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

Maximilian Rokuss, Benjamin Hamm, Yannick Kirchhoff, Klaus H. Maier-HeinarXiv 2025 · DOI · arXiv

Abstract

Automated analysis of breast MRI often struggles to reliably distinguish between left and right breasts, a critical requirement for clinical workflows involving symmetry comparison and lesion tracking. This paper introduces BreastDivider, the largest publicly available breast MRI dataset to date, containing over 13,000 annotated 3D volumes. By employing a multi-stage active learning pipeline with expert refinement, the authors developed a side-specific segmentation model that achieves near-perfect accuracy, establishing a new gold standard for the field.

Introduction

In clinical breast imaging, "close enough" is insufficient for high-precision diagnostic tasks. While automated segmentation has advanced, most existing algorithms treat the chest as a single anatomical block. This lack of side-specificity hinders tasks such as comparing symmetry, tracking lesions over time, or generating structured reports. Furthermore, standard models often fail in complex clinical scenarios, such as unilateral mastectomies or the presence of breast implants.

By failing to "divide and conquer" the anatomy, models lose the spatial priors necessary for high-precision downstream tasks like tumor detection. This work addresses this gap by providing explicit, side-specific labels for left and right breasts alongside a robust deep-learning model.

Methods

The researchers developed a multi-stage pipeline to create a high-quality, large-scale resource:

  1. Standardization: Scans from four major public sources were aggregated and harmonized into a consistent RAS (Right-Anterior-Superior) orientation and NIfTI format.
  2. Initial Labeling: Starting with 100 cases, the team applied a Center-of-Mass (COM) splitting technique. By calculating the centroid of the breast tissue and identifying the sagittal splitting plane, they automatically separated left from right.
  3. Active Learning: To scale to over 13,000 cases, an iterative active learning strategy was used. An nnU-Net model trained on the initial set generated "pseudo-ground truth" labels for the remaining data.
  4. Expert Refinement: A team of three experts manually inspected 100% of the 13,752 cases, refining or discarding segmentations to ensure a "gold standard" level of quality.
Visual examples of side-specific breast segmentation across diverse patient anatomies.

Datasets

The study introduces a massive compiled dataset by aggregating four primary sources:

  • Duke-Breast-Cancer-MRI: Single-center DCE-MRI data.
  • MAMA-MIA: Multicenter benchmark dataset.
  • Advanced-MRI-Breast-Lesions: Single-institution data including T1, T2, and FLAIR sequences.
  • EA1141: A large multicenter trial dataset (48 clinical sites) featuring abbreviated protocols.

The final collection comprises 13,752 3D MRI volumes, making it one of the largest segmentation datasets in medical imaging. The dataset specifically includes challenging cases like breast asymmetry, implants, and post-mastectomy scans, ensuring the model is "battle-hardened" for real-world clinical use.

Key Results

The performance of the trained nnU-Net model was evaluated using 5-fold cross-validation, yielding remarkably high and consistent metrics across all folds. The model achieved a Dice Similarity Coefficient (DSC) of 99.08% ± 0.48, indicating almost perfect overlap with ground truth. Boundary precision was equally high, with a Normalized Surface Dice (NSD@1mm) of 99.10% ± 0.92.

Performance metrics of the nnU-Net model across 5-fold cross-validation.

Notably, the 95th Percentile Hausdorff Distance (HD95) was recorded at 0.0196 mm, demonstrating that even the most significant errors in boundary placement were clinically negligible.

Discussion and Future Work

The authors argue that explicit left-right labels are a foundational asset for "temporally informed segmentation," where a reliable Region of Interest (ROI) from a baseline scan acts as a spatial prior for subsequent contrast-enhanced images. Because of its sheer size and anatomical variety, this dataset is also a prime candidate for pretraining large-scale foundational models in medical AI.

While the segmentation of breast tissue itself is nearly solved by this work, the next step involves integrating these high-quality ROIs into downstream tasks like automated tumor detection, classification, and longitudinal tracking. The dataset and model are publicly available at github.com/MIC-DKFZ/BreastDivider.

Summary updated

BibTeX
@misc{rokuss2025divide,
  author = {Maximilian Rokuss and Benjamin Hamm and Yannick Kirchhoff and Klaus H. Maier-Hein},
  title = {{Divide and Conquer: A Large-Scale Dataset and Model for Left-Right Breast MRI Segmentation}},
  year = {2025},
  eprint = {2507.13830},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.13830},
}
Figure