Breast cancer is the most frequently diagnosed cancer among women worldwide. Accurate histological grading is critical for treatment decisions, as it provides information about how aggressive the tumor is likely to be. The standard method, the Nottingham Grading System (NGS), has been in use for over 50 years and scores tumors based on tubule formation, nuclear pleomorphism, and mitotic activity.
Despite its long track record, the NGS relies heavily on subjective visual interpretation by pathologists, which leads to significant interobserver variability. Two pathologists examining the same tumor slide may assign different grades, leading to inconsistent treatment recommendations and potentially suboptimal patient outcomes.
The emergence of digital pathology and whole slide imaging (WSI) technology has opened new possibilities for standardizing the grading process. When tissue sections are digitized at high resolution, the resulting images can be analyzed computationally, removing some of the variability introduced by different microscopes, optical lenses, and individual observer thresholds.
Artificial intelligence and machine learning, particularly convolutional neural networks (CNNs), offer the possibility of automating this analysis in a way that is objective, consistent, and scalable. The central question this study addresses is whether a deep learning model can grade breast ductal carcinoma in a way that meaningfully agrees with trained pathologists.
The study used 100 whole slide images (WSIs) of infiltrating ductal adenocarcinoma of the breast (DAC-NOS), randomly selected from the publicly available TCGA-BRCA dataset (The Cancer Genome Atlas Breast Invasive Carcinoma dataset). Patients ranged in age from 27 to 90, and all tumor stages were included. Eighteen of the 100 cases had no Nottingham grade recorded in the original pathology reports, leaving 82 graded cases for comparison.
Four general pathologists, none specialized in breast pathology, were trained in a standardized session to grade all 100 WSIs using QuPath, an open-source digital pathology software. They independently evaluated each case in a standardized region of interest (ROI) of 0.374 mm2, equivalent to 10 high-power microscope fields, which is the standard for mitotic counting under NGS.
A separate set of 10 ambassador WSIs (not included in the test set) was used to train the deep learning model. An experienced thoracic pathologist performed manual segmentation of these 10 slides, annotating tissue regions, normal versus tumor tissue, solid and tubular components, and individual pleomorphic cells and mitotic figures.
Agreement between pathologists and between the AI model and human assessors was measured using Cohen's kappa coefficient, a statistical metric that accounts for chance agreement. Kappa values were interpreted as: slight (below 0.2), fair (0.21-0.40), moderate (0.41-0.60), good (0.61-0.8), and very good (0.81-1.0).
The CNN-based algorithm was developed using Aiforia Create, a cloud-based platform for computational pathology. The model comprised multiple nested and sequentially operating CNNs, each focused on the output of the previous layer, mimicking the hierarchical way a pathologist progressively zooms in and identifies features from tissue architecture down to individual cell morphology.
The model was structured in layers: the outermost layer detected total tissue area, a second layer distinguished normal from tumor tissue, a third layer classified tumor into solid and tubular components, and a final layer used object detection to identify individual mitotic figures and pleomorphic cells. This hierarchical approach allowed the model to derive a final Nottingham grade based on its segmentation and detection outputs at each level.
Training was performed with extensive image augmentation to compensate for the small number of training slides. Augmentations included random scaling, rotation, shearing, luminance and contrast changes, noise addition, and JPEG compression artifacts. These artificial variations help the model generalize to the wide range of staining qualities and imaging conditions encountered in real clinical practice.
The entire training process was completed in approximately 17.5 hours over 19,000 iterations. Performance on the training set was excellent, with precision, sensitivity, and F1 scores above 94% for all tissue elements, reaching up to 99.85% for whole-tissue detection.
Among the four human pathologists, agreement on tubule formation was moderate, with Cohen's kappa values ranging from 0.41 to 0.49. This is consistent with published literature suggesting that glandular structure is the most reproducible of the three NGS components when pathologists are given standardized guidance.
Agreement on nuclear pleomorphism was lower, ranging from slight to fair (kappa 0.096 to 0.24). Nuclear features such as size variation and irregular chromatin patterns are inherently subjective assessments, and even trained pathologists often disagree significantly on how to score borderline cases.
Surprisingly, mitotic count showed the lowest interobserver agreement among the four pathologists, with kappa values ranging from 0.005 to 0.29. The authors attribute this to the absence of guidance specifying which area of the tumor to count, meaning each pathologist independently selected their region, introducing substantial variability in the baseline mitotic activity seen.
For overall Nottingham grade, pairwise agreement ranged from slight to moderate (kappa 0.075 to 0.5). The best agreement was between pathologists P2 and P4 (kappa 0.5), while the worst was between P1 and P2 (kappa 0.075), illustrating just how inconsistent grade assignments can be even among trained clinicians examining the same digital images.
The agreement between the deep learning model and human pathologists was lower than the agreement among pathologists themselves. For mitotic count, the AI-human agreement was slight to fair (kappa 0.016 to 0.24), making it paradoxically the highest-agreement component for the AI-human comparison, even though it was the most variable among pathologists alone.
For tubule formation, agreement between the AI model and pathologists was slight (kappa up to 0.19), below the moderate interobserver agreement seen among the human evaluators. For nuclear pleomorphism, agreement was essentially absent, with kappa values near zero and even slightly negative for some pathologist-AI pairs, indicating no meaningful concordance.
In terms of overall Nottingham grade, only one pathologist (P1) achieved fair agreement with the AI model (kappa 0.21). The other three pathologists showed only slight agreement with the model (kappa 0.11 to 0.16). These results suggest that the current model, trained on only 10 slides, has not yet reached a level of performance that could substitute for human grading.
Notably, the AI model consistently tended to assign higher grades more frequently than the human pathologists, especially in severe cases. For lower-grade cases, however, the model's assessments more closely aligned with the average human assessment. This asymmetric behavior may reflect training data characteristics or deliberate calibration toward sensitivity for high-grade disease.
The study's findings confirm that interobserver variability in breast cancer grading is a genuine and persistent clinical problem, even when standardized digital slides and training protocols are used. The results align with prior literature showing that mitotic counting, in particular, is highly sensitive to the region of tumor selected for analysis.
A key limitation acknowledged by the authors is that none of the four grading pathologists were breast pathology specialists. Research has shown that pathologists without subspecialty training in breast disease tend to underscore tumors, which may have influenced the interobserver comparisons and the apparent lower agreement with the AI model, which was calibrated toward higher grades.
Another significant methodological limitation is that the AI model and the human pathologists did not evaluate identical regions of interest. The AI analyzed an ROI selected by the thoracic pathologist who trained it, while each of the four grading pathologists independently chose their own ROI. This mismatch means that differences in agreement could partly reflect ROI selection rather than true grading disagreement.
The training dataset of only 10 slides is very small for a deep learning application. While the model achieved impressive segmentation metrics on its training set, the diversity of morphological features across 100 test slides was almost certainly not adequately captured. Training on a far larger and more diverse set of slides would be necessary before clinical deployment could be considered.
This study demonstrates both the promise and current limitations of AI for automated breast cancer grading. The deep learning model successfully trained on whole slide images to evaluate all three NGS components simultaneously, which is an advance over prior work that typically addressed only individual components in isolation.
The modest agreement between the AI model and human pathologists, ranging from slight to fair, should not be interpreted as a failure of the approach but rather as a baseline from which to improve. The agreement is comparable to the inter-pathologist agreement seen in the study, suggesting the model captures some genuine grading signal even if not yet at clinical accuracy.
Future studies should involve breast pathology subspecialists as the human comparators, use larger and more diverse training datasets, and establish standardized ROI selection protocols to reduce confounding variables. Expanding the training dataset from 10 to hundreds or thousands of slides would likely substantially improve AI performance and AI-human concordance.
The broader significance of this work is its contribution to an emerging body of evidence that AI has a role to play in reducing the subjectivity of cancer grading. Even at this early stage, the approach raises important questions about how AI mechanisms should be trained, validated, and integrated into pathology workflows to improve diagnostic consistency.