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

Abstract

Multi-view mammography benefits from complementary craniocaudal and mediolateral oblique projections, yet differences in acquisition quality and anatomical coverage often lead to conflicting diagnostic evidence between views. Most existing multi-view learning approaches address this variability through deterministic feature fusion, implicitly assuming that disagreement can be resolved by stronger representations, an assumption that can yield brittle or overconfident predictions in safety-critical screening settings. We present Conflict–Evidential Inference Network (CEI-Net), an uncertainty-aware framework that reframes multi-view mammography classification as evidential reasoning under disagreement. Rather than enforcing consensus at the feature level, CEI-Net represents each view as a source of uncertain evidence and performs fusion by selectively moderating class hypotheses that are unsupported across views. This design allows disagreement to manifest as calibrated uncertainty when appropriate while still enabling confident decisions when reliable evidence is available from a single view. Empirical evaluation of the CBIS-DDSM and VinDr-Mammo benchmarks shows that this conflict-aware inference paradigm yields consistent improvements in discrimination and calibration over deterministic fusion methods, without adding auxiliary routing/quality networks. These findings suggest that explicitly reasoning about disagreement, rather than absorbing it, is critical for robust and interpretable multi-view decision-making in clinical screening. Our code is publicly available at https://github.com/rayhan-ahmed91/CEI-Net.

Links to Paper and Supplementary Materials

Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/2986_paper.pdf

SharedIt Link: Not yet available

SpringerLink (DOI): Not yet available

Supplementary Material: Not Submitted

Link to the Code Repository

https://github.com/rayhan-ahmed91/CEI-Net

Link to the Dataset(s)

vindr-mammo dataset: https://physionet.org/content/vindr-mammo/1.0.0/ cbis-ddsm dataset: https://www.cancerimagingarchive.net/collection/cbis-ddsm/

BibTex

@InProceedings{AhmMd_When_MICCAI2026,
        author = { Ahmed, Md Rayhan AND Lasserre, Patricia},
        title = { { When Views Disagree: Conflict-Aware Evidential Inference for Multi-view Mammography } },
        booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
        year = {2026},
        publisher = {Springer Nature Switzerland},
        volume = {LNCS 16882},
        month = {September},
        page = {pending}
}

Reviews

Review #1

  • Please describe the contribution of the paper

    This paper addresses a clinically important problem in multi-view mammography: inter-view conflict and overconfident deterministic fusion. The proposed CEI-Net introduces a novel uncertainty-aware framework based on evidential deep learning and class-conditional conflict gating, which explicitly models disagreement rather than simply absorbing it. The work is technically sound and well-motivated. It achieves better performance than existing CNN, Transformer, and Mamba-based methods in both discrimination and calibration.

  • 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 work correctly identifies a critical limitation of existing multi-view fusion methods: they ignore view disagreement and produce overconfident, poorly calibrated predictions. The problem is clinically relevant. 2.The method is novel. 3.The paper is well-organized.

  • 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 paper is motivated by conflicting evidence between CC and MLO views, but the quantitative analysis mainly reports dataset-level AUROC, ECE, and NLL improvements, without stratifying performance by clinically important disagreement regimes such as low-quality views, partial breast coverage, dense breasts, subtle lesions, or artifact-heavy cases. As a result, the manuscript does not clearly demonstrate in which concrete clinical scenarios the proposed conflict-aware evidential inference is most beneficial, rather than simply improving average benchmark performance. 2.The whole framework relies on the assumption that inter-view disagreement can be adequately characterized by class-wise Jensen-Shannon divergence between per-view Dirichlet means and then moderated by a scalar conflict gate. However, the paper does not analyze whether this design remains reliable in harder cases such as severe asymmetry of lesion visibility, view-specific false positives, very weak evidence in both views, or anatomically inconsistent but still clinically meaningful complementary findings. Thus, the robustness of the core conflict-modeling assumption is still not fully demonstrated. 3.The method assumes consistent paired CC/MLO views and standard preprocessing, and although the paper includes a missing-view test-time ablation, it does not evaluate real-world issues such as view mispairing, acquisition mismatch, corrupted views, variable compression quality, or imperfect preprocessing and breast cropping. Since such factors are common in practical mammography workflows, the real clinical robustness and deployability of the framework remain unclear.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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?

    The problem is clinically relevant and important, and the method is novel. However, the validation requires further improvement.

  • 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 have made a satisfactory rebuttal that can address my comments.



Review #2

  • Please describe the contribution of the paper

    The main contribution of this paper is the proposal of the Conflict-Evidential Inference Network (CEI-Net). By leveraging evidential deep learning, CEI-Net provides an uncertainty-aware framework specifically tailored to the multi-view nature of mammography, effectively incorporating inter-view disagreement.

  • 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 driven by a strong motivation, as effectively leveraging the multi-view nature of mammography is of clinically relevant. The proposed method cleverly learns from incorporating two types of uncertainty by extracting view-specific features via an encoder and tokenization, computing Dirichlet evidence for each view, and applying class-wise conflict weighting. Furthermore, a major strength of this work is its comprehensive evaluation, successfully demonstrating improved AUROC and statistical significance compared to various recent SOTA models.

  • 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 major weaknesses of this paper are as follows:

    Baselines for Uncertainty: The baseline models used for comparison are deterministic and rely solely on simple temperature scaling, which may not provide a sufficiently rigorous baseline for evaluating uncertainty estimation. Visualizations: In the lower section of Figure 2, the clinical focus is somewhat ambiguous. It is unclear which specific regions or features within the mammogram the reader is expected to observe.

  • Please rate the clarity and organization of this paper

    Good

  • Please comment on the reproducibility of the paper. Please be aware that providing code and data is a plus, but not a requirement for acceptance.

    The authors claimed to release the source code and/or dataset upon acceptance of the submission.

  • 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
    1. While using 224x224 images is an understandable concession to computational constraints and aligns with the strongest baseline (Mammo-Clustering), mammograms are inherently ultra-high-resolution. This aggressive downsampling inevitably causes significant information loss. I would appreciate the authors’ perspective on this limitation. 2.Edge Cases: The proposed framework appears highly conservative, penalizing weights first for uncertainty and then again for inter-view disagreement. I am curious about the model’s robustness and reliability when handling complex edge cases, such as early-stage malignancies with ambiguous margins in dense breasts.
  • 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?

    I believe this is a good EDL study that effectively leverages the unique characteristics of mammography. Rather than simply adopting existing methodologies, the authors propose an architecture specifically tailored for this domain. My recommendation is based on their rigorous evaluation, which includes validation on two distinct open datasets, extensive comparisons with various baselines, and thorough ablation studies that clearly demonstrate the effectiveness of each component.

  • 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.

    This study is driven by a strong motivation, cleverly redefining multi-view disagreement as epistemic uncertainty to formulate and train EDL. The authors have successfully addressed my previous concerns. They provided a sufficient explanation of how the deterministic baseline was adapted for the EDL comparison and committed to improving the visualizations. As pointed out by myself and other reviewers, including an analysis of vulnerable edge cases related to false negatives would certainly strengthen the paper. However, since the authors have already demonstrated clear performance improvements over various SOTA baselines and provided a solid ablation study, I consider the current evaluation to be sufficient for MICCAI.



Review #3

  • Please describe the contribution of the paper

    The main contribution of this manuscript lies in its conceptual reframing of multi-view mammography classification, shifting the paradigm from deterministic feature fusion to an uncertainty-aware evidential reasoning framework. Specifically, author make three methodological contributions:

    1.Reframing Inter-View Conflict as Epistemic Uncertainty Unlike conventional methods that tend to absorb or average out disagreements between views , the proposed CEI-Net models each mammographic view as an uncertain evidence source via Dirichlet distributions. This approach successfully translates multi-view inconsistencies into interpretable epistemic uncertainty, which is highly valuable for preventing overconfident and potentially brittle predictions in ambiguous clinical scenarios.

    2.Class-Conditional Conflict Gating To intelligently aggregate conflicting evidence, the authors introduce a novel gating mechanism driven by Jensen-Shannon (JS) divergence. Rather than applying a uniform penalty across an entire view, this mechanism quantifies conflict at the level of individual diagnostic hypotheses. This allows the model to selectively attenuate evidence only for highly disputed classes, offering a more nuanced and interpretable fusion strategy.

    3.Uncertainty-Derived Reliability Weighting The framework dynamically computes a reliability score for each view based on the strength of its accumulated Dirichlet evidence. This design not only enables adaptive weighting during view fusion but also elegantly accommodates common clinical “missing-view” scenarios by naturally assigning zero weight to absent inputs without requiring architectural modifications.

  • 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 manuscript presents a well-motivated and methodologically sound approach to multi-view mammography classification. The major strengths of this work include:

    1.Novel Problem Formulation (Evidential Reasoning under Disagreement): The paper elegantly reframes multi-view mammography from a standard deterministic feature fusion problem to an uncertainty-aware evidential reasoning paradigm. Instead of forcing a consensus at the feature level—which can lead to overconfident predictions —the CEI-Net models each view as a source of uncertain evidence via Dirichlet distributions. This appropriately translates inter-view diagnostic conflict into calibrated epistemic uncertainty, which is highly valuable for safety-critical clinical screening.

    2.Methodological Elegance (Class-Conditional Conflict Gating): The proposed fusion mechanism uses Jensen-Shannon (JS) divergence to quantify conflict at the class hypothesis level. This is particularly innovative because it selectively attenuates only the disputed diagnostic hypotheses, rather than indiscriminately penalizing the entire view. Furthermore, the model achieves this directly in the evidence space without relying on auxiliary routing or image-quality sub-networks, ensuring high parameter efficiency.

  • 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.

    Q1: Risk of False Negatives Induced by the JS Gating Mechanism. The authors employ a Jensen-Shannon (JS) gating mechanism, defined as $g_k = \exp(-\gamma JS_k)$, to attenuate evidence when the CC and MLO views disagree. However, in clinical practice, it is common for a lesion to be visible in only one view (e.g., deep in the breast, captured only by MLO) due to anatomical occlusion. In this scenario, the CC view might confidently predict “benign” (seeing only normal tissue), while the MLO view confidently predicts “malignant.” This creates a high JS divergence, triggering the gating mechanism to strongly attenuate the “malignant” evidence, which could lead to a fatal false negative. How does the proposed method distinguish between “true asymmetric lesion visibility” and “spurious conflict caused by noise” to prevent the model from compromising and missing the diagnosis?

    Q2: Limited Spatial Interpretability of the Conflict Mechanism. While Figure 2 effectively demonstrates the interpretability of the JS gating mechanism at the mathematical and predictive level (showing how uncertainty increases in high-conflict cases like Fig. 2e), it lacks spatial/clinical interpretability. Since CEI-Net relies on extracting “Top-M salient spatial tokens” from each view, the model inherently possesses spatial information about where the evidence originates. Could the authors expand their qualitative analysis to visualize which specific spatial tokens or anatomical regions are driving the high JS divergence? Providing token-level heatmaps or spatial overlays alongside the probability charts would significantly bridge the gap between mathematical conflict resolution and actionable clinical interpretability.

    Q3: Dataset Discrepancy and Potential View Dominance. Could the authors explain why the proposed method achieves a significantly higher performance gain on CBIS-DDSM compared to VinDr-Mammo (whereas baseline methods perform relatively consistently across both)? Furthermore, in Table 2, the Deterministic Prob-Avg achieves an AUROC almost identical to the missing-MLO (CC only) setting. Does this suggest that the model’s decisions heavily rely on the CC view? Additionally, the Expected Calibration Error (ECE) for CC only is lower than that of MLO only, which seems to contradict the textual claim regarding the higher discriminative value of the MLO view.

  • 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

    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?

    This paper presents a novel uncertainty-aware evidential reasoning framework (CEI-Net) for multi-view mammography classification. The transition from deterministic feature fusion to epistemic uncertainty modeling is a strong conceptual contribution. However, critical concerns regarding clinical safety and empirical interpretability preclude a higher initial score.

    Key Strengths (Driving the Accept):

    Evidential Reframing: Effectively models inter-view conflict as epistemic uncertainty via Dirichlet distributions, mitigating overconfident predictions in ambiguous cases.

    Class-Conditional Gating: The JS divergence-based gate provides a mathematically elegant, hypothesis-level conflict resolution mechanism, offering more nuanced fusion than global pooling approaches.

    Inherent Robustness: Uncertainty-derived reliability weighting naturally handles missing clinical views without requiring architectural modifications.

    Major Reservations (Driving the “Weak” modifier):

    Clinical Safety (False Negatives): The conflict gating mechanism risks treating true asymmetric lesion visibility (e.g., a tumor visible only in MLO due to depth) as “conflict,” potentially suppressing critical malignant evidence and causing fatal false negatives.

    View Dominance: Ablation results (Deterministic Prob-Avg vs. CC only) strongly suggest the model’s decision-making is overwhelmingly biased toward the CC view, contradicting text assertions regarding the higher discriminative value of the MLO view.

    The methodological innovation is sound. I am open to raising my score if the authors can adequately address the false-negative edge cases and provide spatial evidence of interpretability during the rebuttal.

  • 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



Author Feedback

We thank the reviewers and AC for recognizing the clinical motivation, novelty, clarity, and empirical evaluation of CEI-Net. We revised Fig. 2 and related text to clarify conflict behavior and spatial evidence. Safety under asymmetric visibility (R1/R2/R3). A lesion visible in only one view is an important safety case. CEI-Net does not force CC/MLO consensus or convert disagreement into a benign decision. Each view first contributes evidence according to its Dirichlet strength, r(v)=S(v)/(S(v)+K). The JS gate is then applied to reliability-weighted evidence, not directly to class probabilities or to an entire view. Thus, when one view is confident and the other is weak or uncertain, the confident view can still dominate. When two confident views disagree, the gate reduces Dirichlet concentration for the disputed hypothesis and increases epistemic uncertainty rather than suppressing malignancy. We now clarify that such cases are uncertain/high-risk for review/triage. For single-view inference, JS is set to zero, so no artificial conflict is introduced. Missing-view results support this: CC-only and MLO-only remain functional, while full model performs best. Disagreement regimes and robustness (R1/R2/R3). Dense breasts, partial coverage, artifacts, mismatch, corrupted views, and view-pairing errors matter in deployment. We do not claim full prospective robustness to all acquisition failures, and the conclusion now clarifies this limitation. The current evidence tests the central claim that conflict modeling improves inference under disagreement or shift. Fig. 2 stratifies cases by JS disagreement and shows that stronger disagreement is associated with higher error and NLL. Table 2 evaluates missing-view inference without retraining, and cross-dataset results show improved calibration and NLL under domain shift. We will study view mispairing, corruption, imperfect cropping, and dense/subtle-lesion cases in future robustness tests. Uncertainty baselines and ablations (R2). To separate temperature scaling from uncertainty modeling, we compare CEI-Net with evidential variants using the same backbone and tokenization, including equal-weight fusion, reliability-only fusion, JS-gating-only fusion, and removal of tokenized evidence. Temperature scaling is applied only for evaluation calibration, while the actual uncertainty mechanism is the Dirichlet evidence strength plus JS-based conflict modulation. We now state this explicitly. Spatial interpretability and revised Fig. 2 (R2/R3). The previous figure made belief-level conflict clearer than clinical localization. We revised Fig. 2 to overlay selected Top-M spatial tokens and distinguish malignant, benign, and disputed tokens in specific regions. The revised panels connect image regions selected by CEI-Net with per-view belief, JS disagreement, and class-wise gates, which links conflict to spatial evidence.s View dominance and dataset discrepancy (R3). The similarity between deterministic probability avg. and CC-only, especially on CBIS-DDSM, reflects naive averaging, not CEI-Net collapsing to CC. Full model outperforms CC-only, MLO-only, and deterministic avg. on both datasets. Compared with CC-only, full CEI-Net improves AUROC from 0.796 to 0.855 on CBIS-DDSM and from 0.804 to 0.829 on VinDr-Mammo, while reducing ECE from 0.135 to 0.109 and from 0.300 to 0.073.This indicates that fusion uses complementary information from both views. We corrected the wording about MLO being generally more discriminative: in these splits, CC has stronger standalone signal, while full CEI-Net remains best. Image resolution (R2). 224x224 downsampling can remove subtle findings. We used this resolution for controlled comparison with baselines and feasible multi-run evaluation. Since all methods use the same preprocessing, gains reflect conflict-aware evidential inference rather than a resolution advantage. We now state this limitation and will explore high-resolution patch or multi-scale extensions.



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 motivation and novelty of this work are consistently recognized; while, some concerns (especially those regarding empirical evaluations under typpical situations) should be addressed.

  • 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.

    Overall, a strongly motivated paper, and the authors provided a solid rebuttal, which addressed major concerns rasied by the reviwers.



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 three reviewers recognize a novel formulation of multi-view disagreement in terms of epistemic uncertainty, and note an elegant method and a strong contribution. Several of the issues raised by the reviewers were adressed in the rebuttal through an enriching discussion. This clarifications should be also included in 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 paper presents a novel method for handling multi-view conflict in breast mammography, a particularly relevant and frequent clinical case. To do so, the authors model this conflict as a source of epistemic uncertainty using Dirichlet distributions and propose a gating mechanism based on Jensen-Shannon (JS) divergence. Experiments are conducted on two large-scale mammography datasets and demonstrate the framework’s benefits in managing conflicting cases. In the first round of reviews, the reviewers’ main concerns were: i) management of asymmetric visibility cases (when a lesion is not visible on a view), disagreement regimes, and treatment of single-view cases; ii) uncertainty baselines and interpretation; and iii) view dominance and potential bias. The authors clearly responded to all the reviewers’ points, promising to better clarify asymmetric visibility, single-view, and robustness to acquisition features in the camera-ready paper. In addition, the authors revised Fig. 2 to address R2 and R3 concerns about the spatial interpretability of tokens. R1 and R2 were satisfied with these answers and recommended acceptance. While R3 did not provide an updated recommendation, I believe the authors addressed his questions in the rebuttal and therefore also recommend acceptance.



back to top