Breast cancer affects approximately one in eight women in the United States over their lifetime. It encompasses a spectrum of disease types, from confined, slow-growing lesions to invasive cancers that spread throughout the body. Two of the most clinically significant presentations are Invasive Ductal Carcinoma (IDC) and metastatic breast cancer, both of which require early and accurate identification for effective treatment.
Invasive Ductal Carcinoma (IDC) begins in the milk ducts of the breast and invades surrounding breast tissue. Unlike ductal carcinoma in situ (DCIS), which remains confined to the duct, IDC has broken through the duct wall. IDC is the most common type of breast cancer and, if not detected early, can progress to involve nearby blood or lymph vessels and spread to distant organs.
When breast cancer cells reach lymph vessels or the bloodstream, they can travel to other parts of the body and establish new tumor sites. This process, called metastasis, creates secondary tumors at distant sites - most commonly the bones, liver, lungs, and brain. Initially these may form tiny clusters called micrometastases that develop their own blood supply. Metastatic (stage IV) breast cancer is much harder to cure and significantly shortens patient survival compared to localized disease.
This paper, published in Cancer Medicine (2024) by researchers from Bangladesh's Military Institute of Science and Technology and collaborators in Malaysia and the USA, proposes a customized two-branch CNN model that simultaneously addresses IDC detection and metastasis classification using microscopic histopathological images, aiming to improve accuracy for both tasks in a computationally accessible framework.
The study used two different datasets for the two branches of the model. For IDC detection, the researchers used a widely cited collection of breast histopathology images from the Hospital of the University of Pennsylvania and the Cancer Institute of New Jersey, covering 162 patients. The original whole-slide images were processed into 277,524 patches of 50x50 pixels: 198,738 IDC-negative and 78,786 IDC-positive. This patch-based approach allows the model to learn local cellular features at the cost of losing whole-slide context.
For metastasis detection, the researchers used the PatchCamelyon (PCam) benchmark dataset - a widely used, clinically curated dataset for classifying metastatic tissue in lymph node histopathology images. PCam contains 327,680 color images of 96x96 pixels taken from lymph node segment histopathologic scans. Each image is labeled as positive if there is at least one pixel of tumor tissue in the central 32x32 pixel region of the patch.
Data preprocessing was minimal by design - a key advantage over competing approaches. Images were resized and normalized using standard mean-standard deviation normalization (subtracting the pixel mean and dividing by the standard deviation channel-by-channel). No complex stain normalization, hand-crafted feature extraction, or multi-step augmentation pipeline was required. This computational simplicity was a deliberate design goal.
Both datasets were split 80% for training and 20% for testing, with 15% of training data reserved for validation. The IDC branch trained on 22,201 images for evaluation; the metastasis branch used 16,000 test images (8,000 per class, balanced). The Adam optimizer with learning rate decay was used for both branches, with batch size 128 and 40 epochs for IDC and 20 epochs for metastasis.
The proposed model uses a two-branch architecture, with each branch customized for its specific classification task. Both branches share the same general structure - three stages of convolutional and pooling layers - but differ in input size and output layer configuration to match their respective datasets and classification goals.
Branch 1 (IDC detection) takes 50x50 pixel patches as input and processes them through three stages, each containing three consecutive Conv2D layers followed by MaxPooling2D and Dropout. The filter counts increase from 32 (stage 1) to 64 (stage 2) to 128 (stage 3), allowing the network to learn progressively more complex and abstract features. After the convolutional stages, a Flatten layer converts the spatial feature maps to a 1D vector, followed by two Dense layers (256 units and 2 outputs). Total trainable parameters: 514,306.
Branch 2 (metastasis detection) takes the larger 96x96 pixel PCam patches and uses the same three-stage structure with identical filter counts (32, 64, 128). The larger input size requires a larger Dense layer at the classification stage (1,179,904 parameters in the first Dense layer alone), resulting in a total of 1,661,186 trainable parameters. The sigmoid activation function is used throughout both branches for binary output classification.
The Adam optimizer was used for training both branches. Adam combines adaptive learning rates with momentum by tracking exponential moving averages of both gradients and squared gradients, making it particularly effective for training CNNs on medical image datasets. Dropout with rate 0.02 was applied after each pooling stage to prevent overfitting, and the model was implemented and trained on Google Colab using GPU resources.
Branch 1 (IDC detection) achieved an overall accuracy of 89% on 22,201 test images. The confusion matrix showed 5,736 true positive (IDC present, correctly identified) and 13,655 true negative (IDC absent, correctly identified) predictions, with 576 false positives and 2,234 false negatives. The per-class F1 scores were 0.96 for IDC-absent and 0.91 for IDC-present, with an overall precision of 89%, recall of 84%, and F1 of 94%.
Branch 2 (metastasis detection) significantly outperformed Branch 1, achieving 95% overall accuracy on 16,000 balanced test images (8,000 per class). The confusion matrix showed 7,509 true positives and 7,629 true negatives with 491 false positives and 371 false negatives. Precision, recall, and F1 score were each 0.95 for both classes. The ROC curve showed an AUC of 0.985, indicating excellent discriminative ability between metastasis-present and metastasis-absent cases.
The higher accuracy of the metastasis branch (95%) compared to IDC (89%) likely reflects the inherent difficulty of the IDC task. IDC detection in 50x50 pixel patches must distinguish subtle local cellular features against a background of similar-looking tissue. Metastasis detection in the PatchCamelyon dataset benefits from larger 96x96 patches and a more clearly defined label criterion (at least one pixel of tumor in the central region), making positive cases more visually distinct.
Importantly, both models showed training and validation curves that converged without signs of overfitting - with training and validation accuracy tracking closely and loss steadily decreasing. This is notable given the relatively lightweight preprocessing pipeline: the model learned effective representations from normalized patches without requiring extensive augmentation or stain normalization.
The comparison with state-of-the-art methods is instructive. For metastasis detection, the proposed 95% accuracy exceeds handcrafted feature methods (Kahya et al.: 94.54%), CSDCNN-based approaches (93.2%), VGG16-based transfer learning (92.60%), ResNet-152 transfer learning (91.35%), and DCNN-SVM-AlexNet combinations (87.2%). Against DenseNet-based metastasis classification, the improvement is substantial: 95% versus 83.1%.
For IDC detection, the proposed 89% accuracy is lower than some published approaches - notably the IRRCNN architecture achieved 96.84%. However, most higher-performing IDC models require substantially more complex preprocessing pipelines, longer training times, or high-end hardware resources that are not always available. The proposed model achieves competitive accuracy with minimal preprocessing, which has practical importance for clinical deployment.
The key claimed advantage of this approach is computational accessibility. Unlike transfer learning approaches that initialize from large pre-trained networks like ResNet-152 or VGG16, this model is trained from scratch with a relatively compact architecture. The two-branch design also enables simultaneous IDC and metastasis classification within a single framework, potentially reducing the need for separate specialized models in clinical workflows.
The authors note that a limitation is the absence of explainability tools. Unlike some competing approaches that include attention maps or Grad-CAM visualizations, the current model does not provide visual explanations of its predictions. Integrating Explainable AI (XAI) techniques like Grad-CAM and SHAP is explicitly identified as a priority for future work, recognizing that oncologists need to understand model decisions before trusting them in clinical settings.
This study demonstrates that a customized two-branch CNN model, trained with minimal preprocessing on publicly available histopathology patch datasets, can achieve 89% accuracy for IDC classification and 95% accuracy for metastasis detection - performance levels competitive with many state-of-the-art approaches while being computationally more accessible.
The most significant clinical implication is the potential for real-time diagnostic assistance. Computer-aided diagnosis systems that can automatically classify IDC and metastasis from histopathology images could help oncologists identify cancer type and stage more rapidly, directly guiding decisions about treatment intensity and approach. The low preprocessing requirements make this particularly relevant for resource-constrained clinical settings.
However, the authors acknowledge that moving from benchmark dataset performance to genuine clinical deployment requires addressing current limitations. The absence of model explainability is the most pressing gap: clinicians cannot trust predictions they cannot understand. Future integration of Grad-CAM and SHAP analysis will be necessary to show pathologists which image features are driving each prediction.
Future research directions include developing more robust and generalizable models using private institutional datasets alongside the public benchmarks, expanding the model to classify additional breast cancer subtypes beyond IDC and metastasis, and conducting prospective clinical validation studies to measure the model's impact on diagnostic accuracy and patient care pathways.