Breast cancer is responsible for approximately 40,000 deaths annually in the United States, with most deaths resulting from metastatic disease. The first site of spread is typically the ipsilateral axillary lymph nodes - the lymph nodes under the arm on the same side as the tumor. Whether these nodes contain cancer cells is one of the most important factors in determining prognosis and treatment.
Knowing the nodal status guides critical treatment decisions, including whether to use neoadjuvant chemotherapy (treatment before surgery), radiation therapy, and the extent of surgical lymph node removal. The pathologic node stage (pN) - confirmed by surgery - is the gold standard, while the clinical node stage (cN) is assessed preoperatively through imaging and biopsy.
Current methods for preoperative assessment are imperfect. Ultrasound-guided biopsy correctly identifies only about 50% of patients with axillary metastasis. Breast MRI, while useful, achieves a mean sensitivity of only 60% for detecting nodal involvement. When imaging is inconclusive, patients undergo sentinel lymph node biopsy (SLNB) - a surgical procedure that carries meaningful risks of lymphedema, chronic pain, sensory disorders, and limited arm motion, even though 52-82% of patients who undergo it ultimately have benign (cancer-free) nodes.
This study developed and validated a four-dimensional (4D) convolutional neural network (CNN) that predicts axillary lymph node metastasis non-invasively from breast MRI data. The model integrates both imaging features and clinical/pathologic variables to assess whether a patient's lymph nodes are cancer-positive or cancer-negative before any surgery.
Unlike prior studies limited to a single institution and MRI scanner, this model was trained and tested on data from two institutions with different MRI equipment and a racially diverse patient population. This multi-institutional design is critical for creating a model that could actually generalize to real-world clinical settings.
The approach uses dynamic contrast-enhanced (DCE) MRI - a type of scan that captures how contrast dye flows in and out of tumor tissue over time. By analyzing this temporal sequence as a 4D dataset (three spatial dimensions plus time), the model captures information about tumor vascularity and enhancement patterns that a single snapshot misses.
The retrospective study included 350 female patients (mean age 51.7 years) with newly diagnosed invasive breast cancer who underwent DCE breast MRI at a university hospital and a safety-net hospital within the same health system between 2013 and 2016. Clinical data collected included age, race, tumor size, estrogen receptor (ER) and HER2 status, Ki-67 index, and tumor grade.
Each patient underwent MRI with one precontrast scan and four successive contrast-enhanced scans after contrast injection, yielding a temporal series that captures how the tumor takes up and washes out the contrast agent. A fellowship-trained breast radiologist with 15 years of experience manually delineated the primary tumor boundary, though the final model only requires a rough bounding box that can be drawn in under a minute.
Images were preprocessed to a consistent 1mm isotropic resolution, harmonized with histogram equalization across scanner types, and cropped to 120x120x120 voxel cuboidal volumes centered on the tumor. Crucially, the model analyzed both the primary tumor region and the ipsilateral axillary region to capture both tumor characteristics and any detectable changes in the axilla itself.
Four model variants were developed to systematically isolate the contribution of each data type: (1) a clinical-only model using age, ER/HER2 status, Ki-67, and tumor grade; (2) a 3D model using only the contrast difference image at a single time point; (3) a 4D image model using all DCE time points; and (4) the 4D hybrid model combining all temporal imaging data with clinical variables.
The 4D hybrid model incorporated a vision transformer-inspired patch-based approach - dividing the 3D volume into small overlapping patches and learning features from each, similar to how transformers process words in a sentence. Hyperparameter optimization used Heteroscedastic Evolutionary Bayesian Optimization (HEBO), which adapts more flexibly to noisy medical imaging data than traditional optimization methods.
Model selection and evaluation used nested fivefold cross-validation, where the data was divided into five groups with each serving once as the held-out test set. This rigorous approach prevents overly optimistic performance estimates that can arise from simpler train/test splits and provides a reliable estimate of real-world clinical performance.
For predicting pathologic node status (pN), the 4D hybrid model achieved an AUC of 0.87 (95% CI: 0.83-0.91), sensitivity of 89%, and specificity of 76% - substantially outperforming the clinical-only model (AUC 0.63), the 3D imaging model (AUC 0.69), and the 4D imaging-only model (AUC 0.71).
For predicting clinical node status (cN), the 4D hybrid model achieved an AUC of 0.79, sensitivity of 80%, and specificity of 62%. The model's performance exceeded standard clinical imaging, which achieved a true-positive rate of only 77.6% for detecting axillary metastasis in this dataset.
At a slightly reduced specificity threshold of 71%, the model achieved 91% sensitivity and only 9% false-negative rate - matching the benchmark standards applied to SLNB itself. Applied clinically, this model would have avoided more than half (51%) of benign sentinel node biopsies while correctly detecting 95% of patients with axillary metastasis when operating at a sensitivity of 95%.
The study confirmed several known associations between tumor characteristics and lymph node involvement. Larger tumors (higher T category) strongly predicted both cN+ and pN+ status. The difference in imaging sensitivity between T1 and T2 tumors was particularly striking: 27.8% vs 81.7%, highlighting how standard imaging misses most nodal metastases in small cancers.
HER2-positive tumors had 7.6-fold higher odds of pN+ disease compared to HER2-negative tumors (odds ratio 7.6, p less than 0.001). Higher Ki-67 index (a measure of tumor cell proliferation) and higher tumor grade also strongly predicted nodal positivity. Patients with node-positive disease were significantly younger.
Saliency mapping - a technique that highlights which image regions most influenced the model's predictions - showed the model focused on primary tumor and peritumoral voxels rather than distal tissue. Notably, the model appeared to learn to approximate tumor boundaries on its own without being explicitly given that information.
When trained and tested on data from a single institution at a time, the model maintained strong performance: AUC 0.84 when trained on the safety-net hospital and tested on the university hospital, and AUC 0.81 in the reverse direction. These similar results (p = 0.68) across two institutions with different patient populations and MRI equipment suggest meaningful generalizability.
Compared to prior published models for axillary metastasis prediction from DCE-MRI, this model's performance was competitive. A comparable study by Yu et al. achieved an AUC of 0.9 but required detailed manual segmentation of axillary lymph nodes - a time-consuming process impractical for routine clinical use. The current model achieves similar diagnostic performance using only a rough bounding box around the primary tumor.
Limitations include the requirement for a manual tumor bounding box (which could be automated in future work), scanner field strength limited to 1.5T, and the absence of an independent external prospective validation cohort. The racial diversity of the patient cohort (36% non-Hispanic White, 35% Hispanic, 26% Black) is a notable strength for developing models applicable to heterogeneous real-world populations.
The practical clinical value of this model lies in its potential to safely reduce the number of patients who undergo unnecessary sentinel lymph node biopsy. In this study's cohort, the model could have avoided 51% of benign SLNBs while maintaining 95% sensitivity for detecting true metastases - meaning very few patients with actual cancer spread would be missed.
SLNB carries real risks: subclinical lymphedema occurs in 24% of patients, clinical lymphedema in about 5.6%, chronic axillary pain in 16%, sensory disorders in 2-22%, and limited arm motion in up to 9%. Avoiding SLNB in the large proportion of patients with benign nodes would prevent these complications for thousands of patients annually.
The model is particularly valuable as a clinical decision support tool that could help select patients who can safely forgo SLNB and proceed directly to surgery or systemic therapy. Patients with dense breasts, hormone receptor-negative tumors, or those eligible for neoadjuvant therapy - groups where nodal assessment has greatest impact on treatment planning - are likely to benefit most.