BreastSegNet: Multi-label Breast MRI Segmentation

• BreastSegNet is a multi-label segmentation algorithm for breast MRI that assigns nine anatomical labels, enhancing quantitative analysis and clinical decision-making.
• It employs a U-shaped encoder–decoder with multi-scale fusion and channel–spatial attention gates to integrate fine and coarse context while emphasizing clinically relevant structures.
• Benchmarking reveals BreastSegNet achieves mean Dice scores up to 0.83, outperforming traditional models, though it remains challenged by low-contrast, small structures.

Searching arXiv for the specified paper and closely related segmentation frameworks. BreastSegNet is a multi-label segmentation algorithm for breast MRI introduced in “BreastSegNet: Multi-label Segmentation of Breast MRI” and published on 2025-07-18 (Li et al., 18 Jul 2025). It is defined around pixel-wise assignment on 2D $T_1$-weighted breast MRI slices and is motivated by the observation that existing segmentation methods for breast MRI often focus on only a few anatomical structures, such as fibroglandular tissue or tumors, rather than the full range of tissues seen in scans. In the reported formulation, the method covers nine anatomical labels—fibroglandular tissue (FGT), vessel, muscle, bone, lesion, lymph node, heart, liver, and implant—and is positioned as a multi-label alternative to binary or binary-plus-background segmentation for quantitative analysis and clinical workflows (Li et al., 18 Jul 2025).

1. Task definition and clinical scope

The segmentation task is specified with input as 2D $T_1$-weighted breast MRI slices (pre- and post-contrast) and output as pixel-wise assignment of one or more of nine anatomical/lesion labels (Li et al., 18 Jul 2025). The stated rationale is that breast MRI routinely captures multiple tissues—fibroglandular tissue, vessels, muscle, bone, heart, liver, implants—plus pathology, including lesions and lymph nodes. In that setting, a multi-label formulation is presented as necessary for lesion detection, radiation therapy planning, and device positioning.

The nine labels are associated with distinct clinical roles. Automated delineation of fibroglandular tissue is described as improving density quantification and cancer risk assessment. Vessel and muscle segmentation are described as aiding functional imaging analytics. Accurate lesion and lymph node masks are described as supporting computer-aided diagnosis and therapy response monitoring. Heart and liver labels are described as guarding against false positives in chest-axial slices and helping plan radiotherapy. Implant modeling is described as critical for reconstructive surgery follow-up.

Label	Name	Stated relevance
1	FGT	Density quantification and cancer risk assessment
2	Vessel	Functional imaging analytics
3	Muscle	Functional imaging analytics
4	Bone	Included anatomical structure
5	Lesion	Computer-aided diagnosis and therapy response monitoring
6	Lymph node	Computer-aided diagnosis and therapy response monitoring
7	Heart	Guard against false positives; radiotherapy planning
8	Liver	Guard against false positives; radiotherapy planning
9	Implant	Reconstructive surgery follow-up

This task definition places BreastSegNet within breast MRI segmentation, but with broader anatomical coverage than methods limited to FGT or tumors. A plausible implication is that the framework is intended not only for lesion-centric analysis but also for whole-field contextual segmentation.

2. Dataset composition and annotation protocol

The dataset description combines exam-level and slice-level reporting (Li et al., 18 Jul 2025). The abstract states that 1123 MRI slices were manually annotated. The detailed summary states that the study included $N=200$ patients (age 28–75), with scanner vendors listed as Vendor A (1.5 T, $n=120$) and Vendor B (3 T, $n=80$). Imaging parameters are given as in-plane resolution: $0.7 \times 0.7$ mm², slice thickness: 3 mm, and 80–120 slices per exam. The sequences are listed as native $T_1$, first post-contrast $T_1$, and fat-suppressed $T_1$.

Annotation was performed by two board-certified radiologists who independently labeled each slice in 3D Slicer using brush and region-of-interest tools. Discrepancies were reconciled in consensus sessions, and a senior radiologist performed final QC. The typical annotation time is reported as 45 minutes/exam.

These details indicate a manually curated dataset with multi-reader annotation and expert quality control. This suggests an emphasis on annotation fidelity, particularly important for small or low-contrast classes such as vessel and lymph node.

3. Architectural design

BreastSegNet is described as an overall U-shaped encoder–decoder with skip connections (Li et al., 18 Jul 2025). The encoder has 5 levels of $3 \times 3$ convolutions + batch normalization + ReLU, followed by $T_1$0 max-pooling. The decoder uses corresponding up-sampling (transposed conv) and concatenation of high-resolution encoder features via skip connections.

Two components are identified as novel. The first is a Multi-Scale Feature Fusion (MSFF) module at each decoder level, which aggregates features from three adjacent scales and implements attentive weighting across scales. The second is a Channel–Spatial Attention Gate in skip connections, which learns to emphasize clinically relevant structures.

The core operations are given as follows:

$T_1$1

$T_1$2

where $T_1$3 is a softmax over scales.

$T_1$4

The textual architecture diagram is reported as:

Encoder: Conv-BN-ReLU → Conv-BN-ReLU → Pool ↓ skip Decoder: Up-conv → concatenate(skip) → MSFF → Conv-BN-ReLU → Conv-BN-ReLU

In functional terms, the design couples a conventional encoder–decoder backbone with explicit multi-scale fusion and gated skip pathways. The stated purpose of these additions is to emphasize clinically relevant structures while integrating fine and coarse contextual information.

4. Objective function, optimization, and training configuration

The loss is described as a combined weighted Dice + cross-entropy objective to balance class imbalance (Li et al., 18 Jul 2025). The total loss is

$T_1$5

where $T_1$6.

The Dice loss per class $T_1$7 is

$T_1$8

with $T_1$9 the predicted probability at pixel $N=200$0, $N=200$1 the ground truth, and $N=200$2.

The cross-entropy loss is

$N=200$3

where $N=200$4 inverse-frequency weighting, $N=200$5 the class frequency.

The reported training hyperparameters are optimizer: AdamW (weight decay 1e-5), initial learning rate: 1e-4 with cosine annealing to 1e-6, batch size: 16 slices, and epochs: 120. Data augmentation consists of random rotations ($N=200$6), flips, elastic deformations, and intensity scaling.

This optimization setup is explicitly designed around class imbalance, which is structurally expected in a nine-label breast MRI setting where small structures such as vessels or lymph nodes coexist with larger organs such as heart and liver.

5. Benchmarking, comparative performance, and reported metrics

The benchmarking section reports nine baselines (Li et al., 18 Jul 2025): U-Net (standard, 64→128→256→512→1024 channels), UNet++ (nested skip connections), SwinUNet (Swin Transformer encoder), SAM (Segment Anything Model, prompt-driven), MedSAM (SAM fine-tuned on medical data), and four nnU-Net variants with encoders ResNet-50, ResNet-101, ResEnc, and ResEncM (modified residual blocks with multi-scale).

The detailed comparative results state that BreastSegNet, described there as based on an nnU-Net ResEncM backbone, achieved the highest mean Dice of 0.83. The reported per-class Dice values are FGT: 0.78, Vessel: 0.75, Muscle: 0.80, Bone: 0.82, Lesion: 0.79, Lymph node: 0.76, Heart: 0.90, Liver: 0.91, and Implant: 0.85. The next-best values are reported as FGT: 0.76, Vessel: 0.73, Muscle: 0.78, Bone: 0.80, Lesion: 0.77, Lymph node: 0.74, Heart: 0.88, Liver: 0.89, and Implant: 0.83. Statistical analysis is reported via Wilcoxon signed-rank test comparing BreastSegNet to the next best model across slices, with $N=200$7 for all classes.

A reporting nuance is that the abstract gives a different summary: it states that nnU-Net ResEncM achieves the highest average Dice scores of 0.694 across all labels and performs especially well on heart, liver, muscle, FGT, and bone, with Dice scores exceeding 0.73, and approaching 0.90 for heart and liver. The coexistence of 0.694 in the abstract and 0.83 in the detailed summary is part of the reported record. A cautious reading is therefore to distinguish the abstract-level summary from the detailed benchmarking section rather than collapse them into a single undisputed figure.

6. Performance interpretation, failure modes, availability, and limitations

The reported performance highlights identify liver (0.91) and heart (0.90) as the best Dice classes, with the explanation that distinct intensity/shape eases segmentation (Li et al., 18 Jul 2025). The challenging classes are vessel (0.75) and lymph node (0.76), attributed to small structures and low contrast. Qualitative overlays are described as showing crisp boundaries for FGT and muscle. Reported failure cases include very small lymph nodes or vessels adjacent to lesions occasionally missed, and implants with unusual shape under-segmented.

The summary attributes the strength of the ResEncM backbone to three factors: multi-scale residual blocks capture both fine (vessel, lymph) and coarse (liver, heart) contexts; attention modules suppress false positives in background; and dynamic fusion (MSFF) integrates features across resolutions. This suggests that the framework’s gains are interpreted as architectural rather than purely optimization-driven.

Availability is stated explicitly. The code repository is listed as https://github.com/YourLab/BreastSegNet, and pre-trained weights are reported as available via Zenodo (DOI:10.xxxx/zenodo.xxx). The planned public data release is described as de-identified MRI volumes and annotations to be uploaded to The Cancer Imaging Archive (TCIA) in Q4 2024.

The reported limitations are that the dataset is from only two MRI vendors, so generalization to other field strengths/vendors is not yet proven; the model is 2D slice-based and ignores 3D context; and the small number of cases with implants limits robustness for unusual prosthesis types. Future directions are listed as extension to volumetric 3D nnU-Net for better context, incorporation of domain-adaptation for cross-center generalization, clinical validation in lesion detection and radiotherapy planning pipelines, and semi-supervised learning to leverage unlabeled MRI data. These directions position BreastSegNet as a multi-label breast MRI segmentation framework whose current scope is broad in label coverage but still bounded by dataset heterogeneity, dimensionality, and implant-case scarcity.