Divide and Conquer: A Large-Scale Dataset and Model for Left-Right Breast MRI Segmentation

Benjamin Hamm; Klaus Maier-Hein; Maximilian Rokuss; Yannick Kirchhoff

REVIEW 3 major objections 5 minor 13 references

Divide and Conquer: A Large-Scale Dataset and Model for Left-Right Breast MRI Segmentation

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper releases the first public breast MRI dataset with separate left and right breast labels—more than 13,000 cases—along with a model that separates the breasts at 99 percent Dice.

desk verdict A genuinely useful left-right breast MRI dataset, but the 99% Dice is self-referential and the COM splitting method may mishandle asymmetry and mastectomy cases. read the letter →

arxiv 2507.13830 v1 pith:MOWTQFXR submitted 2025-07-18 eess.IV cs.CV

Maximilian Rokuss , Benjamin Hamm , Yannick Kirchhoff , Klaus Maier-Hein This is my paper

classification eess.IVcs.CV

keywords breastMRIleft-rightsegmentationmedicalimagingdatasetnnU-Netactivelearningpseudo-labelingregionofinterestwomen'shealth

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper introduces the first publicly available breast MRI dataset in which each of more than 13,000 3D volumes carries separate left and right breast segmentation labels. The labels were generated by splitting existing whole-breast masks at the center of mass, then scaled to the full dataset through active-learning pseudo-label transfer and reviewed in their entirety by three experts. Alongside the dataset, the authors release an nnU-Net model that segments the left and right breast as distinct regions, reporting a 5-fold cross-validated Dice Similarity Coefficient of 99.08 ± 0.48%. The resource is meant to give downstream tasks—per-breast classification, lesion localization, dynamic contrast-enhanced temporal analysis, and model pretraining—a reliable side-specific region of interest. If the label quality holds up, this closes a practical gap in breast MRI analysis, especially for unilateral abnormalities and post-mastectomy cases.

What carries the argument

The load-bearing object is the side-specific segmentation mask produced by a center-of-mass splitting plane: for each seed case, the centroid of the whole-breast mask defines a sagittal plane, and voxels on either side are assigned to the right and left breast. This turns coarse whole-breast annotations into explicit left/right ROIs without new manual delineation. The model is nnU-Net in its low-resolution configuration, chosen so that a single patch can contain both breasts at once. Active learning then transfers labels from non-contrast T1 predictions to co-registered sequences, and iterative retraining plus full manual review scales the 100 seed cases to 13,752 labeled volumes. This machinery is what lets the paper claim both large scale and a consistent side-specific standard.

What would settle it

Take a random sample of roughly 100 released cases, have two or more independent experts draw left and right breast masks on the original MRI volumes without seeing the model outputs, and compute Dice against the released labels; a large drop from the reported 99.08% would show the self-generated labels do not carry the claimed accuracy.

Watch

Extended reading notes

Core claim

On its own terms, the paper's claim is that explicit left-right breast segmentation is feasible at scale and useful as a divide-and-conquer front end for breast MRI analysis. The authors aggregate 13,752 scans from four public sources, harmonize them to RAS (right-anterior-superior) orientation in NIfTI format, and produce side-specific masks, starting from 100 whole-breast segmentations whose centroids define sagittal splitting planes. An active-learning loop propagates these labels across the remaining scans, using non-contrast T1 predictions as pseudo-ground truth for co-registered sequences, with every case visually inspected and either accepted, discarded, or refined. In 5-fold cross-validation, the resulting nnU-Net model achieves a mean Dice of 99.08 ± 0.48%, a Normalized Surface Dice at 1 mm of 99.10 ± 0.92%, and a 95th-percentile Hausdorff distance of 0.0196 mm. The discovery is that this pipeline yields a consistent, side-specific ROI representation across heterogeneous protocols, including challenging cases with asymmetry, implants, and unilateral breast absence.

Load-bearing premise

The load-bearing premise is that the pseudo-labels produced by the active-learning pipeline, after visual review by three experts, are accurate enough to serve as ground truth both for validating the model and for benchmarking others, since the reported Dice scores are computed against these same self-generated labels rather than against independent anatomical segmentations.

Editorial extensions

If this is right

Per-breast classifiers and prompt-based lesion localizers can use the released ROIs as a spatial prior instead of carving regions out of whole-breast masks.
In dynamic contrast-enhanced MRI, the first-timepoint ROI can be propagated across the temporal sequence, reducing the number of full-volume segmentations needed.
The dataset's size and harmonized RAS/NIfTI format make it a plausible pretraining resource for larger breast MRI models.
Researchers working on asymmetry, unilateral abnormalities, or mastectomy cases finally have a benchmark that does not assume both breasts are always present.
Third parties can reproduce or contest the claimed accuracy by running the released model and comparing with the released labels.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because the model and its cross-validation labels come from the same generation pipeline, the 99% Dice is a measure of self-consistency; comparing against independent expert segmentations on an external sample would test anatomical accuracy directly.
The center-of-mass splitting plane may be less reliable in asymmetric or implant cases; a natural extension is to compare it with anatomical midlines such as the sternum.
The dataset could be reused for side-specific functional analysis in DCE MRI, where left-right symmetry of enhancement is a clinically meaningful prior, though that use is not explored in the paper.
Releasing per-case confidence scores or inspection flags would let downstream users weight cases by estimated label quality, which would increase the dataset's value as a benchmark.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The dataset is the real contribution here, and it fills a genuine gap: no public breast MRI resource at this scale has explicit left-right labels. Aggregating 13,752 scans from multiple sources, harmonizing orientations, and releasing both masks and a trained nnU-Net model is useful for per-breast analysis, tumor detection, and pretraining. That part deserves credit.

The evaluation, however, does not back up the headline claim of 99% Dice. The labels were largely generated by the model itself through active learning, and the 5-fold cross-validation compares the model to those same pseudo-labels. That measures self-consistency, not anatomical accuracy. The visual inspection by three experts for every case is better than nothing, but subjective review of 13,752 3D volumes is not a quantitative validation and can miss a systematically shifted split line.

The stress-test note also lands. The COM splitting technique computes the centroid of the whole-breast mask and draws a sagittal plane through it. That works for roughly symmetric breasts, but in unilateral mastectomy or marked asymmetry the centroid sits inside the remaining or larger breast, so the plane cuts through breast tissue and creates a spurious contralateral segment. The paper explicitly cites unilateral absence as a handled case, but the described algorithm cannot handle it in a principled way. That is a concrete geometric flaw, not just an abstract worry about circularity.

So the soft spots are real and load-bearing. If the split is biased in common clinical cases, the dataset actively misleads downstream users who treat it as ground truth. The paper would be materially improved by a held-out set of manually refined labels, ideally from multiple annotators, with inter-observer agreement reported. Without that, the DSC numbers are not informative about true accuracy.

Who is this for? Researchers in breast imaging who want a large pretraining set or coarse side-specific ROIs. They will get value from the resource, but they should verify the labels before relying on them for anything quantitative. The paper is a resource paper, and the resource is worth having. It deserves a serious referee, not a desk rejection, but the referee should require external validation before publication.

Referee Report

3 major / 5 minor

Summary. The paper presents a large-scale breast MRI dataset with left/right breast segmentation labels, aggregating 13,752 scans from four public sources and harmonizing them to RAS orientation and NIfTI format. The authors train an nnU-Net in a low-resolution configuration using an initial 100-case whole-breast segmentation set, then employ active learning to propagate labels to the remaining cases. They report a 5-fold cross-validation DSC of 99.08 ± 0.48%, NSD@1mm of 99.10 ± 0.92%, and HD95 of 0.0196 ± 0.1849 mm, and release the dataset and model publicly.

Significance. If the labels are anatomically correct, this is a valuable community resource: the scale, the harmonized multi-source compilation, and the public release are concrete strengths. The paper also gives a clear divide-and-conquer motivation for side-specific ROIs. However, the central quantitative claim is currently not independently supported: the evaluation is performed against labels generated by the model itself through active learning, so the reported 99% DSC largely reflects self-consistency of the labeling process. The geometric labeling rule described for the initial 100 cases also has a known failure mode for asymmetric and post-mastectomy anatomy. With a held-out manually labeled validation set and a corrected label-generation procedure, the contribution would be much stronger.

major comments (3)

[Model Training and Active Learning; Table I] The 5-fold cross-validation is circular. The ground truth for the vast majority of the 13,752 cases consists of pseudo-labels produced by the model itself during active learning, and high-confidence predictions were added to the training set. Computing DSC, NSD, and HD95 against these same pseudo-labels measures the model's ability to reproduce the labeling process, not its agreement with expert anatomical truth. To support the claim of 'high-quality annotations', the authors should evaluate on a held-out set of cases manually segmented by expert radiologists, ideally with multiple raters, and report DSC/NSD/HD95 on that set. They should also report the provenance of each label (manual, model-refined, or raw pseudo-label) per case.
[Left-Right Breast Label Generation] The center-of-mass splitting rule is not anatomically valid for the edge cases highlighted in Figure 1. When the two breasts are asymmetric or one is absent (e.g., after mastectomy), the centroid of the whole-breast mask lies inside the larger or remaining breast, and the sagittal plane through that centroid will cut through that breast, creating a spurious contralateral region. The paper states that unilateral absence is a handled case, but the algorithm as described cannot handle it in a principled way. Visual inspection of 13,752 volumes by three experts is not a quantitative substitute for a measurement: a systematically shifted split line can look plausible on individual slices. Please provide a quantitative validation on a stratified sample of cases with mastectomy and marked asymmetry, and describe how the final labels for such cases were corrected.
[Table I] The reported HD95 values (mean 0.0196 mm, with fold means between 0.0086 and 0.0270 mm) are implausibly low for MRI volumes whose voxel spacing is at best 3x3x3 mm and often 1 mm or finer. Even a single-voxel boundary shift should produce a 95th-percentile Hausdorff distance on the order of the voxel spacing. These values suggest either a metric implementation error or that the model predictions and the pseudo-label ground truth are nearly identical. Please clarify how HD95 was computed (e.g., whether surface distances were computed in physical space and whether spacing was applied correctly) and verify the metric on a small manually labeled subset.

minor comments (5)

[Abstract] The claim 'first publicly available breast MRI dataset with explicit left and right breast segmentation labels' should be qualified with 'to our knowledge' or supported by a literature search, since such claims are difficult to verify.
[Model Training and Active Learning] The sentence about using non-contrast-enhanced T1-weighted images as pseudo-ground truth for co-registered sequences lacks details: please specify which sequences were co-registered, how registration was performed, and how many cases received labels through this transfer path.
[Materials and Methods] The 'low-resolution configuration' of nnU-Net is not described precisely; please provide the patch size, spacing, and batch size so that the training setup is reproducible.
[Results and Discussion] NSD@1mm is reported but never defined in the text; a one-sentence definition would help readers interpret the metric.
[Manual Inspection and Quality Control] No inter-rater agreement is reported for the three experts who inspected the segmentations; reporting a sample-based agreement measure would strengthen the quality-control claim.

Circularity Check

1 steps flagged · score 8.0 of 10

The reported 99% DSC is a self-consistency score: most ground-truth labels were generated by the model itself and then reused as the reference in cross-validation.

self definitional [Materials and Methods, Model Training and Active Learning; Results and Discussion, Table I]
"High-confidence predictions were incorporated into the training set for the next training round, ultimately producing a final dataset of 13,752 cases with high-quality left–right breast segmentation labels. ... The segmentation accuracy was evaluated using the Dice Similarity Coefficient (DSC), a standard metric for measuring the overlap between the predicted segmentation and the ground truth."

The final dataset used as ground truth in the 5-fold cross-validation is largely composed of predictions generated by the model itself through the active-learning loop. The model is trained on its own high-confidence outputs, and the reported DSC 99.08 ± 0.48% then measures how well the model reproduces labels produced by the same prediction pipeline. This is not an independent test of anatomical accuracy; it is a self-agreement score between the model and its own pseudo-labels. The manual visual inspection is qualitative and cannot break this statistical circularity, and no held-out, independently annotated test set is reported.

full rationale

The central quantitative claim—5-fold cross-validation DSC of 99.08%—reduces to self-agreement. The paper states that the final 13,752-case dataset was produced by incorporating the model's own high-confidence predictions into the training set, then evaluates the model against that same dataset as ground truth. This is the classic pseudo-label circularity: the reference labels are generated by the predictor being evaluated. The initial center-of-mass splitting rule adds a further concern because left/right labels are defined by a sagittal plane through the combined-mask centroid; in asymmetric or post-mastectomy cases this can split a single remaining breast, but the evaluation cannot detect such anatomical errors because the 'ground truth' is itself the output of that same geometric rule. The paper provides no external, manually labeled benchmark to break the loop. The resource may still be practically useful, but the reported performance numbers do not independently validate anatomical correctness. Score 8 reflects that the main result is forced by the self-generated labeling/evaluation loop.

Assumptions & free parameters 0 free parameters · 4 assumptions · 0 invented entities

The central contribution rests on the assumption that model-generated labels, after visual review, are accurate enough to serve as ground truth. No external benchmark or manual re-segmentation is used to validate the labels, so the dataset's reliability is an open empirical question.

assumptions (4)

domain assumption The whole-breast segmentation masks from the Duke dataset are anatomically accurate and suitable for generating left-right labels.
Used as the seed for the COM splitting and initial model training (Dataset Compilation section).
domain assumption A single sagittal plane through the center of mass of the whole breast mask correctly separates left and right breasts in all cases, including asymmetry and mastectomy.
The label generation method assumes this geometric rule; the paper acknowledges challenging cases but relies on visual inspection to correct errors.
domain assumption Cross-sequence pseudo-label transfer from non-contrast T1 images to co-registered sequences is reliable, as previously demonstrated in brain metastasis detection.
Cited as [12] and used to propagate labels; the current paper does not independently validate this on breast MRI.
domain assumption Expert visual inspection of all cases is sufficient to guarantee label quality for use as ground truth and for public release.
The quality control section describes this review but no quantitative inter-observer agreement or error rate is reported.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Divide and Conquer: A Large-Scale Dataset and Model for Left-Right Breast MRI Segmentation." pith.science (2026). https://pith.science/paper/MOWTQFXR

@misc{pith2026250713830,
  author       = {Pith},
  title        = {Pith review of: Divide and Conquer: A Large-Scale Dataset and Model for Left-Right Breast MRI Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MOWTQFXR}},
  note         = {Machine review of arXiv:2507.13830}
}

read the original abstract

We introduce the first publicly available breast MRI dataset with explicit left and right breast segmentation labels, encompassing more than 13,000 annotated cases. Alongside this dataset, we provide a robust deep-learning model trained for left-right breast segmentation. This work addresses a critical gap in breast MRI analysis and offers a valuable resource for the development of advanced tools in women's health. The dataset and trained model are publicly available at: www.github.com/MIC-DKFZ/BreastDivider

Figures

Figures reproduced from arXiv: 2507.13830 by the authors.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

13 extracted references · 10 canonical work pages

[1]

Weakly supervised 3d deep learning for breast cancer classification and localization of the lesions in mr images,

J. Zhou, L.-Y . Luo, Q. Dou, H. Chen, C. Chen, G.-J. Li, Z.-F. Jiang, and P.-A. Heng, “Weakly supervised 3d deep learning for breast cancer classification and localization of the lesions in mr images,” Journal of Magnetic Resonance Imaging , vol. 50, no. 4, pp. 1144–1151, 2019. [Online]. Available: https://onlinelibrary .wiley.com/ doi/abs/10.1002/jmri.26721

doi:10.1002/jmri.26721 2019
[2]

Classification of breast cancer in mri with multimodal fusion,

M. Morais, F. M. Calisto, C. Santiago, C. Aleluia, and J. C. Nascimento, “Classification of breast cancer in mri with multimodal fusion,” in 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI) , 2023, pp. 1–4

work page 2023
[3]

Le- sionlocator: Zero-shot universal tumor segmentation and tracking in 3d whole-body imaging,

M. Rokuss, Y . Kirchhoff, S. Akbal, B. Kovacs, S. Roy, C. Ulrich, T. Wald, L. T. Rotkopf, H.-P. Schlemmer, and K. Maier-Hein, “Le- sionlocator: Zero-shot universal tumor segmentation and tracking in 3d whole-body imaging,” in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 30 872–30 885

work page 2025
[4]

Large language model with region-guided referring and grounding for ct report generation,

Z. Chen, Y . Bie, H. Jin, and H. Chen, “Large language model with region-guided referring and grounding for ct report generation,” IEEE Transactions on Medical Imaging , 2025

work page 2025
[5]

Longitudinal segmentation of ms lesions via temporal difference weighting,

M. R. Rokuss, Y . Kirchhoff, S. Roy, B. Kovacs, C. Ulrich, T. Wald, M. Zenk, S. Denner, F. Isensee, P. V ollmuth et al. , “Longitudinal segmentation of ms lesions via temporal difference weighting,” in International Conference on Medical Image Computing and Computer- Assisted Intervention. Springer, 2024, pp. 64–74

work page 2024
[6]

How well do supervised 3d models trans- fer to medical imaging tasks?

W. Li, A. Yuille, and Z. Zhou, “How well do supervised 3d models trans- fer to medical imaging tasks?” in The Twelfth International Conference on Learning Representations

work page
[7]

Dynamic contrast- enhanced magnetic resonance images of breast cancer patients with tumor locations [data set],

A. Saha, M. R. Harowicz, L. J. Grimm, J. Weng, E. H. Cain, C. E. Kim, S. V . Ghate, R. Walsh, and M. A. Mazurowski, “Dynamic contrast- enhanced magnetic resonance images of breast cancer patients with tumor locations [data set],” The Cancer Imaging Archive, vol. 10, 2021

work page 2021
[8]

A large- scale multicenter breast cancer dce-mri benchmark dataset with expert segmentations,

L. Garrucho, K. Kushibar, C.-A. Reidel, S. Joshi, R. Osuala, A. Tsirikoglou, M. Bobowicz, J. Del Riego, A. Catanese et al., “A large- scale multicenter breast cancer dce-mri benchmark dataset with expert segmentations,” Scientific data, vol. 12, no. 1, p. 453, 2025

work page 2025

Show all 13 references

[9]

nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,

F. Isensee, P. F. Jaeger, S. A. A. Kohl, J. Petersen, and K. H. Maier-Hein, “nnu-net: a self-configuring method for deep learning-based biomedical image segmentation,” Nature Methods, vol. 18, no. 2, pp. 203–211, 2021

2021
[10]

Standard and delayed contrast-enhanced mri of malignant and benign breast lesions with histological and clinical supporting data (advanced-mri- breast-lesions) (version 2),

D. Daniels, D. Last, K. Cohen, Y . Mardor, and M. Sklair-Levy, “Standard and delayed contrast-enhanced mri of malignant and benign breast lesions with histological and clinical supporting data (advanced-mri- breast-lesions) (version 2),” https://doi .org/10.7937/C7X1-YN57, 202...

doi
[11]

Abbreviated breast mri and digital tomosyn- thesis mammography in screening women with dense breasts (ea1141),

C. Comstock, C. Gatsonis, G. Newstead, B. Snyder, I. Gareen, J. Bergin, H. Rahbar, J. Sung et al., “Abbreviated breast mri and digital tomosyn- thesis mammography in screening women with dense breasts (ea1141),” The Cancer Imaging Archive: Little Rock, AR, USA , 2023

2023
[12]

Enhancing deep learning methods for brain metastasis detection through cross-technique annotations on space mri,

T. Wald, B. Hamm, J. C. Holzschuh, R. El Shafie, A. Kudak, B. Ko- vacs, I. Pfl ¨uger, B. von Nettelbladt, C. Ulrich, M. A. Baumgartner et al., “Enhancing deep learning methods for brain metastasis detection through cross-technique annotations on space mri,” European Radiology ...

2025
[13]

nninteractive: Redefining 3d promptable segmentation,

F. Isensee, M. Rokuss, L. Kr ¨amer, S. Dinkelacker, A. Ravindran, F. Stritzke, B. Hamm, T. Wald, M. Langenberg, C. Ulrich et al. , “nninteractive: Redefining 3d promptable segmentation,” arXiv preprint arXiv:2503.08373, 2025

2025 arXiv

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

Weakly supervised 3d deep learning for breast cancer classification and localization of the lesions in mr images,

J. Zhou, L.-Y . Luo, Q. Dou, H. Chen, C. Chen, G.-J. Li, Z.-F. Jiang, and P.-A. Heng, “Weakly supervised 3d deep learning for breast cancer classification and localization of the lesions in mr images,” Journal of Magnetic Resonance Imaging , vol. 50, no. 4, pp. 1144–1151, 2019. [Online]. Available: https://onlinelibrary .wiley.com/ doi/abs/10.1002/jmri.26721

doi:10.1002/jmri.26721 2019

[2] [2]

Classification of breast cancer in mri with multimodal fusion,

M. Morais, F. M. Calisto, C. Santiago, C. Aleluia, and J. C. Nascimento, “Classification of breast cancer in mri with multimodal fusion,” in 2023 IEEE 20th International Symposium on Biomedical Imaging (ISBI) , 2023, pp. 1–4

work page 2023

[3] [3]

Le- sionlocator: Zero-shot universal tumor segmentation and tracking in 3d whole-body imaging,

M. Rokuss, Y . Kirchhoff, S. Akbal, B. Kovacs, S. Roy, C. Ulrich, T. Wald, L. T. Rotkopf, H.-P. Schlemmer, and K. Maier-Hein, “Le- sionlocator: Zero-shot universal tumor segmentation and tracking in 3d whole-body imaging,” in Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 30 872–30 885

work page 2025

[4] [4]

Large language model with region-guided referring and grounding for ct report generation,

Z. Chen, Y . Bie, H. Jin, and H. Chen, “Large language model with region-guided referring and grounding for ct report generation,” IEEE Transactions on Medical Imaging , 2025

work page 2025

[5] [5]

Longitudinal segmentation of ms lesions via temporal difference weighting,

M. R. Rokuss, Y . Kirchhoff, S. Roy, B. Kovacs, C. Ulrich, T. Wald, M. Zenk, S. Denner, F. Isensee, P. V ollmuth et al. , “Longitudinal segmentation of ms lesions via temporal difference weighting,” in International Conference on Medical Image Computing and Computer- Assisted Intervention. Springer, 2024, pp. 64–74

work page 2024

[6] [6]

How well do supervised 3d models trans- fer to medical imaging tasks?

W. Li, A. Yuille, and Z. Zhou, “How well do supervised 3d models trans- fer to medical imaging tasks?” in The Twelfth International Conference on Learning Representations

work page

[7] [7]

Dynamic contrast- enhanced magnetic resonance images of breast cancer patients with tumor locations [data set],

A. Saha, M. R. Harowicz, L. J. Grimm, J. Weng, E. H. Cain, C. E. Kim, S. V . Ghate, R. Walsh, and M. A. Mazurowski, “Dynamic contrast- enhanced magnetic resonance images of breast cancer patients with tumor locations [data set],” The Cancer Imaging Archive, vol. 10, 2021

work page 2021

[8] [8]

A large- scale multicenter breast cancer dce-mri benchmark dataset with expert segmentations,

L. Garrucho, K. Kushibar, C.-A. Reidel, S. Joshi, R. Osuala, A. Tsirikoglou, M. Bobowicz, J. Del Riego, A. Catanese et al., “A large- scale multicenter breast cancer dce-mri benchmark dataset with expert segmentations,” Scientific data, vol. 12, no. 1, p. 453, 2025

work page 2025