BreastDivider: Left-Right Segmentation in Breast MRI

• BreastDivider is a public dataset and pretrained nnU-Net model that provides explicit left and right segmentation in 3D breast MRI scans.
• It standardizes 13,752 heterogeneous MRI scans from multiple modalities by converting whole-breast masks into side-specific labels using a centroid-based splitting and active learning curation.
• Key performance metrics, including DSC of approximately 99.1% and robust boundary agreement, underscore its value for unilateral disease detection and advanced clinical workflows.

BreastDivider is a publicly available breast MRI dataset and pretrained segmentation model for explicit left-right breast delineation in 3D MRI volumes. It was introduced to address the absence of large-scale, side-specific breast segmentation labels in public breast MRI resources, and it couples a curated dataset of 13,752 annotated scans with an nnU-Net baseline trained for left-right segmentation. The resource is positioned as infrastructure for anatomically informed breast MRI analysis, particularly in workflows where each breast must be processed independently rather than as a single whole-breast region of interest (Rokuss et al., 18 Jul 2025).

1. Definition and scope

BreastDivider denotes both a dataset and a trained deep-learning model for segmenting the left and right breasts separately in breast MRI. Its central problem formulation is narrower than generic whole-breast segmentation: instead of producing a single breast mask, it produces two anatomical labels, one for the left breast and one for the right breast. The project is described as the first publicly available, large-scale breast MRI dataset with explicit left and right breast segmentation labels, thereby filling a specific gap left by prior public resources that either provided no breast masks or only whole-breast masks without side-specific separation (Rokuss et al., 18 Jul 2025).

The practical motivation is clinical and algorithmic. Breast MRI is used not only for tumor identification but also for detection, diagnosis, treatment planning, and longitudinal monitoring. In those settings, side-specific separation supports a “divide and conquer” workflow in which the two breasts are processed independently. This is directly relevant to unilateral abnormalities, asymmetry analysis, prompt-based lesion localization, structured reporting, and cases in which one breast is absent after mastectomy. The significance of BreastDivider is therefore not simply that it enlarges an annotation corpus, but that it formalizes per-breast anatomy as a first-class segmentation target.

2. Dataset composition and standardization

The BreastDivider dataset aggregates 13,752 3D MRI scans from multiple public sources: Duke-Breast-Cancer-MRI, MAMA-MIA, Advanced-MRI-Breast-Lesions, and EA1141. The imaging is intentionally heterogeneous and includes common MRI modalities such as T1-weighted, T2-weighted, T1 with contrast, and FLAIR. The source collections also contribute additional heterogeneity: Advanced-MRI-Breast-Lesions includes T1 DCE with five fat-saturated phases, delayed T1, T2, and T1 non-fat-saturated sequences; Duke contains pre-operative DCE MRI at 1.5T or 3T with pre- and post-contrast series; and EA1141 includes pre-contrast T1, first post-contrast T1, T2, and DWI, sometimes with a second post-contrast T1 (Rokuss et al., 18 Jul 2025).

To impose a minimum quality threshold, only volumes with at least 32 slices per axis and resolution of $3\times3\times3$ mm or finer were retained. All data were standardized into a common RAS orientation and converted to NIfTI. This standardization materially reduces preprocessing burden for downstream use and makes the dataset suitable for direct benchmarking, pretraining, or pipeline integration. Before BreastDivider, the field had resources such as Duke-Breast-Cancer-MRI and MAMA-MIA with whole-breast masks, but not a large public corpus of explicit side-specific labels.

3. Label generation and curation workflow

The label-generation process began with 100 Duke cases that already had whole-breast segmentation labels. These masks were converted into left-right labels through a center-of-mass splitting procedure. For each 3D volume, the centroid of the segmented breast tissue was computed, and the sagittal plane passing through that centroid was used to divide the original mask into left and right parts. The paper describes this as a voxel-wise coordinate averaging procedure followed by voxel assignment according to which side of the sagittal plane each voxel occupies. The generated labels were then visually checked for anatomical plausibility before being used for model development (Rokuss et al., 18 Jul 2025).

After this seed stage, the dataset was expanded through active learning. A model trained on the initial 100-case set produced predictions that were used as pseudo-ground truth for co-registered sequences. The workflow then iteratively applied the model to unlabeled cases, inspected predictions, retained high-confidence examples, refined poor ones, and retrained in loops until the final 13,752-case set was obtained. Every predicted segmentation was reviewed by three experts, with the dataset divided evenly so that 100% of cases were inspected. Cases were accepted, discarded if scan quality or image integrity was insufficient, or flagged for refinement. Refinement was performed either manually with nnInteractive or by returning the case to the active learning loop.

This hybrid curation strategy is one of the defining features of BreastDivider. It combines automatic scalability with explicit anatomical review, producing a resource that is neither purely synthetic nor purely manually traced from scratch. The dataset’s scale is therefore inseparable from its curation design.

4. Segmentation model and reported performance

The released baseline model is based on nnU-Net in a low-resolution configuration. The low-resolution setup was selected so that both breasts could fit into the network patch size even for high-resolution scans. Training began with the 100-case seed set and expanded through the active-learning procedure described above. Final evaluation used 5-fold cross-validation, with each fold serving once as validation and the remaining four folds as training (Rokuss et al., 18 Jul 2025).

The task itself is defined operationally rather than through a bespoke analytical objective: a breast mask is split into two anatomical labels by the sagittal plane passing through the breast-tissue centroid, and model predictions are evaluated with standard segmentation metrics. The paper reports Dice Similarity Coefficient (DSC), NSD@1mm, and HD95.

Metric	Mean across folds
DSC	99.08 ± 0.48%
NSD@1mm	99.10 ± 0.92%
HD95	0.0196 ± 0.1849 mm

Fold-level results are similarly high, with DSC around 99.05–99.11% and NSD@1mm around 99.07–99.15%. Within the curated benchmark, this suggests that the combination of standardized data, centroid-derived initialization, iterative refinement, and nnU-Net yields extremely accurate delineation of left and right breasts. The reported use of NSD@1mm and HD95 is also important because it indicates that performance was assessed not only by overlap but also by boundary agreement and outlier surface error.

5. Downstream role in breast MRI analysis

BreastDivider is best understood as an infrastructural resource for workflows that require per-breast localization. The paper explicitly identifies unilateral disease detection, asymmetry analysis, lesion localization, structured reporting, and preprocessing for classification systems as relevant use cases. It also notes that the pretrained model can be used directly as a breast ROI extractor or adapted as an initialization point for downstream breast MRI pipelines (Rokuss et al., 18 Jul 2025).

Its role is distinct from that of other breast MRI resources. “A Multi-Institutional Open-Source Benchmark Dataset for Breast Cancer Clinical Decision Support using Synthetic Correlated Diffusion Imaging Data” (Tai et al., 2023) provides volumetric CDI$^s$ breast cancer imaging with oncology-relevant annotations such as lesion type, genetic subtype, MRLD, SBR grade, and pCR, but its stated purpose is clinical decision support rather than left-right breast segmentation. “Comparative Analysis of Deep Learning Architectures for Breast Region Segmentation with a Novel Breast Boundary Proposal” (Narimani et al., 2024) addresses whole-breast boundary definition in DCE-MRI, including anterior and posterior boundary rules and comparison of UNet, UNet++, DenseNet, FCNResNet, and DeepLabv3ResNet variants, but it does not target large-scale side-specific labeling. A plausible implication is that BreastDivider supplies a preprocessing layer that such resources and tasks previously lacked.

The same distinction applies to breast MRI registration. “GuidedMorph: Two-Stage Deformable Registration for Breast MRI” (Chen et al., 19 May 2025) uses breast or dense-tissue masks to guide deformable alignment and preserve fine structural detail. This suggests that explicit side-specific masks could be useful whenever registration, longitudinal comparison, or representation learning benefits from separating global bilateral anatomy into independent breast-specific regions.

6. Relation to broader “breast division” literature and limitations

In the broader imaging literature, “breast division” has often meant something different from BreastDivider’s left-right MRI segmentation. In mammography, the term usually refers to background suppression, pectoral muscle removal, ROI extraction, or dense-versus-fat tissue partitioning. Examples include histogram-based 8-neighborhood connected component labeling for breast region extraction and pectoral removal in MLO mammograms (Boss et al., 2013), the anatomically grounded AEPm pipeline for automatic pectoral-muscle elimination (Ayala-Godoy et al., 2020), hybrid CCL/fuzzy/straight-line suppression methods (Aroquiaraj et al., 2014), and the two-stage segmentation-plus-density pipeline of Deep-LIBRA (Maghsoudi et al., 2020). BreastDivider departs from this tradition by treating the left-right split in 3D breast MRI as the primary target rather than a secondary preprocessing step.

Its limitations follow directly from its construction. The labels are not all manually drawn from scratch; they are largely bootstrapped from a centroid-based splitting rule and then refined. The paper therefore identifies unusual anatomy, implants, severe asymmetry, and post-mastectomy changes as challenging edge cases. The dataset is also compiled from heterogeneous public sources rather than a single standardized acquisition protocol. That heterogeneity is valuable for breadth and generalization, but it also introduces domain variation that downstream users must consider (Rokuss et al., 18 Jul 2025).

Despite these limitations, BreastDivider constitutes a standardized public benchmark for an anatomically specific segmentation problem that had previously been underserved. By releasing both the dataset and the trained model through the project repository at www.github.com/MIC-DKFZ/BreastDivider, it establishes explicit left-right breast segmentation as a reusable component in breast MRI analysis rather than an ad hoc internal preprocessing step.