Invasive Ductal Carcinoma (IDC) is the most common form of breast cancer, accounting for over 80% of all cases. Determining the tumor's histological grade is one of the most important steps in treatment planning, as it predicts how aggressively the cancer is likely to grow and spread. Studies have shown that histological grade is as informative as lymph node status in predicting outcomes.
Pathologists use the Nottingham Grading System (NGS) to assign a grade by scoring three features in tissue biopsy slides: the degree of tubule formation (how closely tumor cells resemble normal glandular structures), nuclear pleomorphism (the degree of variation in cell nucleus appearance), and mitotic count (the number of actively dividing cells). These three scores are summed to produce grades 1, 2, or 3, ranging from least to most aggressive.
Despite being the gold standard, manual grading is time-consuming, subjective, and prone to significant interobserver variability. Studies have found only about 75% agreement between individual pathologists and expert consensus diagnoses. This inconsistency can directly affect treatment decisions and patient outcomes.
Automated deep learning approaches, particularly Convolutional Neural Networks (CNNs), offer a way to make grading faster, more consistent, and less dependent on the experience of an individual pathologist. However, most existing studies tested only one or two CNN architectures, leaving clinicians and researchers without a systematic comparison to guide model selection for this specific task.
The study used the Four-Breast-Cancer-Grades (FBCG) dataset, which combines two existing datasets. The first, BreaKHis, contains 7,909 histopathological images from 82 patients at four magnification levels, covering both benign and malignant tumor subtypes. The second, the Breast Cancer Grading (BCG) dataset, contains 300 high-resolution images representing the three IDC grades from 21 patients.
A key methodological decision was to treat benign breast tissue images from BreaKHis as a fourth class, Grade 0. This creates a four-class grading problem: Grade 0 (benign/normal), Grade 1 (low-grade IDC), Grade 2 (intermediate-grade IDC), and Grade 3 (high-grade IDC). This approach more realistically mirrors the clinical scenario where the pathologist must first determine whether cancer is present and then assign a grade.
The combined dataset contained 888 images total, split 80/20 into training and test sets stratified by grade. The training set was further divided into five folds for cross-validation. The dataset is notably imbalanced, with Grade 0 being the largest class and Grade 3 the smallest, which was addressed using class weighting in the loss function.
All images were preprocessed to a uniform resolution compatible with each CNN architecture's input requirements (224 x 224 pixels) and normalized by scaling pixel values to the range 0 to 1. Data augmentation including random flips, rotation up to 20 degrees, and random zoom was applied during training to reduce overfitting on this relatively small dataset.
All seven models used transfer learning, specifically feature extraction, where the convolutional layers of each pretrained network are kept frozen and only a new classification head is trained on the breast cancer data. This approach is well-suited to small datasets because the pretrained network has already learned generalizable visual features from millions of natural photographs, and these features transfer effectively to histopathological images.
The seven CNN architectures evaluated were: EfficientNetB0, EfficientNetV2B0, EfficientNetV2B0-21k, ResNetV1-50, ResNetV2-50, MobileNetV1, and MobileNetV2. These span a range of architectural philosophies and computational requirements, from the lightweight MobileNet family designed for mobile devices to the more computationally intensive ResNet family with 25.6 million parameters.
A standardized classification head was added on top of each pretrained feature extractor: a dropout layer (rate 0.5), a dense layer with 256 neurons and ReLU activation, a second dropout layer (rate 0.4), and an output layer with 4 neurons and softmax activation. This identical head architecture across all seven models ensures that differences in performance reflect the underlying pretrained networks rather than the classifier design.
All models were trained for 100 epochs using the Adam optimizer with a learning rate of 0.001 and a weighted categorical cross-entropy loss function to account for class imbalance. Evaluation used balanced accuracy and macro-averaged precision, recall, and F1 score to give equal weight to all four classes regardless of their size in the dataset.
The top-performing model was EfficientNetV2B0-21k, achieving a balanced accuracy of 0.9666 on five-fold cross-validation and 0.9524 on the held-out test set. This model also achieved the highest macro precision, recall, and F1 score across all seven architectures. These scores reflect strong, consistent performance across all four grade classes, not just the most common one.
What makes this result particularly notable is that EfficientNetV2B0-21k achieves this performance with relatively modest computational requirements: only 0.72 billion floating-point operations and 7.1 million parameters, with an inference time of 0.076 seconds per image and a training time of just over 30 minutes. In contrast, the heavier ResNet models required 4.1 billion FLOPs and 25.6 million parameters while achieving lower accuracy.
The authors attribute EfficientNetV2B0-21k's superiority to its pretraining on the ImageNet-21k dataset, which contains approximately 12.4 million images across 21,000 classes, compared to the 1.2 million images in the more commonly used ImageNet-1k dataset. The broader and more diverse pretraining provides richer initial feature representations that transfer more effectively to new tasks.
The second-best model on the test set was MobileNetV1, which tied EfficientNetV2B0-21k in balanced accuracy (0.9524) while achieving the fastest inference time of 0.042 seconds, nearly half of EfficientNetV2B0-21k's speed. MobileNetV1 is thus a strong alternative for real-time or resource-constrained deployment scenarios.
One of the study's most practically important findings is that larger and more computationally expensive models do not consistently outperform simpler ones. The ResNet models, with 25.6 million parameters and 4.1 billion FLOPs each, achieved only moderate accuracy (0.925 to 0.935) while requiring the longest inference times (0.22 to 0.23 seconds per image).
In contrast, EfficientNetV2B0-21k with 7.1 million parameters achieved the highest accuracy, and MobileNetV1 with just 4.2 million parameters tied for best test set accuracy while predicting in under 0.05 seconds. This suggests that for the IDC grading task on the FBCG dataset, pretraining dataset quality (ImageNet-21k vs. ImageNet-1k) matters more than model architectural complexity.
All seven models achieved average balanced accuracy of 0.936 on cross-validation and 0.931 on the test set, with a standard deviation of less than 0.02. This tight clustering demonstrates that modern CNN architectures are uniformly capable at this grading task when trained with appropriate transfer learning, and that the choice between them can be guided primarily by deployment constraints rather than accuracy requirements.
The study's results have direct implications for how automated IDC grading systems should be designed. For hospital workstations where computational resources are plentiful and accuracy is the top priority, EfficientNetV2B0-21k is the recommended choice. For mobile or embedded applications such as smartphone-based diagnostic tools or resource-limited clinical settings, MobileNetV1 offers comparable accuracy at dramatically lower computational cost.
The fact that all models scored above 0.9 on both cross-validation and test metrics is clinically meaningful, as it suggests that transfer learning is a robust and reliable approach for histopathological grading even with relatively small datasets. The approach does not require thousands of annotated images, making it feasible even for uncommon cancer subtypes or rare grading conditions.
The authors emphasize that the choice of CNN architecture should be driven by the specific deployment context. Resource-intensive models are acceptable and even preferable in standard clinical pathology workstation settings. However, as pathology AI expands to point-of-care testing, mobile applications, or low-resource healthcare environments, lightweight models like MobileNetV1 that can predict a grade in 0.042 seconds on minimal hardware become highly relevant.
This study provides the most comprehensive benchmark of contemporary CNN architectures for IDC grading to date, comparing seven models using a standardized methodology on a four-grade dataset. The conclusion that EfficientNetV2B0-21k achieves the best accuracy-efficiency trade-off fills an important knowledge gap for researchers and clinicians seeking to implement automated grading systems.
Important limitations include that results are specific to the FBCG dataset and may not generalize to other breast cancer datasets or clinical cohorts. The study used feature extraction only and did not explore fine-tuning, which could potentially improve performance but also risks overfitting on the small training set. Stain normalization was also not applied, though prior work suggests this may not be critical when using robust features.
Future directions include testing on larger and more diverse histopathological datasets, evaluating additional recent CNN architectures such as Vision Transformers, and investigating fine-tuning strategies that can safely improve performance without causing overfitting. Prospective clinical validation comparing automated grading against pathologist consensus would be the key step toward clinical deployment.