Preoperative prediction of lymphovascular invasion in breast cancer using multicenter DCE-MRI radiomics and clinical features

Lymphovascular invasion (LVI) is a key histopathological factor driving the metastatic process in invasive breast cancer; however, a definitive diagnosis is typically only possible postoperatively. Radiomics based on dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) has shown promise for the preoperative prediction of LVI. Nevertheless, existing studies are largely single-center, lack systematic comparisons of machine learning algorithms, and have insufficient external validation. This study developed and validated multicenter DCE-MRI radiomics models for preoperative LVI prediction, systematically compared ten machine learning algorithms, and quantified the incremental value of radiomics over clinical variables available before surgery. This retrospective multicenter study included 935 female patients with pathologically confirmed invasive ductal carcinoma from three institutions. All patients underwent preoperative 3.0 T DCE-MRI. After manual three-dimensional delineation of regions of interest on DCE-MRI images, 1,197 radiomic features were extracted with PyRadiomics after image resampling to 2 × 2 × 2 mm³ (B-spline interpolation), intensity normalization, and fixed-bin-width discretization. Feature selection (Mann-Whitney U test, Pearson correlation pruning with |r| > 0.9, and least absolute shrinkage and selection operator [LASSO] regression) was confined to the training cohort. Ten classifiers were compared: logistic regression (LR), support vector machines (SVM), k-nearest neighbors (KNN), random forests (RF), Extremely Randomized Trees (ExtraTrees), XGBoost, LightGBM, Gradient Boosting, AdaBoost, and multi-layer perceptron (MLP). A primary clinical model was built using only variables available before surgery (age, histological grade, Ki-67, MRI axillary lymph node status, and MRI tumor size); a sensitivity clinical model additionally included postoperative sentinel lymph node (SLN) status. Combined clinical-radiomics models and a nomogram were constructed, and all probability thresholds were fixed at the value maximizing the Youden index in the training set. Performance was assessed by the area under the receiver operating characteristic curve (AUC) with the DeLong test, calibration intercept and slope, Brier score, and decision curve analysis (DCA) in a training set, an internal validation set, and two independent external test sets. In the multivariable analysis, SLN status (odds ratio [OR] 5.743, 95% confidence interval [CI]: 3.693–8.931, p < 0.001) and MRI tumor size (OR 1.413, 95% CI: 1.186–1.684, p < 0.001) were independently associated with LVI; because SLN status is established intraoperatively or postoperatively, the primary clinical model used only the five preoperative variables. Among the ten radiomics algorithms, LR and MLP showed the most stable cross-center performance (LR: AUC 0.841, 0.716, 0.632, and 0.586 in the training, internal validation, test 1, and test 2 cohorts, respectively), whereas complex ensemble methods overfit (Gradient Boosting: 0.991 in training vs. 0.696 and 0.532 externally; XGBoost: 0.894 vs. 0.613 and 0.453). The preoperative combined model achieved AUCs of 0.856, 0.721, 0.665, and 0.603, respectively, significantly outperforming the preoperative clinical model in the training, validation, and test 2 cohorts (DeLong p < 0.001, p = 0.013, and p < 0.001) but not in test 1 ( p = 0.306), and significantly outperforming the radiomics model only in test 1 ( p = 0.026). The incremental value of radiomics was significant in the training cohort (integrated discrimination improvement p < 0.001) but not in the external cohorts. Calibration slopes of the combined model declined externally (0.49, 0.28, and 0.21), indicating that probability estimates require local recalibration. In this multicenter study, LR and MLP were the most robust radiomics classifiers for preoperative LVI prediction, whereas complex ensemble algorithms were prone to overfitting on external validation. The preoperative combined model improved discrimination and clinical net benefit over the clinical model internally, but its external incremental value was inconsistent and its calibration deteriorated in the external cohorts. A sensitivity analysis using postoperative SLN status showed higher absolute performance, underscoring that genuinely preoperative tools trade performance for applicability. These findings emphasize rigorous multicenter benchmarking, recalibration before local deployment, and honest reporting of external performance in radiomics research.