TLDR: BreastSegNet is a new multi-label segmentation algorithm for breast MRI that identifies nine anatomical structures: fibroglandular tissue, vessel, muscle, bone, lesion, lymph node, heart, liver, and implant. Developed using a large, expertly annotated dataset, the model, particularly the nnU-Net ResEncM variant, achieved high accuracy (average Dice score of 0.694), especially for larger structures. The code and weights are publicly available, with data release planned, aiming to enhance comprehensive quantitative breast MRI analysis.
Breast magnetic resonance imaging (MRI) is a vital tool for detecting breast cancer early and planning treatments. It offers high-resolution images that are crucial for understanding breast health. However, a significant challenge in breast MRI analysis has been the limited scope of existing segmentation methods. These methods often focus on only a few specific areas, like fibroglandular tissue or tumors, leaving out many other important anatomical structures visible in the scans. This narrow focus restricts their usefulness for detailed quantitative analysis, which is essential for advanced research and clinical applications.
To address this gap, a new study introduces BreastSegNet, a groundbreaking multi-label segmentation algorithm designed for breast MRI. This innovative model is capable of identifying and segmenting nine distinct anatomical labels: fibroglandular tissue (FGT), vessel, muscle, bone, lesion, lymph node, heart, liver, and implant. By covering such a wide range of tissues, BreastSegNet significantly enhances the utility of breast MRI for comprehensive quantitative analysis, allowing researchers to extract more complete body composition parameters, such as muscle quality and bone density, directly from breast MRI scans.
Developing the Dataset and Annotation Process
The success of any robust segmentation model relies heavily on high-quality, meticulously annotated data. For BreastSegNet, the researchers undertook an extensive manual annotation effort, creating a large dataset of 1123 MRI slices. These slices meticulously capture all nine anatomical structures. The annotation process was rigorous, involving four researchers without formal radiology training working under the close supervision and detailed review of an expert fellowship-trained breast radiologist. This iterative, model-assisted annotation workflow ensured the highest level of accuracy and consistency.
The process began with initial manual annotations of a small set of MRIs, which were then reviewed and approved by the radiologist. These initial annotations were used to train a preliminary segmentation model. This model then provided initial predictions for subsequent images, which were manually refined by annotators and re-reviewed by the radiologist. This iterative refinement process, involving multiple rounds of model development and manual correction, led to the creation of a highly accurate and curated dataset. For independent evaluation, a separate test set of 50 patient MRIs was manually annotated without model assistance, with all annotations undergoing radiologist review to serve as the ground truth for performance assessment.
Benchmarking and Performance
To identify the most effective algorithm for this complex task, the study benchmarked nine different segmentation models. These included well-known architectures like U-Net, SwinUNet, and UNet++, as well as foundation models such as fine-tuned SAM and MedSAM. Additionally, several variants of nnU-Net, specifically nnU-Net ResEncM, ResEncL, and ResEncXL, were evaluated. The nnU-Net series, known for its self-configuring capabilities and advanced CNN architectures, demonstrated superior performance in this study.
Among all the models tested, nnU-Net ResEncM emerged as the top performer, achieving the highest average Dice score of 0.694 across all nine labels. The Dice coefficient is a widely used metric that measures the overlap between the model’s predicted segmentation and the actual ground truth, with a score of 1 indicating perfect overlap. BreastSegNet, powered by nnU-Net ResEncM, showed exceptional performance on larger, more clearly defined structures such as the heart, liver, muscle, fibroglandular tissue (FGT), and bone, with Dice scores exceeding 0.73 and approaching 0.90 for heart and liver. While performance varied for smaller or more challenging regions like vessels, lesions, and implants, the model still achieved respectable scores.
The researchers acknowledge that lymph nodes presented the lowest segmentation performance among the nine labels. This limitation is attributed to several factors, including the rarity of lymph nodes in the dataset, their similar attenuation values to blood vessels, and their small size, which makes even slight mislabeling significantly impact the Dice score. Future work aims to address these challenges to further improve the model’s accuracy for these structures.
Also Read:
- OrthoInsight: AI Breakthrough in Automated Rib Fracture Diagnosis and Reporting
- New AI Model Enhances Aneurysm Blood Flow Analysis for Better Diagnostics
Public Availability and Future Impact
A significant contribution of this study is the commitment to making the research publicly accessible. All model code and pretrained weights for BreastSegNet are available on GitHub, fostering transparency and enabling other researchers to build upon this work. The researchers also plan to release the meticulously annotated dataset at a later date, which will be invaluable for advancing quantitative research in breast MRI. This public release of resources is crucial for accelerating progress in breast cancer screening, diagnosis, and personalized treatment strategies.
The development of BreastSegNet represents a substantial step forward in medical image segmentation, offering a comprehensive tool for analyzing breast MRI scans. By providing detailed, multi-label segmentation, this model paves the way for more in-depth quantitative research and potentially more accurate clinical assessments in breast imaging. For more detailed information, you can refer to the full research paper available at https://arxiv.org/pdf/2507.13604.
