Abstract
BACKGROUND: Breast cancer is the most commonly diagnosed malignancy in women, and accurate preoperative assessment of axillary lymph node (ALN) status is pivotal for staging and treatment planning. However, conventional imaging has variable accuracy and sentinel lymph node biopsy (SLNB), though less invasive than ALN dissection (ALND), still entails morbidity, underscoring the need for non-invasive prediction. The objective of this study was to construct and validate a predictive nomogram that integrates multiphase dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI)-derived radiomics features with key clinical indicators, aiming to non-invasively estimate the likelihood of ALN metastasis (ALNM) in individuals diagnosed with breast cancer before surgical intervention. METHODS: A total of 209 patients with histologically confirmed breast cancer were retrospectively enrolled and randomly divided into a training set (n=146) and a validation set (n=63). Radiomics features were extracted from three temporal post-contrast DCE-MRI sequences representing early-enhancement phase [first post-contrast phase (DCE1)], peak-enhancement phase [third post-contrast phase (DCE3)], and delayed-enhancement phase [seventh post-contrast phase (DCE7)]. Using the extremely randomized trees (ExtraTrees) machine learning algorithm, four classification models were developed: one for each individual phase and a combined model incorporating all three phases. Multivariate logistic regression was employed to construct a nomogram by integrating the selected radiomics features with significant clinical predictors. RESULTS: In the training set, the DCE3 model achieved the highest performance among single-phase models [area under the curve (AUC) =0.954], while in the validation set, the combined model integrating DCE1, DCE3, and DCE7 outperformed others (AUC =0.904). Notably, in the validation set, the delayed-phase DCE7 model (AUC =0.733) outperformed both DCE1 (AUC =0.709) and DCE3 (AUC =0.650), indicating its unique role in capturing stromal and microenvironmental features associated with metastasis. The final nomogram integrating radiomics and clinical variables demonstrated excellent discrimination (AUC =0.940 in training and 0.922 in validation), outperforming the clinical-only model (validation AUC =0.572) and radiomics-only model (validation AUC =0.904). It also yielded the highest sensitivity (0.947) and F1 score (0.766) in the validation set. Calibration curves and decision curve analysis (DCA) confirmed its predictive reliability and clinical utility. CONCLUSIONS: This study presents and validates a nomogram that combines multiphase DCE-MRI-derived radiomics features with clinical parameters to non-invasively and accurately predict ALNM in breast cancer patients. The model offers a valuable tool to support personalized surgical decision-making and may help reduce unnecessary axillary interventions.