Breast cancer is the most commonly diagnosed cancer in women worldwide. Identifying axillary lymph node (ALN) status is critical for staging, treatment planning, and prognosis. Whether cancer cells have spread to lymph nodes determines whether patients need additional procedures like further surgery or radiotherapy.
The traditional method, axillary lymph node dissection (ALND), is invasive and can cause serious complications including infection, nerve damage, shoulder dysfunction, arm numbness, and upper limb lymphedema. The more refined alternative, sentinel lymph node biopsy (SLNB), is less invasive but still carries a 3.5 to 10.9% risk of arm numbness or lymphedema.
Current imaging methods such as ultrasound, CT, and MRI have difficulty accurately predicting sentinel lymph node metastasis and often produce false-negative results. There is a clear clinical need for a fully non-invasive, precise preoperative prediction tool that spares patients the risks and delays of intraoperative pathology.
Radiomics is a field that converts medical images into large sets of quantitative, computationally extracted features. These features can capture subtle patterns of tumor appearance -- including texture, shape, and signal intensity -- that are not visible to the naked eye but may correlate with biological behavior.
Radiomic features can reflect intratumor heterogeneity, a hallmark of cancer aggressiveness. By analyzing these features alongside machine learning, researchers can build models that predict biological outcomes -- such as lymph node spread -- based on the tumor's imaging signature rather than requiring tissue samples from lymph nodes themselves.
Among imaging modalities, contrast-enhanced MRI (CE-MRI) offers superior temporal and spatial resolution compared to ultrasound or mammography. Dynamic contrast-enhanced MRI captures how tumor tissue takes up contrast agent over time, revealing information about vascularity and permeability that is closely linked to tumor biology.
This retrospective study enrolled 177 breast cancer patients from a single institution between January 2015 and May 2021. All patients had pathologically confirmed breast cancer and underwent dynamic contrast-enhanced MRI (DCE-MRI) before surgery. Of these, 81 had positive sentinel lymph nodes and 96 had negative nodes.
MRI scans were performed on a 3.0T scanner using a dedicated breast coil. The imaging protocol included multiple sequences, and the second phase of DCE-MRI (61 to 122 seconds after contrast injection) was selected for radiomic analysis because it provided the greatest contrast between the tumor and surrounding tissue.
Patients were divided 70:30 into training (123 patients) and validation (54 patients) sets. Inclusion required complete imaging and pathological data, while patients who had received preoperative chemotherapy, endocrine therapy, or radiotherapy were excluded to avoid confounding the imaging features.
Tumor regions of interest (ROIs) were manually segmented on DCE-MRI images by an experienced radiologist, with review by a senior radiologist. From each 3D ROI, a total of 1,316 radiomic features were extracted, covering first-order statistics (intensity-based), 3D shape features, and multiple texture categories including gray-level co-occurrence and run-length matrices.
All features were normalized using maximum-minimum normalization to bring values into a consistent range, which helps machine learning algorithms converge more reliably. Feature selection then proceeded in two stages: first, a variance-based filter reduced the set to 132 features, and then LASSO logistic regression further narrowed this to 13 optimal predictive features.
The 13 final features included two first-order statistical features, two shape-based features, and nine texture features. These features were deemed most predictive of sentinel lymph node metastasis (SLNM) and served as the inputs for all machine learning models.
Five machine learning classifiers were built and compared using the 13 selected features: Support Vector Machine (SVM), Random Forest (RF), Logistic Regression (LR), Gradient Boosting Decision Tree (GBDT), and Decision Tree (DT). Each model was trained on the training set and tested on the independent validation set.
A combined model was then constructed by augmenting the best-performing radiomic model with two clinically significant variables: tumor size and BI-RADS classification (a standardized imaging assessment score). This combined model used the SVM algorithm and was evaluated using the same framework as the individual radiomic models.
Model performance was evaluated using the area under the ROC curve (AUC), accuracy, sensitivity, and specificity. DeLong's test was used to statistically compare AUC values between models. Ten-fold cross-validation was applied to validate model accuracy on the training set.
In the training set, RF, GBDT, and DT all achieved perfect AUC scores of 1.00, indicating likely overfitting. In the more meaningful validation set, SVM achieved the highest AUC of 0.86 (accuracy 78%, sensitivity 60%, specificity 90%), followed closely by RF (AUC 0.85), LR (AUC 0.84), and GBDT (AUC 0.82). The Decision Tree model performed significantly worse at AUC 0.74.
The combined model -- incorporating SVM radiomic features, tumor size, and BI-RADS classification -- outperformed all individual models with a validation set AUC of 0.88 (accuracy 80%, sensitivity 68%, specificity 90%). This improvement demonstrates that adding clinically available variables enhances predictive power beyond imaging features alone.
Only two clinical variables were significantly different between SLN-positive and SLN-negative patients: tumor size (larger tumors more likely to have lymph node metastasis) and BI-RADS classification. Age, histological type, grade, and molecular subtype did not differ significantly between groups, underscoring the value of radiomic features for capturing information these standard clinical factors miss.
The Support Vector Machine (SVM) algorithm demonstrated the best generalization performance in the validation set, despite not achieving a perfect training score. SVM works by finding an optimal boundary (hyperplane) between classes in high-dimensional feature space, seeking the best balance between model complexity and learning ability.
Decision Tree models, which did achieve perfect training scores, performed worst on the validation set -- a classic sign of overfitting, where a model memorizes training data rather than learning generalizable patterns. The gap between training and validation performance was far smaller for SVM, confirming its stronger real-world applicability.
These findings are consistent with prior studies where SVM consistently outperforms other classifiers in radiomic prediction of lymph node metastasis. SVM handles high-dimensional, small-sample problems particularly well -- characteristics that describe this type of radiomic dataset precisely, with 13 features and only 177 patients.
Prior studies using ultrasound, mammography, and other MRI sequences for radiomic prediction of lymph node metastasis have generally produced lower AUC values than the CE-MRI approach in this study. For instance, a model based on T2-weighted fat-suppression and diffusion-weighted MRI (DWI) achieved an AUC of only 0.805 in its validation cohort.
CE-MRI is particularly informative because contrast enhancement reveals tumor vascularity and perfusion dynamics -- both of which are linked to tumor aggressiveness and metastatic potential. The second phase of dynamic contrast imaging, used here, captures peak enhancement and provides the clearest distinction between tumor and background tissue.
The combined model in this study, while slightly below the 0.90 AUC achieved by more complex multi-sequence radiomics models in the literature, is considerably simpler to implement. Using a single standardized MRI sequence reduces variability in feature extraction and makes the approach more practical for routine clinical adoption.
Radiomic analysis has known technical limitations. Image acquisition parameters such as scanner settings, field strength, and reconstruction algorithms affect feature values, potentially reducing reproducibility across institutions or scanners. This study used a single MRI scanner with a consistent protocol, which minimizes but does not eliminate this concern.
Manual tumor segmentation -- the process of drawing boundaries around the tumor on each image slice -- is time-consuming, subject to user variability, and prone to error. The study used strict inter-reader agreement protocols (Cohen's kappa), with a senior radiologist resolving discrepancies, but automated segmentation systems will be needed for practical scaling.
Feature selection choices also significantly affect model outcomes. Different combinations of filters and selection methods can produce different feature sets and thus different model behaviors. Standardizing these pipelines is an active area of research, and the Imaging Biomarker Standardization Initiative (IBSI) has proposed guidelines to improve reproducibility across radiomic studies.
This study demonstrated that machine learning models built from CE-MRI radiomic features can predict sentinel lymph node metastasis in breast cancer patients with high accuracy non-invasively before surgery. The SVM-based combined model, incorporating 13 radiomic features, tumor size, and BI-RADS classification, achieved an AUC of 0.88 in an independent validation set.
The study has several acknowledged limitations: it is a single-center retrospective study with relatively small sample size (177 patients), which limits generalizability and robustness. The training and validation cohorts came from the same institution, and the results have not yet been tested against data from other hospitals.
Future work will focus on developing automated segmentation tools, enrolling larger patient cohorts, and conducting multicenter validation studies. If confirmed in broader populations, this approach could offer a practical, non-invasive preoperative method to guide surgical decision-making and spare patients from unnecessary sentinel lymph node biopsies.