Tumor grade is one of the most important prognostic factors in breast cancer. It reflects how abnormal cancer cells look compared to normal cells and how quickly the tumor is likely to grow. The widely used Nottingham grading system evaluates three features: nuclear pleomorphism (how irregular the cell nuclei appear), tubule formation (how much the cancer cells form tube-like structures resembling normal glands), and mitotic rate (how quickly cells are dividing). Each component is scored 1 to 3, and the total determines whether the tumor is low, intermediate, or high grade.
Despite its clinical importance, breast cancer grading suffers from significant inter-observer variability between pathologists. Studies show that different pathologists examining the same slide can assign different grades, reducing reproducibility and potentially affecting treatment decisions. Deep learning-based automated grading offers the possibility of consistent, objective grading that does not vary from pathologist to pathologist.
This study trained a deep learning model to classify breast cancer as low/intermediate versus high grade directly from whole slide images (WSIs), without requiring any hand-drawn annotations of tumor regions or individual cells. Using the MIL (multiple instance learning) framework, the model was trained using only the overall tumor grade as a label, derived from the existing pathology report.
The model was developed and validated on 1,392 young (under 40 years) breast cancer patients from the Dutch PARADIGM cohort, achieving a Cohen's Kappa of 0.59 (accuracy 80%) against expert pathologist grading, and demonstrating statistically significant separation of survival outcomes between the model-predicted grade groups.
Breast cancer in young women under 40 tends to be more biologically aggressive and carries a worse prognosis than breast cancer in older women. Because of this, many treatment guidelines recommend systemic therapy for almost all young patients. However, some patients with lower-risk tumors could be managed with local treatment alone, and systemic therapy would represent overtreatment that exposes them to serious side effects unnecessarily. Accurate tumor grading is therefore essential to avoid both under- and overtreatment in this population.
The Nottingham grading system relies on visual assessment by trained pathologists, which is inherently subjective. Large nationwide studies have documented significant inter- and intra-laboratory variation in breast cancer grading, including a Dutch study of over 33,000 patients that found substantial inconsistency across centers. Standardizing grade assessment through deep learning could reduce this variability and improve the reliability of grade as a prognostic marker.
Early automated grading methods relied on hand-crafted features such as textural statistics and morphological measurements, which required careful feature engineering and often performed poorly across institutions. More recent deep learning approaches have overcome feature engineering requirements but typically depended on labor-intensive pixel-level or cell-level annotations of specific tissue regions, which are impractical to obtain at scale from routine clinical datasets.
Multiple instance learning (MIL) resolved the annotation bottleneck by enabling models to train from slide-level labels alone, without knowing which specific regions within the slide determine the grade. Under the MIL assumption, a high-grade slide must contain at least one high-grade tile somewhere, while a low-grade slide contains no high-grade tiles. This weak supervision paradigm makes it feasible to train grading models from existing pathology reports without additional annotation effort.
The deep learning model used a ResNet-34 backbone pretrained on ImageNet. Each whole slide image was divided into 224x224 pixel tiles extracted at 10x magnification, a resolution that captures both individual nuclei and larger tissue architecture simultaneously. Tiles containing mostly white background were automatically filtered out, leaving only tissue-containing tiles for analysis.
Training followed a two-step iterative process. In the inference step, 512 randomly selected tiles per slide were passed through the network to compute each tile's probability of being high grade. The top 5 tiles with the highest predicted high-grade probability were then selected. In the training step, these 5 tiles were used to update the model, each assigned the same grade label as the whole slide they came from. This iterative selection progressively focuses the model on the most diagnostically relevant tissue areas without requiring a pathologist to identify them.
Three model variants were compared using a multi-task learning (MTL) approach: one trained on overall tumor grade alone, one trained simultaneously on tumor grade and all three component grades (nuclear, tubular, and mitotic), and one trained on tumor grade, component grades, and hormone receptor and HER2 status. MTL models share all layers except a final task-specific layer for each output, and the total loss is the sum of individual task losses. Adding auxiliary tasks can improve the quality of learned representations by encouraging the model to capture multiple aspects of the tumor simultaneously.
Data augmentation was applied to reduce overfitting and improve robustness to staining variability, including random 90-degree rotations, flips, and adjustments to brightness, contrast, hue, and saturation. The best-performing model across 300 training iterations was retained based on Cohen's Kappa on the internal validation set, with final evaluation performed on the fully independent test set to ensure unbiased performance estimates.
The MTL model trained on tumor grade and its three grade components achieved the best performance, with a Cohen's Kappa of 0.59 and accuracy of 80% on the independent test set of 686 patients. For comparison, published inter-pathologist agreement for the same binary low/intermediate versus high grade task on a different dataset was kappa 0.78, indicating that the model performs at a level roughly consistent with a less-experienced pathologist rather than matching specialist-level agreement.
Agreement was highest for the mitotic component (kappa 0.53), while tubular formation scored lowest (kappa 0.35). The authors attribute the poor tubular score to a fundamental mismatch between the MIL framework and the task: tubule formation is scored as a percentage of the entire tumor area, but the model evaluates only the 5 highest-grade tiles. This local window is sufficient for nuclear and mitotic assessment but insufficient for evaluating a property that requires estimating the global proportion across the whole tumor.
Kaplan-Meier survival curves for model-predicted low/intermediate versus high grade groups showed significantly different overall survival, distant recurrence-free survival, and recurrence-free survival (all p less than 0.05). Eight-year overall survival rates were 86.4% for the model-predicted low/intermediate grade group versus 72.8% for the high-grade group. These differences were comparable to those observed for pathologist-assigned grades, demonstrating that the model captures biologically meaningful prognostic information.
In univariate Cox regression, model-predicted grade was significantly associated with all three survival endpoints. However, after adjusting for tumor size, lymphovascular invasion, locoregional treatment, and stratifying for molecular subtype in multivariate analysis, the model's hazard ratios showed only a non-significant trend while pathologist grades retained significance for overall and distant recurrence-free survival. This suggests the model captures some but not all of the prognostic information contained in expert pathologist grading.
The model's primary practical advantage over prior automated grading approaches is that it works directly on entire whole slide images without requiring tumor region selection or pixel-level annotations. This weakly supervised, annotation-free design means the model can be directly applied to any new digitized slide without additional preparation, enabling straightforward integration into digital pathology workflows.
An important limitation is that the model can only distinguish low/intermediate from high grade but cannot separate grade 1 from grade 2 tumors. Grade 1 tumors have a significantly more favorable prognosis than grade 2 tumors, and this clinically important distinction is lost in the binary classification framework. Extending the model to three-class grading would require either stronger supervision or a different learning framework capable of capturing the more subtle morphological differences between grade 1 and grade 2.
The study was conducted on a relatively homogeneous patient population: young Dutch women with invasive ductal carcinoma, stained and scanned at a single institute using two specific scanners. Whether the model generalizes to older patients, different ethnic populations, different staining protocols, or different scanner types remains untested. Scanner variability is a known source of performance degradation in deep pathology models, and the study's independence was based on training/testing data split rather than cross-institution validation.
Interpretability tools including tile-level heatmaps and visualization of the top 5 selected tiles were developed to allow pathologists to inspect which tissue regions drove the model's grade assignment. Importantly, even though no tumor annotations were used during training, the model spontaneously learned to focus on tumor regions rather than stroma or fat tissue. This emergent attention is a positive sign for model trustworthiness and a requirement for clinical adoption as a second-reader tool.
This study demonstrates that deep learning can automate breast cancer grading from whole slide images using only slide-level grade labels, without any hand-drawn tumor annotations or region selection by pathologists. The model achieves 80% accuracy and produces survival stratification that is significant in univariate analysis, establishing a proof of concept for automated grading that can be further developed toward clinical integration.
The non-significant trend in multivariate survival analysis indicates the model is not yet ready to replace pathologist grading in clinical practice. Future work should focus on extending the model to full three-class grading, training on more diverse multi-institution datasets to improve generalizability, and using consensus grades from multiple pathologists as training labels to reduce noise from individual-pathologist variability in the ground truth.
An alternative direction is to train models directly on survival endpoints rather than pathologist-assigned grade labels, using Cox proportional hazards loss functions. Survival-optimized models could potentially capture prognostic morphological signals that correlate with patient outcomes even when they do not align perfectly with the subjective visual criteria that pathologists use for grading. This approach could ultimately produce more powerful prognostic tools from histopathology images than grade-prediction models allow.
As digital pathology adoption expands globally, automated grading systems like this one could help standardize breast cancer assessment across institutions with varying levels of pathology expertise, reduce diagnostic variability, and flag cases with borderline morphology for priority expert review. The combination of interpretability, annotation-free training, and demonstrated survival stratification positions this model as a meaningful step toward clinically useful AI pathology tools for breast cancer.