Abstract
BACKGROUND: Axillary lymph node metastasis (ALNM) is pivotal for breast cancer treatment and prognosis. Invasive tests may carry complications, while non-invasive methods like physical examination have poor accuracy. Existing AI-based models rely mostly on tumor-centric features. However, metastatic lymph nodes show neoangiogenesis and altered hemodynamics, leading to time-intensity curve (TIC) profiles similar to those of adjacent vessels. This study aimed to quantify these hemodynamic disparities from dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) and enhance ALNM prediction accuracy. METHODS: A retrospective study included 186 patients (92 ALNM+, 94 ALNM-). Axillary vessels and lymph nodes were semi-automatically segmented via Hessian matrix algorithms. Four TIC-derived features [lymph-node TIC area (LNTICA), the difference in TIC area between the vessel ROI and lymph-node ROI (DIFF), non-negative area difference (NON-NEG DIFF), ratio of area difference to lymph node TIC area (RATIO TO LYMPH)] were extracted. ResNet50 extracted image features, and a Stacking framework integrated image, clinical, and TIC features, using support vector machine (SVM) and nomogram as classifiers. Statistical tests (F-test, t-test, and Kolmogorov-Smirnov test) validated feature discriminability. RESULTS: All TIC features differed significantly between groups (P<0.001 for F-test, P<0.001 for t-test/Kolmogorov-Smirnov test for NON-NEG DIFF and RATIO TO LYMPH). RATIO TO LYMPH (mean: 0.11 vs. 0.36) showed optimal discriminability. Integrating TIC features improved area under the receiver operating characteristic curve (AUC): SVM (0.876→0.914) and nomogram (0.902→0.941) in the test set. SHapley Additive exPlanations (SHAP) analysis confirmed RATIO TO LYMPH as one of the top predictive features. CONCLUSIONS: Lymph node-vessel hemodynamic disparities are robust ALNM biomarkers. Integrating these TIC-derived features with clinical and image data significantly enhances prediction accuracy, providing a non-invasive tool for clinical decision-making.