BreastStage: Computational Breast Cancer Staging

• BreastStage is a term for modular computational workflows that predict breast cancer stage from imaging and pathology inputs.
• It integrates diverse methods including digital pathology regression, mammographic quality control, and MRI-based anatomical segmentation.
• The ecosystem supports clinical decision making by linking stage prediction with anatomical, geometric, and surgical planning tools.

BreastStage is best understood as a context-dependent label for computational workflows that estimate breast-cancer stage or support stage-related assessment from breast imaging and pathology. In the current literature, it does not denote a single standardized public benchmark or universally adopted architecture. The closest direct realization is breast cancer stage prediction from digital pathology biopsy slides using pretrained vision models (Dossou et al., 2023). Other works contribute adjacent components that are often necessary in a staging pipeline, including mammographic positioning assessment (Tanyel et al., 2024), lesion detection and segmentation in mammography (K et al., 2020, Yan et al., 2020, Cao, 2020), left-right breast MRI decomposition (Rokuss et al., 18 Jul 2025), longitudinal MRI registration (Chen et al., 19 May 2025), breast morphology modeling (Weiherer et al., 2021), asymmetry simulation (Montenegro et al., 8 Feb 2025), and patient-specific surgical planning (Mazier et al., 2021, Nascimento et al., 2019). This suggests that “BreastStage” functions less as a single model name than as an ecosystem of breast-specific computational methods whose outputs range from ordinal stage estimates to anatomical, geometric, and surgical decision support.

1. Terminological scope and conceptual boundaries

Within the literature, BreastStage spans at least three distinct meanings. First, it refers to direct prediction of breast cancer stage from image-derived evidence, most explicitly from biopsy slide images in digital pathology (Dossou et al., 2023). Second, it can denote staged decision processes in breast imaging quality assessment, such as landmark-based mammographic positioning evaluation in mediolateral oblique views (Tanyel et al., 2024). Third, it is used more loosely for stage-adjacent infrastructure: lesion detection, breast-region segmentation, left-right anatomical separation, deformable registration, shape analysis, and surgical localization.

A central distinction is between clinical stage prediction and stage-enabling analysis. The pathology-based work predicts one of five breast cancer stages—Stage 0 through Stage 4—at biopsy level (Dossou et al., 2023). By contrast, several imaging papers improve anatomical localization, quality control, or registration without producing an oncologic stage label. Left-right MRI segmentation, for example, supports side-specific assessment and surgical planning but does not itself output a cancer stage (Rokuss et al., 18 Jul 2025). Similarly, MRI registration improves alignment of dense tissue and whole-breast structure across time points, which is useful for tracking tumor progression and treatment response but is not equivalent to staging (Chen et al., 19 May 2025).

A second distinction is between formal clinical staging and simplified risk surrogates. The MATLAB GUI system for “type and stage of Breast Tumor” discusses Stage 0 through Stage 4 and AJCC T-category size criteria in its background, but its implemented decision logic is a simplified size-threshold-based risk scheme rather than full AJCC staging (Banerjee et al., 2024). This difference is important because many BreastStage-like systems operate on partial surrogates—tumor diameter, lesion visibility, or biopsy morphology—rather than the complete TNM framework.

2. Direct stage prediction from digital pathology

The most explicit BreastStage formulation in the supplied literature is the prediction of breast cancer stage from digital pathology biopsy slides in the Nightingale Open Science Dataset of Digital Pathology. The task is not binary cancer detection; it is prediction of the cancer stage of a patient or biopsy from biopsy slide images, with the stage space defined as Stage 0, Stage 1, Stage 2, Stage 3, and Stage 4 (Dossou et al., 2023).

A biopsy is represented by multiple slide images, and each slide receives a stage prediction. The biopsy-level prediction is the mean of slide-level predictions:

$PCS(B) = \frac{1}{|B|}\sum_i PCS(S_i),$

where $B$ is a biopsy and $PCS(S_i)\in\{0,1,2,3,4\}$. Because of this averaging, the biopsy prediction is continuous in $[0,4]$ (Dossou et al., 2023). The ensemble extension averages biopsy-level outputs across models:

$PCS(B)_E = \frac{1}{|E|}\sum_{k=1}^{|E|} PCS(B)_k.$

The dataset scale reported for this task is 72,400 biopsy slide images, 4,335 breast biopsies, and 3,425 unique patients, with data observed from 2014 to 2020 (Dossou et al., 2023). Slide images are linked to cancer stage and mortality information, and the dataset also contains biopsies without stages. The original training set was split into $D_{Train}$ and $D_{Eval}$ with an 80:20 ratio, while the original evaluation set was used as $D_{Test}$. Preprocessing consisted of downsampling to $(224,224)$, randomized cropping, horizontal flipping, and normalization for training, with normalization only for validation and test.

The study fine-tuned pretrained computer vision models including ResNet-18, ResNet-50, ResNet-152, EfficientNet (M), ConvNeXt (Base), Wide ResNet-101, VGG, RegNet, Swin Transformer (B), and MaxVit, using AdamW, batch size 32, and 50 epochs (Dossou et al., 2023). The metric was Mean Square Error,

$MSE = \frac{1}{n}\sum_{j=1}^{n}(\hat y_j-y_j)^2,$

reflecting the regression-like treatment of ordinal stage. EfficientNet (M) was the best single model with $B$0. An ensemble of all models achieved $B$1, and a filtered ensemble using only models with $B$2 achieved $B$3, which was the winning solution (Dossou et al., 2023).

These results establish an important technical point: in direct BreastStage prediction, ordinal stage estimation can be operationalized as continuous biopsy-level regression aggregated from slide-level predictions. The paper also notes that Stage 2 is already potentially very advanced, whereas Stages 0 and 1 are considered early. At the same time, the workflow carries identifiable limitations. Unlabeled biopsies were assigned Cancer Stage 1 because it was “the most frequent cancer stage,” which introduces potential label noise; the evaluation reported MSE but not calibration, AUROC, per-class recall, or ordinal accuracy; and the interpretability discussion remained conceptual, proposing causal inference without implementing it (Dossou et al., 2023).

3. Simplified stage assignment and its divergence from formal oncology staging

A separate strand of BreastStage-like work uses image processing and lesion size as a surrogate for type and stage. The MATLAB GUI system processes mainly mammogram images, applies median filtering with a $B$4 neighborhood, Otsu thresholding, watershed segmentation, and morphological erosion and dilation, then uses the MATLAB imdistline object to measure tumor diameter in pixels and convert it to centimeters (Banerjee et al., 2024). On that basis, it classifies a case as healthy, benign, or malignant and assigns a risk stage.

The implemented decision rules are explicitly simplified. If tumor size is less than 1 cm and greater than 0, the system states that it has the possibility to be a benign tumor; if tumor size is more than 1 cm, it has the possibility to be a malignant tumor; otherwise it is considered a healthy breast with no lump. The stage logic is likewise size-based: within 1 cm corresponds to low risk stage, within 2 cm to low to medium risk stage, and greater than 2 cm to high risk stage (Banerjee et al., 2024).

This framework is clinically distinct from the full staging descriptions presented in the same paper’s background section. That literature review reproduces a common overview in which Stage 0 is confined to the milk duct, Stage 1 is less than 2 cm and not beyond the breast, Stage 2 is 2–5 cm and may involve lymph nodes, Stage 3 may include inflammation and chest-wall or nodal spread, and Stage 4 indicates distant spread regardless of size. It also reproduces AJCC T-category size criteria such as $B$5, $B$6, $B$7, and $B$8 invasion of chest wall or skin (Banerjee et al., 2024).

The crucial point is that the GUI does not implement this full staging system. Its stage output is a simplified risk categorization derived from measured diameter, and the paper does not provide a formal accuracy table, sensitivity, specificity, ROC analysis, or cross-validation for the proposed GUI itself (Banerjee et al., 2024). A common misconception is therefore to treat any image-derived stage label as equivalent to formal oncologic staging. The cited evidence does not support that equivalence. At most, the GUI is a proof-of-concept image-processing tool for risk-oriented reporting.

4. Mammography-based upstream components

Mammography contributes several BreastStage-adjacent functions: positioning quality control, suspicious-region localization, image-level malignancy discrimination, and full-image lesion delineation. These systems do not directly predict Stage 0–4, but they condition the reliability of downstream staging and biopsy workflows.

In positioning assessment, the relevant work explicitly states that BreastStage is not introduced as a separate public dataset name; instead, it operationalizes breast positioning assessment as a staged landmark-based decision process on mediolateral oblique images (Tanyel et al., 2024). The pipeline detects the nipple and the two endpoints of the pectoralis muscle line, then draws the posterior nipple line (PNL) from the nipple perpendicular to the pectoralis muscle. If that perpendicular intersects the pectoralis muscle inside the image boundaries, the view is treated as good positioning; otherwise it is poor positioning. The study used 1000 exams from VinDr-Mammography, yielding 2000 MLO images in an 80%/10%/10% split, and an expert radiologist labeled the nipple and pectoral line. The best model, CoordAtt UNet, achieved accuracy $B$9, specificity $PCS(S_i)\in\{0,1,2,3,4\}$0, sensitivity $PCS(S_i)\in\{0,1,2,3,4\}$1, and the smallest angular error of $PCS(S_i)\in\{0,1,2,3,4\}$2 degrees (Tanyel et al., 2024). In a BreastStage context, this is best interpreted as explainable image quality control rather than disease staging.

For lesion analysis, a two-stage multiple instance learning framework on INbreast first localizes candidate mass regions with a fully convolutional network similar to U-Net and then performs image-level benign-versus-malignant classification by attention-weighted aggregation of patch embeddings (K et al., 2020). The mammogram-level feature is

$PCS(S_i)\in\{0,1,2,3,4\}$3

with softmax-normalized attention weights. On five-fold subject-level cross-validation, the localization stage achieved precision $PCS(S_i)\in\{0,1,2,3,4\}$4 and recall $PCS(S_i)\in\{0,1,2,3,4\}$5, while the full system achieved $PCS(S_i)\in\{0,1,2,3,4\}$6 for benign-versus-malignant classification (K et al., 2020).

For full-mammogram contouring, a two-stage multi-scale system based on YOLOv3 plus multi-scale fusion followed by v19U-Net++ segmentation addressed interaction-free mass delineation from native full mammograms (Yan et al., 2020). The best reported end-to-end performance on INbreast was $PCS(S_i)\in\{0,1,2,3,4\}$7 Dice with the two-stage system plus multi-scale fusion, versus substantially worse one-stage full-image segmentation baselines (Yan et al., 2020). In parallel, the one-stage anchor-free BMassDNet used truncation normalization, CLAHE, natural deformation data augmentation, and a modified FSAF detector with ResNet101 and FPN. It reported $PCS(S_i)\in\{0,1,2,3,4\}$8 at $PCS(S_i)\in\{0,1,2,3,4\}$9 on INbreast and $[0,4]$0 at $[0,4]$1 on DDSM (Cao, 2020).

Taken together, these mammographic systems show that BreastStage often depends on an upstream stack: acquisition quality assurance, candidate-region localization, lesion delineation, and malignancy-oriented triage. A plausible implication is that direct staging models will inherit failure modes from these earlier modules whenever stage estimation is attempted from mammographic evidence alone.

5. MRI-based anatomical decomposition and longitudinal alignment

MRI-oriented BreastStage components focus on anatomical separation and temporal correspondence rather than direct stage labeling. A large-scale example is the introduction of the first publicly available breast MRI dataset with explicit left and right breast segmentation labels, containing 13,752 3D MRI scans assembled from Duke-Breast-Cancer-MRI, MAMA-MIA, Advanced-MRI-Breast-Lesions, and EA1141 (Rokuss et al., 18 Jul 2025). Labels were derived initially by center-of-mass splitting of whole-breast masks and then expanded through active learning and expert review. The model is based on nnU-Net, specifically a low-resolution configuration, and achieved mean cross-validation performance of DSC $[0,4]$2, NSD@1mm $[0,4]$3, and HD95 $[0,4]$4 mm (Rokuss et al., 18 Jul 2025).

This resource is not a staging model, but its clinical relevance to staging-related workflows is explicit. Separate left/right regions of interest support unilateral lesion localization, asymmetry analysis, breast-specific classification, prompt-based lesion localization, longitudinal analysis in DCE-MRI, and structured reporting. The paper further notes difficult edge cases such as unilateral breast absence after mastectomy, breast implants, and marked asymmetry (Rokuss et al., 18 Jul 2025). In BreastStage terminology, this is best treated as a foundational anatomical prior.

Longitudinal registration introduces another layer of stage-adjacent infrastructure. GuidedMorph uses a two-stage deformable registration design in which a first network estimates a global deformation field for whole-volume alignment and a second network refines alignment in dense tissue or masked local regions (Chen et al., 19 May 2025). The framework defines deformable registration as

$[0,4]$5

and decomposes the transform into $[0,4]$6 and $[0,4]$7. A Dual Spatial Transformer Network fuses these fields to avoid two interpolation steps, and an EDT-based mask warping method preserves fine structure. Evaluated on 100 randomly selected patients from ISPY2 and 100 from an internal dataset, the method reported improvements over the best learning-based baseline of more than $[0,4]$8 in dense tissue Dice, $[0,4]$9 in breast Dice, and $PCS(B)_E = \frac{1}{|E|}\sum_{k=1}^{|E|} PCS(B)_k.$0 in breast SSIM (Chen et al., 19 May 2025).

A sustainability-oriented extension appears in deep active learning for breast region segmentation on the Stavanger breast MRI dataset of 59 patients, using FCN-ResNet50 and Breast Anatomy Geometry-based sample selection (Narimani et al., 5 Jan 2026). The paper is explicit that it is not about breast stage classification. Its significance is procedural: it addresses how to reduce annotation burden, computational cost, and carbon footprint while preserving segmentation quality. The reported conclusion was that combining the Nearest Point strategy with 30% of the training data provided the best balance between segmentation performance, efficiency, and environmental sustainability (Narimani et al., 5 Jan 2026). This suggests that scalable BreastStage pipelines may increasingly depend on annotation-efficient infrastructure rather than brute-force labeling.

6. Shape modeling, asymmetry simulation, and surgical localization

A further branch of BreastStage concerns breast morphology, aesthetic change, and patient-specific geometry. The Regensburg Breast Shape Model is an open-access 3D statistical shape model built from 110 standing-position breast scans (Weiherer et al., 2021). After dense correspondence, generalized Procrustes analysis, and PCA, the model is written as

$PCS(B)_E = \frac{1}{|E|}\sum_{k=1}^{|E|} PCS(B)_k.$1

with shape coefficients typically constrained by $PCS(B)_E = \frac{1}{|E|}\sum_{k=1}^{|E|} PCS(B)_k.$2. A key technical contribution is the breast probability mask, which weights the registration objective to fit accurately inside the breast region while only roughly outside it, thereby reducing thoracic variance. The resulting model reported generalization $PCS(B)_E = \frac{1}{|E|}\sum_{k=1}^{|E|} PCS(B)_k.$3 mm and specificity $PCS(B)_E = \frac{1}{|E|}\sum_{k=1}^{|E|} PCS(B)_k.$4 mm; without BPMs, generalization degraded to $PCS(B)_E = \frac{1}{|E|}\sum_{k=1}^{|E|} PCS(B)_k.$5 mm (Weiherer et al., 2021). The paper explicitly notes relevance to breast shape staging or grading, quantitative shape analysis, symmetry and asymmetry assessment, and tracking of changes over time.

A complementary image-synthesis approach models postoperative asymmetry by framing asymmetry transfer as inpainting (Montenegro et al., 8 Feb 2025). The method masks a target breast in a pre-operative image, inserts explicit cues for nipple position and lower breast contour, and reconstructs the missing region using a Double U-Net/Double GAN architecture that also views the contralateral breast. On the Champalimaud Foundation dataset with 2,000 pre- and post-operative breast images and the public Breast Cosmetic dataset with 3,762 post-operative images, the Double GAN achieved on CP SSIM $PCS(B)_E = \frac{1}{|E|}\sum_{k=1}^{|E|} PCS(B)_k.$6, LPIPS $PCS(B)_E = \frac{1}{|E|}\sum_{k=1}^{|E|} PCS(B)_k.$7, PSNR $PCS(B)_E = \frac{1}{|E|}\sum_{k=1}^{|E|} PCS(B)_k.$8, and on BC SSIM $PCS(B)_E = \frac{1}{|E|}\sum_{k=1}^{|E|} PCS(B)_k.$9, LPIPS $D_{Train}$0, PSNR $D_{Train}$1 (Montenegro et al., 8 Feb 2025). Glow-based invertible models enabled annotation-free inference and achieved IoU $D_{Train}$2 on CP/BC for segmentation mode. Although this is not disease staging, it encodes treatment-related morphological change in clinically meaningful coordinates such as nipple displacement and lower contour deformation.

For preoperative planning, a rigged deformable torso model combines linear blend skinning with scalable bones and 55 blendshapes to register a patient-specific breast surface scan in seconds (Mazier et al., 2021). The final deformation is

$D_{Train}$3

Across 7 patients scanned in standing and supine poses, the registration converged in about 2 seconds on average, 3 seconds maximum, with mean absolute error $D_{Train}$4 mm for mesh registration and $D_{Train}$5 mm for breast anatomical landmarks (Mazier et al., 2021). The output is not an oncologic stage but a transferable patient-specific canvas for surgical reference patterns.

The most explicit surgical localization formulation comes from virtual mammography with Surface Evolver, which estimates tumor position in surgical coordinates from CC and MLO mammograms and patient-specific breast measurements (Nascimento et al., 2019). The final target is given in nipple-centered coordinates $D_{Train}$6, where $D_{Train}$7 is geodesic radius from the nipple, $D_{Train}$8 is phase angle, and $D_{Train}$9 is depth below the breast surface:

$D_{Eval}$0

The method relies on a layer factor $D_{Eval}$1 to characterize how close the tumor is to the skin and uses an iterative reverse-localization workflow to match simulated and observed mammographic projections (Nascimento et al., 2019). Here, BreastStage is best interpreted as spatial staging of lesion location under compression rather than pathological disease stage.

7. Limitations, misconceptions, and future directions

The primary misconception surrounding BreastStage is that it names a single, standardized system. The literature does not support that reading. One paper states explicitly that BreastStage is not introduced as a separate public dataset name in mammographic positioning assessment (Tanyel et al., 2024), while the direct pathology study uses the term only implicitly through its breast-stage prediction target (Dossou et al., 2023). The term therefore remains heterogeneous across subfields.

A second misconception is that all stage outputs are clinically equivalent. They are not. The digital pathology system predicts one of five cancer stages via biopsy-level regression (Dossou et al., 2023). The MATLAB GUI instead outputs size-based risk categories and does not implement full AJCC staging (Banerjee et al., 2024). Mammography quality, detection, segmentation, MRI separation, and registration papers typically do not predict cancer stage at all (Tanyel et al., 2024, Rokuss et al., 18 Jul 2025, Chen et al., 19 May 2025). Any comparison of “BreastStage” methods must therefore distinguish oncologic stage prediction from anatomical preprocessing and from procedural decision support.

Methodological limitations also recur. In pathology-based stage prediction, unlabeled biopsies were assigned Stage 1, there was class imbalance toward Stages 0, 1, and 2, and external validation under distribution shift was not presented (Dossou et al., 2023). In the MRI left-right segmentation resource, labels were partly generated through algorithmic splitting rather than fully manual annotation for all cases (Rokuss et al., 18 Jul 2025). In asymmetry synthesis, strong lower-breast deformation and annotation errors can produce artifacts (Montenegro et al., 8 Feb 2025). In surgical localization under compression, the model relies on simplified breast geometry and approximate deformation assumptions (Nascimento et al., 2019).

Future directions are correspondingly diverse. The pathology work proposes causal inference to improve interpretability, robustness, and trust (Dossou et al., 2023). The MRI left-right segmentation dataset is positioned as a pretraining resource for downstream tools in women’s health (Rokuss et al., 18 Jul 2025). GuidedMorph shows that dense-tissue-aware, two-stage registration can better support longitudinal breast MRI analysis (Chen et al., 19 May 2025). Active learning for breast region segmentation suggests that annotation-efficient and lower-carbon pipelines may become important design constraints (Narimani et al., 5 Jan 2026). Across these threads, the most plausible long-term interpretation is that BreastStage will evolve not into a single monolithic model, but into a modular stack that links pathology, mammography, MRI, morphology, and surgery through breast-specific anatomical priors and stage-aware prediction targets.