts, we follow Yala et al.’s own external-validation paradigm. We see the point of R1 about “matches or slightly exceeds”, and we propose that camera-ready will say “comparable, with overlapping CIs”. To clarify, MIRAI is included as a strong external reference, evaluated with its released model and intended preprocessing. Thus, we do not use MIRAI to argue architectural superiority. The controlled evidence for our design comes from ablations under the same preprocessing and input settings: DINO-only 0.621, SE-ResNeXt-only 0.668, max aggr. 0.727, concat. 0.731, and BilateralMixer 0.736.(III) Ablation gains(R1,R2,R3). The core contribution is the integrated architecture, not any single block. Tab. 1 reports three converging ablations at matched 512×512: DINO-only 0.621→MamaDino 0.736 (+0.115); SE-ResNeXt-only 0.668→MamaDino 0.736 (+0.068); and max aggr.→BilateralMixer: 0.727→0.736 in-distribution and 0.666→0.677 OOD. §4 ¶1 already states that these gains are not attributable to any single component. The +0.005 BilateralMixer-vs-concat gain is small in isolation; its value lies in being one of three stacked architectural changes, not a standalone gain. We will mark this more explicitly. (IV) DINOv3(R2). We agree that fine-tuning is a natural next step. We froze DINOv3 to test it as a self-supervised prior, consistent with our Introduction and recent medical DINOv3 studies using frozen/off-the-shelf encoders [9,13]. Task-specific learning was instead concentrated in SE-ResNeXt, cross-attention, and BilateralMixer. This kept the four-view model tractable under our compute budget and avoided adding a large ViT fine-tuning/hyperparameter-search dimension. We do not claim freezing is optimal; full or parameter-efficient fine-tuning is future work. (V) Data, case-control, calibration(R1,R2,R3). The in-distribution test set is held out and patient-disjoint from train/val.; the OOD cohort is from Oxford, an unseen OPTIMAM site, not an additional dataset. The 3-year target is defined in §2.1 as cancer within three years after baseline screening or screen-detected malignancy at next routine screen. We agree that matched case-control design supports discrimination assessment, but not calibration, PPV, or deployment thresholds under real screening prevalence; §4 already states this limitation. R3: 53,883 women refers to the training cohort, while test cohorts are reported separately in §2.1.(VI) Clarity, citations(R3,R2). Staged training is described in §2.3: Stage1 trains SE-ResNeXt plus cross-attention with DINOv3 frozen; Stage2 freezes the fusion encoder and trains only BilateralMixer. We will reference this earlier, clarify Fig.1 to show view-specific paths and fusion points, and cite AsymMirai and Mendes review.



Meta-Review

Meta-review #1

  • Your recommendation

    Invite for Rebuttal

  • 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 reviews were mixed for this paper. While there was some consensus about the importance of the problem, there were concerns about the clarity of the technical description, the motivation for the approach, as well as possibly unfair comparisons to MIRAI. Lastly, I have a concern about the very modest improvements demonstrated in the ablation study related to the core contribution of the method. I ask the authors to provide a rebuttal to address these concerns.

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

    N/A

  • Please justify your recommendation.

    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.

    The authors have adequately addressed all of the reviewers’ concerns. Accordingly, I recommend acceptance of the paper.



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.

    The authors’ rebuttal thoroughly clarified key questions regarding input resolution, frozen DINOv3 design and comparative evaluation settings. They properly acknowledged existing limitations and will adjust related claims in revision.



back to top