List of Papers Browse by Subject Areas Author List
Abstract
Multiview mammography relies on paired craniocaudal (CC) and mediolateral oblique (MLO) views to provide complementary projections of a 3D breast volume, enabling precise anomaly localization. However, acquiring high-quality, balanced datasets remains challenging for deep learning applications. We propose a novel method to synthesize multiview mammograms by leveraging the inherent geometric relationship between CC and MLO views. To enforce an implicit 3D consistency prior during generation, we develop an alignment module that searches a 2D affine transformation subspace to establish optimal anatomical correspondence. Leveraging this alignment, we introduce a pixel-space self-consistency loss based on the Earth Mover’s Distance (EMD) between the 1D anteroposterior (AP) axis tissue distributions of the generated images. Integrated into a pretrained flow matching model, MammoFlow forces synthesized pairs to share physically plausible tissue distributions from the chest wall to the nipple. To our knowledge, this is the first work to guide multiview mammogram generation using implicit geometric tissue correspondence. Our method demonstrates superior image quality, passes expert radiologist evaluation, and generates physically consistent pairs that improve downstream classification AUC by 5%.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026/paper/3745_paper.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to the Code Repository
https://github.com/XYPB/MammoFlow
Link to the Dataset(s)
N/A
BibTex
@InProceedings{DuYue_MammoFlow_MICCAI2026,
author = { Du, Yuexi AND Barrientos, Leya AND Sheiman, Laura AND Lewin, John AND Tagare, Hemant D. AND Dvornek, Nicha C.},
title = { { MammoFlow: Multiview Mammogram Synthesis with Anatomically Consistent Flow Matching } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 16890},
month = {September},
page = {pending}
}
Reviews
Review #1
- Please describe the contribution of the paper
This paper presents MammoFlow, a multiview mammogram synthesis framework that explicitly leverages the geometric relationship between craniocaudal and mediolateral oblique views. The key contribution lies in introducing an alignment module that searches within a 2D affine transformation subspace to establish anatomical correspondence between views. Based on this alignment, the authors propose a self-consistency constraint using Earth Mover‘s Distance to enforce consistency of tissue distributions along the anteroposterior axis. By integrating this constraint into a flow-based generative model, the method aims to produce anatomically consistent multiview mammograms. The generated data are further shown to improve downstream classification performance.
- 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 a clinically relevant problem in multiview mammography, where paired CC and MLO views are essential for accurate diagnosis but are often limited in availability for data-driven models. 2.The proposed method incorporates geometric prior knowledge into the generative process. In particular, the use of an alignment module combined with a distribution-level consistency constraint based on EMD provides a meaningful way to enforce cross-view anatomical consistency. 3.The experimental results demonstrate improved image quality and include both quantitative evaluation and expert radiologist assessment, which strengthens the practical relevance of the work. The reported improvement in downstream classification performance further supports the utility of the generated data.
- 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 methodological description lacks clarity in several aspects. While the paper introduces multiple components, the overall training pipeline is not clearly described. In particular, the inputs and outputs during both training and inference stages are not explicitly defined, making it difficult to fully understand how the system operates end-to-end. 2.The presentation of the framework is somewhat fragmented, with individual modules described in isolation. A more structured explanation of the data flow and optimization process would improve readability and reproducibility. 3.Figure 3 is difficult to interpret. In particular, it is unclear what the inputs and corresponding outputs are in the figure, making it hard to follow the intended process. The main text also does not provide sufficient explanation to guide the reader through this figure. 4.The alignment process is based on a 2D affine transformation search, but the physical plausibility of such transformations in the context of breast deformation is not sufficiently discussed. Given the complex deformation between CC and MLO views, more justification or analysis of the alignment strategy would strengthen the method.
- 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.
(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?
Overall, the paper tackles an important and clinically relevant problem and proposes a method that integrates geometric priors into a generative framework for multiview mammogram synthesis. The idea of enforcing anatomical consistency through alignment and distribution-based constraints is interesting and shows a certain level of novelty. The experimental results, including radiologist evaluation and downstream task improvement, further support the potential impact of the work. However, the paper has several limitations in terms of clarity and methodological detail. The training and inference procedures are not sufficiently described, and the overall framework lacks a clear and coherent presentation. In addition, some experimental figures are difficult to interpret, and the assumptions behind the alignment module require further justification. Considering these strengths and weaknesses, I believe the paper is marginally above the acceptance threshold, and I assign a score of Weak Accept.
- 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.
The paper addresses an interesting and clinically relevant problem of multi-view breast mammography generation between CC and MLO views. The proposed framework is technically meaningful and aligns well with practical clinical requirements. In the rebuttal, the authors adequately addressed the concerns raised in my initial review, including clarifications on the methodology and experimental details. I believe the paper has potential impact to merit acceptance.
Review #2
- Please describe the contribution of the paper
This study proposes the MammothFlow multi-view mammography image synthesis method, which ensures anatomical consistency between the craniocaudal (CC) and mediolateral oblique (MLO) views.
- 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.Using AP-axis projection matching combined with EMD (Equations 4–6) is a simple and physically sound constraint method for dual-view mammography. 2.This method integrates seamlessly with flow matching. The plug-and-play loss function using a cosine scheduler in Equation 7 is highly practical.
- 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.Although the motivation for this study is sound, the architectural contribution is not significant given that the model relies on a large SD 3.5 backbone and most of its components are frozen. The paper should more clearly explain the fundamental difference between the AP-axis EMD regularizer and previous dual-view constraints other than the Mammo-RGB channel stacking. 2.Using true-label-based alignment during training can lead to ambiguity. Section 2.3 applies pre-computed ground-truth alignment parameters (θ̂, τ̂) to samples generated during training. While this is reasonable for ensuring consistency in supervision, it remains unclear how the alignment issue is handled during the inference stage when ground-truth labels are unavailable. 3.Please provide the exact formulas for ΔEMD/ΔJSD, the aggregation protocol, and whether any alignment methods were applied to synthetic images during evaluation.
- 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.
(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?
MammoFlow employs an image synthesis method based on anatomical consistency, effectively leveraging this physical property while improving accuracy in downstream cancer classification with small datasets.
- 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.
I thank the authors for their thoughtful rebuttal. Their response has addressed my main concerns in a satisfactory manner and has clarified the key issues I raised in my review. I am therefore convinced of the merit of the work and support acceptance.
Review #3
- Please describe the contribution of the paper
The authors propose a method for mammography image synthesizing.
The proposed method rely on density profile adjustment intended to generate more realistic and plausible multi-view images.
The method is assessed on a downstream task of multiview classification.
- 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.
Using Earth Mover Distance to optimize dense tissue distribution for more plausible images appears to be an interesting approach.
That is, the clinical and physical plausibility is an important concern for the synthesized medical imaging.
- 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.
While the authors propose an interesting method of a physically plausible image synthesizing, there is a lack of discussion about the need of such an approach, both from the point of view of the state of the art and from the point of view of the methodological relevance. While the proposed method appears to allow for improving the classification tasks performances, the is a lack of discussion about how the proposed technique compares to other state of the art techniques.
- 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
State of the art: Other synthesizing methods may be explored or commented Such as[1, 2] [1] M. Sorkhei et al., ‘CSAW-M: An Ordinal Classification Dataset for Benchmarking Mammographic Masking of Cancer’, Dec. 13, 2025, arXiv: arXiv:2112.01330.doi: 10.48550/arXiv.2112.01330.
[2] R. Walsh and M. Tardy, ‘A Comparison of Techniques for Class Imbalance in Deep Learning Classification of Breast Cancer’, Diagnostics, vol. 13, no. 1, Art. no. 1, Jan. 2023, doi: 10.3390/diagnostics13010067.
Downstream classification task: Could the authors provide more details about the downstream task? In particular, how the synthesized images get the labels and what labels do they get?
- 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?
The image synthesizing in medical imaging remains a challenging and doubtful task. I would expect more discussion about the limitations and advantages of the proposed method.
- 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 answers provided by the authors are well structured and comprehensive. That is, the authors provide some reasonable arguments with regard to the positioning and clinical need of the paper as asked in the review. Therefore, I change my opinion to “accept”.
Author Feedback
We thank reviewers for recognizing our work’s clinical relevance (R1), EMD-guided flow matching value (R2), and physical plausibility (R3).
R1W1.Training & inference pipeline: Training input is a pair of raw mammograms and the corresponding text prompt, encoded for L_CRF. The images are also used to compute the alignment for L_EMD on the one-step reconstructed image. Inference takes only Gaussian noise and a prompt to generate novel paired views; no alignment is needed. R1W2.Presentation: We will revise Sec. 2 to clarify the full pipeline: training uses real pairs to compute losses, while inference uses only the trained model with noise and pre-defined prompts. R1W3.Fig. 3: Fig. 3 shows 2 random CC&MLO pairs generated from Gaussian noise with the same prompt per method. Col. 1 shows GT images. Rows 1-2 show synthetic cases. The final row plots the aligned view and AP-axis tissue distribution for case 2.Arrows highlight artifacts. R1W4.2D affine alignment: Affine alignment corrects macroscopic positioning, e.g., acquisition angle. Nonrigid deformation is mitigated by spatial relaxation with 1D Gaussian smoothing (Sec. 2.2). Removing it greatly degrades generation (Tab. 2; FID 53.3 to 74.7). Since EMD targets macro-level optimization, we use efficient affine alignment over complex nonrigid estimation.
R2W1.Contribution: We add a novel anatomical regularizer, orthogonal to the base model. Since Mammo-RGB and Vanilla also use SD3.5, improvements in Tab. 1 isolate our method’s direct impact. R2W2.Prior constraints: Prior methods treat views separately[1,6], stack them[7], or translate from a GT reference with oversimplified fixed rotation[12]. We model the physical multiview relation: aligned AP-axis tissue distributions should match because they share a breast volume. EMD gives a direct, differentiable anatomical training signal. R2W3.GT alignment: Pre-computed alignment is a training-only prior used to measure AP-axis consistency on one-step predictions. At inference, no GT labels or alignments are used, as the model has learned the correspondence. R2W4.ΔEMD/ΔJSD: We compute EMD/JSD, denoted as D, between CC and aligned MLO AP-axis distributions, and average over the test set. Metric: ΔD=100×|mean(D_syn)−mean(D_GT)|/mean(D_GT). Lower values mean synthetic pairs better match real image statistics. R2W5.Evaluation alignment: No alignment during inference. We apply affine alignment only when computing EMD/JSD for paired images. All other evaluations use raw synthetic images without alignment.
R3W1.Clinical need: MammoFlow augments data-limited, imbalanced training. Cancer cases and reliable synthetic CC/MLO pairs are rare. Since radiologists rely on cross-view correspondence, single-view realism is insufficient; synthesized pairs must be anatomically consistent to be useful(p. 2). R3W2.Relation to SOTA: Single-view synthesis[1,6] lacks CC/MLO consistency. View translation[12] requires a GT reference. Tumor artifacting[R3.2] generates single-view images from existing data and may lead to incorrect anatomy. [7] ignores physical relationships. We instead regularize AP-axis correspondence to synthesize novel images. R3W3.Suggested works: [R3.1] resamples real data rather than synthesizes new data, and [R3.2] injects GT tumors to healthy images. Both operate on existing data and don’t create new paired morphologies. We provide constrained paired synthesis with better diversity, which can be combined with these methods. R3W4.Downstream task: We classify binary breast cancer. Prompts control healthy/cancerous label during generation. We fine-tune MammoFlow only on malignant cases to generate malignant pairs, used only to balance classifier training distribution (Sec. 3.1). R3W5.Strengths & limitations: Synthetic mammograms don’t replace clinical data, but anatomical consistency improves their utility (Tab. 1,3,4). Limitations (memory cost, affine approximation, imperfect preprocessing) are in Sec. 4 and will be expanded.
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.
This paper proposes a Multiview mammogram synthesis framework that enforces anatomical consistency between CC and MLO views via a geometry-aware alignment module and an EMD-based distribution constraint. Everything is integrated into a flow-matching generative model.
The reviewers agree that the motivation is sound. The use of geometric priors and EMD to enforce cross-view consistency seems to be novel and physically meaningful. The inclusion of radiologist assessment strengthen the work.
At the same time, the reviewers identified multiple issues, that must be addressed in a rebuttal. These include that the training and inference pipeline are insufficiently described. Also the 2D affine alignment model requires further justification given complex breast deformations. The presentation of the method needs to be improved. Additionally, discussion about the clinical need of such synthesized data is required, also in view of other related works, as listed by R#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 addressed the main issues raised by the reviewers in the rebuttal and all reviewers agree on the acceptance of this work.
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’s contribution lies in integrating Earth Mover’s Distance as a cross-view consistency objective within a flow matching–based framework for multi-view mammogram generation. The formulation is technically sound and well-motivated for the problem setting. The reviewers converged on a positive recommendation, and I concur with this consensus in supporting 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.
Accept
- Please justify your recommendation.
The reviewers highlight the clinical relevance of the targeted problem, the novel use of EMD-based distribution constraint to enforce cross-view anatomical consistency, and the solid experimental results supported by both quantitative evaluation and radiologist assessment.
The reviewers also raised concerns about the clarity of the training and inference pipeline, the justification of the 2D affine alignment under complex breast deformations, and the lack of discussion on related work. These concerns were adequately addressed in the rebuttal, where the authors clarified the pipeline, the role of alignment at training versus inference, and the positioning of the work relative to existing methods.
Considering the strengths of the work and the satisfactory rebuttal, the paper is recommended for acceptance.