Breast cancer has become the most common cancer threat to women globally. A critical factor in staging and treatment planning is whether cancer has spread to the axillary lymph nodes (ALN) -- the lymph nodes located in the armpit that serve as early stations in breast cancer's path to distant organs.
Sentinel lymph node biopsy (SLNB) has become the standard method for evaluating axillary node status, replacing the more invasive full axillary lymph node dissection in patients with clinically negative nodes. However, SLNB still carries risks including lymphedema, seroma, paraesthesia, and impaired shoulder movement, and depends heavily on the availability of radionuclide tracers and surgeon expertise.
If the ALN status could be predicted reliably from information already gathered during standard preoperative workup -- without any additional invasive procedure -- many patients could safely forgo SLNB entirely. This would reduce complications, hospitalization time, and costs, while preserving quality of life for patients whose nodes turn out to be negative.
This study took a fundamentally different approach from most prior work on lymph node prediction, which has relied on imaging modalities like ultrasound, MRI, or CT. The researchers investigated whether digitized histopathology slides from core-needle biopsies (CNB) of the primary breast tumor -- images that are already routinely created during preoperative diagnosis -- could predict axillary lymph node metastasis.
Core-needle biopsy is the standard method for diagnosing breast cancer before surgery, providing both tissue for histological diagnosis and molecular profiling (ER, PR, HER2 status). The digitized H&E-stained slides from these biopsies contain rich morphological information about tumor cell architecture, nuclear features, and tissue organization that has traditionally been interpreted only by pathologists for diagnosis, not for lymph node prediction.
To the best of the authors' knowledge, this is the first study to apply deep learning to primary breast tumor CNB whole-slide images specifically for predicting axillary lymph node status -- a clinically impactful use of data that is already collected in standard care. A total of 1,058 patients from Beijing Chaoyang Hospital were enrolled across a 10-year period.
Whole-slide images (WSIs) of core-needle biopsy samples are gigapixel-scale images -- far too large to process directly with standard convolutional neural networks. The study used attention-based multiple instance learning (AMIL), a framework designed specifically for this challenge.
In AMIL, each WSI is treated as a 'bag' of smaller image patches (each 256x256 pixels). The model does not have labels for individual patches -- only the overall ALN status of the patient is known. By learning to assign attention weights to each patch, the model identifies which local tissue regions contribute most to the prediction, effectively learning to focus on diagnostically important areas without explicit patch-level annotation.
The attention mechanism outputs a weight for each patch in the bag. These weights are multiplied back into the patch feature vectors, emphasizing the contributions of the most predictive regions. This process not only improves classification accuracy but also enables post-hoc visualization through heat maps that highlight which regions of the biopsy slide drove the prediction -- providing a degree of model interpretability that is valuable for pathologist review.
Five CNN architectures were tested as feature extractors: AlexNet, VGG16 with batch normalization (VGG16_BN), ResNet50, DenseNet121, and Inception-v3. All were pre-trained on ImageNet and fine-tuned on the biopsy patch data. VGG16_BN outperformed all others with the highest AUC on both the validation and independent test cohorts.
In the enhanced model (DL-CNB+C), clinical variables were incorporated alongside image features. These included patient age, tumor size, tumor type, and molecular status (ER, PR, HER2). Numerical variables were standardized and categorical ones were one-hot encoded. To account for the large dimensional difference between clinical vectors and image features, clinical feature vectors were replicated 10 times before concatenation with image features.
Training used the Adam optimizer with cosine annealing warm restarts for learning rate scheduling, cross-entropy loss, and a 4:1 train/test split with 25% of the training set held as a validation cohort. The dataset of 1,058 patients yielded 840 in the training cohort and 218 in the independent test cohort. Both cohorts were well matched on all baseline characteristics (all p > 0.05).
In the independent test cohort, the imaging-only model (DL-CNB) achieved an AUC of 0.816, sensitivity of 81.0%, and specificity of 70.9% for predicting positive ALN status. This substantially outperformed the clinical-data-only model (AUC 0.613, p < 0.0001), demonstrating that biopsy slide morphology contains unique predictive information beyond standard clinical variables.
Adding clinical data to produce the combined model (DL-CNB+C) raised the AUC slightly to 0.831 with improved accuracy (75.7%), sensitivity (89.3%), and negative predictive value (90.9%). The combined model had the best overall accuracy and NPV, meaning it was particularly good at confirming the absence of lymph node metastasis -- the most clinically valuable prediction for avoiding unnecessary SLNB.
A striking subgroup result emerged: for patients aged 50 or younger, the DL-CNB+C model achieved an AUC of 0.918 (95% CI: 0.825, 0.971) -- far exceeding its performance in patients older than 50 (AUC 0.794). This suggests the model captures tumor biology most relevant to metastatic potential in younger patients, where aggressive tumor phenotypes are more common and standard clinical variables alone perform poorly.
The researchers went beyond performance metrics to investigate which biological features drove the model's predictions. They extracted eight specific nuclear morphometric features from each biopsy slide -- including nucleus density, major and minor axis length, area, circumference, circularity, orientation, and rectangularity -- and re-ran the AMIL framework using these quantitative features as instances instead of raw image patches.
The attention weights from this interpretability analysis revealed that the top predictive features were nuclear density (p = 0.015), circumference (p = 0.009), circularity (p = 0.010), and orientation (p = 0.012). Specifically, tumors with lymph node metastasis had higher nuclear density and more directional orientation (polarity) but lower circularity and smaller circumference compared to node-negative tumors.
This biological pattern makes intuitive sense: poorly differentiated cancer cells undergoing rapid growth tend to form tightly clustered, elongated, and less rounded nuclei with consistent orientation patterns -- features associated with invasive tumor phenotypes. This analysis provides the first quantitative confirmation that nuclear disorder patterns, previously linked to patient survival outcomes, are also informative for lymph node metastasis prediction.
Beyond the binary classification of positive versus negative, the study attempted a more clinically granular three-class prediction: distinguishing node-negative (N0), low metastatic burden (N+(1-2) positive nodes), and high metastatic burden (N+(3+) positive nodes) groups. This distinction matters because patients with 1-2 positive nodes may safely omit full axillary dissection, while those with 3+ require it.
The combined model achieved an overall multi-class AUC of 0.791. It performed best at identifying N0 patients (precision 0.747, recall 0.947), but struggled at the boundary between N+(1-2) and N+(3+) subgroups. This pattern of high sensitivity for node-negative detection but limited resolution of metastatic burden level was also reported by other research groups using different imaging modalities.
The binary task of distinguishing N0 from N+(1-2) (AUC 0.878) was more accurate than distinguishing N0 from N+(3+) (AUC 0.838), suggesting that the imaging features most predictive of any metastasis are somewhat distinct from features that predict heavy nodal burden. Further research into this specific classification challenge will be needed before the model can guide full surgical treatment decisions.
This study extends the rapidly growing field of computational pathology -- the use of AI to extract clinically relevant information from histology slides -- into a new application domain. Prior AI work on pathology slides has focused on diagnosis, tumor grading, molecular subtype classification, and prediction of survival outcomes, but not on preoperative lymph node status prediction from biopsy material.
An important practical advantage of this approach is that it uses data already available in standard clinical care. Core-needle biopsies are performed for all breast cancer diagnoses, and WSIs are increasingly generated as part of digital pathology workflows. Adding a lymph node prediction analysis to this existing pipeline requires no additional imaging, contrast agents, or specialized equipment.
The study also demonstrates that biopsy-based WSI analysis can outperform imaging modalities for some prediction tasks. Unlike ultrasound or MRI, which assess the tumor from outside the tissue, histopathology directly visualizes the cellular architecture and nuclear morphology that drive metastatic behavior. This biological directness may explain why even a relatively simple feature set -- nuclear shape parameters -- captures significant metastatic signal.
This study demonstrated that deep learning applied to routine core-needle biopsy H&E slides can predict axillary lymph node metastasis in early breast cancer with meaningful accuracy (AUC 0.831), particularly for younger patients. The combination of image-based features with clinical data produced the best results, and nuclear morphometry analysis provided biologically interpretable insights into what the model learned.
Key limitations include the retrospective single-center design, the requirement for pathologist-guided region annotation (which limits scalability), and the restriction to H&E-stained slides -- immunohistochemical stains that visualize tumor microenvironment markers such as tumor-infiltrating lymphocytes (TILs) could provide additional predictive signal. The small number of TILs observed on CNB slides also prevented their inclusion in the current model.
Future priorities include prospective multicenter validation, development of automated tumor region segmentation to remove pathologist annotation dependency, and exploration of additional staining modalities. If validated prospectively, this approach could become a valuable complement to standard preoperative workup, helping to identify patients who can safely forgo sentinel lymph node biopsy and its associated procedural risks.