Axillary lymph node (ALN) status is one of the most important prognostic factors in breast cancer, directly influencing decisions about the extent of surgery, the type of systemic therapy, and the role of radiotherapy. The current gold standard for evaluating this status is sentinel lymph node biopsy (SLNB), which while less invasive than full axillary dissection, still carries meaningful risks including lymphedema, shoulder dysfunction, and paralysis.
Beyond complications, SLNB has practical limitations: its false-negative rate is approximately 5.5%, intraoperative pathology review prolongs anesthesia time, and the procedure adds costs and increases the risk of postoperative venous thrombosis. For patients at genuinely low risk of nodal metastasis, the harm-to-benefit ratio of SLNB may be unfavorable -- a consideration that has driven growing interest in non-invasive preoperative prediction tools.
Prior non-invasive prediction models have had important limitations: many incorporate postoperative pathological data (such as lymphovascular invasion or detailed immunohistochemical status) that are unavailable before surgery, or rely on complex machine learning systems impractical for community hospitals. This study aimed to develop a simpler, broadly accessible model using only imaging and clinical data available in the preoperative setting.
This retrospective study enrolled 4,276 breast cancer patients operated at the First Affiliated Hospital of Zhengzhou University between January 2020 and September 2022. After applying strict inclusion and exclusion criteria -- excluding ductal carcinoma in situ, cases without preoperative MRI, patients who received neoadjuvant chemotherapy, and those with incomplete data -- 999 patients (1,006 cases) remained for analysis.
The study combined two data sources: standard clinicopathological variables (age, BMI, menstrual status, tumor location, multifocality, MRI-BI-RADS classification) and quantitative MRI-derived measurements (tumor size on dynamic contrast-enhanced MRI, apparent diffusion coefficient (ADC) value from diffusion-weighted imaging, and morphological characteristics of the most suspicious axillary lymph node including long axis, short axis, and cortical thickness).
The analytical approach used logistic regression to identify independent predictors and construct a prediction nomogram, supplemented by a recursive partitioning analysis (RPA) to create a simple three-tier risk stratification system. This dual approach provides both individualized probability estimates (via the nomogram) and a practical low/intermediate/high risk classification (via RPA) for clinical decision-making.
All MRI examinations were performed on a 3.0 Tesla scanner using a dedicated 8-channel breast coil, with patients positioned prone. The imaging protocol included axial T1-weighted, axial T2-weighted, dynamic contrast-enhanced MRI (DCE-MRI), diffusion-weighted imaging (DWI), and sagittal contrast-enhanced sequences. Images from the second DCE phase (61 to 122 seconds post-injection) were selected for analysis as they provided optimal tumor-to-background contrast.
For each patient, radiologists measured five MRI parameters: tumor size (maximum diameter), ADC value (the apparent diffusion coefficient, a measure of how freely water molecules move through tissue -- restricted in dense, cellular tumors), and for the most morphologically suspicious level I axillary lymph node, the maximum long axis, maximum short axis, and maximum cortical thickness. When multiple suspicious nodes were present, measurements were taken from the largest one.
Measurement reproducibility was assessed in a subset of 32 patients by two independent radiologists (9 and 12 years of breast MRI experience). The intraclass correlation coefficient for tumor size measurements was 0.89 (95% CI: 0.83 to 0.94), indicating good-to-excellent inter-reader agreement. Both radiologists were blinded to pathological type and sentinel lymph node status during measurements to prevent interpretation bias.
Univariable logistic regression identified eight variables significantly associated with sentinel lymph node (SLN) positivity: tumor size on MRI, tumor location, multifocality, MRI-BIRADS classification, ADC value, long axis, short axis, and cortical thickness (all p less than 0.05). Age, BMI, menstrual status, and time-intensity curve type were not significant predictors.
Multivariable logistic regression confirmed six independent predictors: tumor size on MRI, multifocality, MRI-BIRADS classification, ADC value, short axis of the suspicious lymph node, and cortical thickness. Patients with tumors larger than 5 cm had a 51.5% SLN metastasis rate compared to 26.3% for tumors up to 2 cm. Multifocal tumors had a metastasis rate of 60%, approximately twice that of unifocal tumors (33.4%). Higher BIRADS classification showed progressively higher rates: 10.3% for BIRADS 3, 31.6% for BIRADS 4, and 48.4% for BIRADS 5.
Individual lymph node measurements on MRI were each independently predictive when analyzed alone. Cortical thickness showed the strongest individual discriminatory power with AUC 0.854, sensitivity 78.3%, and specificity 78.2% at a cutoff of 2.7 mm. Short axis (AUC 0.800, cutoff 7.0 mm) outperformed long axis (AUC 0.751, cutoff 9.7 mm), and ADC values were significantly lower in metastatic nodes (0.92 versus 0.98 x 10-3 mm2/s), reflecting the denser, more cellular structure of tumor-infiltrated lymph node tissue.
The prediction nomogram combining all six independent predictors achieved a concordance index (AUC) of 0.847 (95% CI: 0.822 to 0.872), indicating excellent discriminatory ability between patients with and without SLN metastasis. This substantially outperforms any individual predictor alone (best single AUC 0.854 for cortical thickness alone, but the nomogram integrates information from all six features simultaneously).
Calibration curves confirmed good agreement between the nomogram's predicted probability of SLN metastasis and the observed frequency in the data, validating that the nomogram's outputs are well-calibrated rather than systematically over- or underestimating risk. Decision curve analysis (DCA) further confirmed clinical utility, showing the highest net benefit when the clinical decision threshold (the minimum probability that would prompt SLNB) was set between 0.04 and 0.081 -- a realistic range for clinical practice.
The nomogram works by assigning a point score to each of the six predictor variables based on their values, summing these to a total score, and reading off the corresponding probability of SLN metastasis. This visual and intuitive format allows any clinician with the patient's preoperative MRI and clinical data to generate an individual predicted risk without requiring a computer or complex software -- an important practical advantage for broad clinical implementation.
Beyond the continuous probability estimate from the nomogram, the study developed a practical three-category risk stratification using recursive partitioning analysis (RPA) -- a tree-based method that identifies the most informative binary decision rules from the available data. The resulting classification uses only two variables from MRI: cortical thickness and ADC value.
The three risk groups were: low risk (cortical thickness of the suspicious lymph node 2.7 mm or less), with a 15.8% SLN metastasis probability; intermediate risk (cortical thickness greater than 2.7 mm and ADC 1.164 x 10-3 mm2/s or above), with 28.6% probability; and high risk (cortical thickness greater than 2.7 mm and ADC below 1.164 x 10-3 mm2/s), with 69.8% probability. The striking separation between low and high risk groups (15.8% vs. 69.8%) is clinically actionable.
The logic behind this stratification is biologically coherent: cortical thickness is the most visually accessible indicator of lymph node involvement on any imaging modality, and ADC value reflects cellular density within the lymph node itself -- a low ADC indicating a densely packed, metastasis-laden node that restricts water diffusion. Using these two parameters in sequence provides a rapid and reproducible triage that requires no complex calculation.
The most direct clinical implication is for the low-risk group (cortical thickness 2.7 mm or less), where the SLN metastasis probability is only 15.8%. The authors suggest that patients in this category -- representing approximately 61% of the study cohort -- might potentially be candidates for SLNB omission, pending prospective multicenter validation. A predicted metastasis rate of 15.8% is in a range where, depending on the clinical context and patient preferences, the risks of SLNB may outweigh the benefits.
The model offers particular advantages over axillary ultrasound in complex clinical scenarios. Standard axillary ultrasound is highly operator-dependent, and its accuracy is reduced in patients with deep-seated nodes or dense breast tissue. MRI provides a more comprehensive three-dimensional evaluation of both the primary tumor and the regional lymph nodes simultaneously, potentially better characterizing multifocal disease and capturing tumor biology through ADC measurements that ultrasound cannot provide.
The study's design -- deliberately excluding molecular markers like HER2, ER, PR, and Ki67 that may not be consistently available preoperatively at all clinical sites -- reflects a focus on creating a widely applicable tool rather than the most predictively powerful one. This pragmatic choice means the model is implementable at any center with 3T MRI capability, without requiring specific preoperative biopsy panels.
The ADC value from diffusion-weighted MRI reflects the degree to which water molecules can move freely within tissue. In normal lymph nodes, water diffuses relatively freely through the lymphoid tissue. When a lymph node is invaded by cancer cells, the high cellular density and altered tissue architecture restrict water movement, resulting in a lower ADC value. The inverse relationship between ADC and SLN metastasis probability observed in this study is therefore mechanistically grounded in tissue biology.
Cortical thickness reflects the degree of cortical expansion that occurs as malignant cells colonize a lymph node. Normal lymph node cortices are typically thin and uniform; as metastatic cells accumulate, the cortex thickens and becomes irregular. A threshold of 2.7 mm identified in this study is consistent with prior literature that has identified cortical measurements as among the most reliable ultrasound and MRI criteria for malignancy assessment.
The finding that multifocal tumors carry approximately twice the SLN metastasis risk of unifocal tumors is biologically plausible: multifocal disease represents a broader tumor-tissue interface with more opportunity for lymphovascular invasion, and multifocality often indicates more aggressive tumor biology. This association aligns with results from prior studies (Song et al., Xue et al.) and supports the inclusion of multifocality as a standard preoperative assessment criterion.
This study developed a preoperative prediction model for sentinel lymph node metastasis that achieved AUC 0.847, using six MRI-derived and clinical predictors available before surgery. The accompanying RPA-based three-tier risk stratification provides a simple framework for patient counseling and surgical decision-making. Together, the nomogram and risk classification offer a clinically accessible tool for individualized preoperative assessment.
Key limitations include the single-center retrospective design, the exclusion of preoperative molecular markers (HER2, ER, PR, Ki67) that could improve model performance, and the absence of direct one-to-one confirmation between the MRI-measured nodes and the surgically identified sentinel nodes. Because sentinel node identification was not performed preoperatively, the MRI measurements were taken from the most morphologically suspicious level I node, which may not always correspond to the actual sentinel node harvested during surgery.
Future work should include prospective multicenter validation to confirm model performance across different patient populations and imaging equipment, integration of preoperative molecular marker data, and studies incorporating preoperative sentinel node localization to enable direct imaging-to-pathology correlation. If validated, this approach could provide a foundation for identifying a substantial proportion of breast cancer patients who can safely avoid SLNB entirely.