MammoFlow: Multiview Mammogram Synthesis with Anatomically Consistent Flow Matching

Du, Yuexi; Barrientos, Leya; Sheiman, Laura; Lewin, John; Tagare, Hemant D.; Dvornek, Nicha C.

MICCAI 2026

MammoFlow: Multiview Mammogram Synthesis with Anatomically Consistent Flow Matching

Yuexi Du, Leya Barrientos, Laura Sheiman, John Lewin, Hemant D. Tagare, Nicha C. Dvornek

Department of Biomedical Engineering and Department of Radiology & Biomedical Imaging, Yale University

MammoFlow generates paired CC and MLO mammograms from noise while preserving cross-view tissue correspondence through an EMD-driven anatomical consistency loss.

arXiv Code Citation

MammoFlow teaser showing paired multiview mammograms and AP-axis tissue correspondence — Multiview mammography projects a shared breast volume into complementary CC and MLO views.

How the AP-axis tissue distribution is computed

Each profile summarizes relative tissue mass from the chest wall to the nipple.

Preprocess the views. Crop each image to the breast region and mask the pectoral muscle in the MLO view so it does not bias the tissue profile.
Align the AP axes. Keep the CC view fixed while testing MLO rotations and horizontal translations that orient its AP axis horizontally.
Project to one dimension. At every AP position, sum pixel intensities vertically, normalize by total image intensity, and apply 1D Gaussian smoothing for spatial relaxation.
Compare the profiles. EMD = Σᵢ |P[i] − Q[i]| compares the cumulative CC and MLO distributions; the transform with the lowest EMD is selected.

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 MammoFlow, a method for synthesizing multiview mammograms by leveraging the inherent geometric relationship between CC and MLO views. An alignment module searches a compact affine transformation subspace to establish anatomical correspondence, and a pixel-space Earth Mover's Distance (EMD) loss encourages generated pairs to share physically plausible anteroposterior tissue distributions.

Integrated into a pretrained flow matching model, MammoFlow produces high-fidelity paired mammograms, passes expert radiologist evaluation, and improves downstream classification AUC by up to 5% when synthetic malignant pairs are added to training.

Method

Training uses real paired views to compute an anatomical alignment prior. Inference only requires Gaussian noise and a prompt.

MammoFlow pipeline with conditional rectified flow, EMD-driven alignment, and temporal loss scheduling

Conditional flow matching

Paired CC/MLO mammograms are concatenated and encoded into a latent path between image data and Gaussian noise. The model predicts the velocity toward clean mammogram pairs.

EMD-driven alignment

A compact search over MLO rotation and horizontal translation aligns AP-axis tissue distributions with the CC view while accounting for pectoral masking.

Multiview regularization

A differentiable EMD loss penalizes cross-view tissue displacement in decoded one-step reconstructions, with a cosine schedule that emphasizes low-noise structure.

Results

MammoFlow improves both image fidelity and multiview correspondence across CSAW, VinDr-Mammo, and RSNA.

1.08% CSAW relative EMD difference

97.5% expert-rated pairing success

+5% downstream CSAW AUC gain

Image Generation And Multiview Correspondence

Method	CSAW				VinDr				RSNA
Method	FID	FrD	Δ EMD	Δ JSD	FID	FrD	Δ EMD	Δ JSD	FID	FrD	Δ EMD	Δ JSD
GT oracle	8.89	3.02	-	-	8.36	5.42	-	-	9.76	12.0	-	-
CA3D-Diff*	63.3	8.72	14.12	16.89	70.5	6.72	173.83	498.86	64.5	8.73	37.20	90.98
Mammo-RGB	88.6	14.9	60.67	188.53	102.8	21.5	33.01	49.57	148.6	15.1	15.43	20.98
Vanilla	73.4	17.7	10.82	68.52	70.5	48.8	18.37	54.61	66.3	16.4	7.75	28.86
MammoFlow	53.3	6.60	1.08	9.57	67.5	12.4	4.19	15.22	65.4	12.7	2.73	33.48

*CA3D-Diff requires a ground-truth reference image. Lower is better for FID, FrD, Δ EMD, and Δ JSD.

Downstream Classification AUC

CSAW real.7452

CSAW + 5k synth..7904

VinDr real.7775

VinDr + 5k synth..8196

RSNA real.7676

RSNA + 5k synth..7822

Reader Study

Authenticity37.5 ± 12.5

MammoFlow images rated as real

Pairing97.5 ± 2.5

Generated pairs judged anatomically paired

Qualitative Results

Rows 1–2 compare generated CC/MLO pairs, while the final row shows the aligned views and AP-axis tissue distributions for Case 2.

Qualitative comparison of ground truth, CA3D-Diff, Mammo-RGB, and MammoFlow samples with AP-axis distribution plots

Orange arrows: cross-view artifacts. Baselines may reproduce contours from the other view or place high-density tissue in only one projection, indicating mismatched anatomy across the CC/MLO pair.

Purple arrows: tissue-distribution mismatch. Baseline AP-axis profiles show clear density differences between views, whereas MammoFlow more closely follows the native mass distribution of the ground-truth pair.

Code Release

The repository contains data preprocessing, SD3.5 multiview training, generation inference, alignment evaluation, and downstream classification scripts.

Pretrained model: We are working on it, and it will be released soon.

Research use only. MammoFlow is not a medical device and must not be used for clinical diagnosis or treatment decisions.

git clone https://github.com/XYPB/MammoFlow.git
cd MammoFlow
conda env create -f environment.yml
conda activate mammoflow

Citation

@inproceedings{mammoflow2026,
  title     = {MammoFlow: Multiview Mammogram Synthesis with Anatomically Consistent Flow Matching},
  author    = {Du, Yuexi and Barrientos, Leya and Sheiman, Laura and Lewin, John and Tagare, Hemant D. and Dvornek, Nicha C.},
  booktitle = {International Conference on Medical Image Computing and Computer Assisted Intervention},
  year      = {2026},
  eprint    = {2606.28537},
  archivePrefix = {arXiv}
}