Magnetic resonance imaging-based radiomics analysis: predicting vascular invasion in breast invasive ductal carcinoma using different machine learning models.

Background: Lymphovascular invasion (LVI) is an adverse prognostic factor; preoperative prediction by imaging is difficult, but radiomics extracts quantitative tumor biology features. Investigating the value of combining magnetic resonance imaging (MRI)-based radiomics features with multiple machine learning (ML) models in predicting LVI status in invasive ductal carcinoma (IDC) of the breast. Methods: A retrospective cohort of 678 female patients with pathologically confirmed IDC of the breast was collected from June 2021 to June 2025. All patients underwent preoperative MRI. Based on postoperative pathology, patients were categorized into LVI-positive (n=258) and LVI-negative (n=420) groups. Using ITK-SNAP software, regions of interest (ROIs) were delineated in phase-3 dynamic contrast-enhanced MRI images to extract radiomics features. Feature selection and dimensionality reduction were performed using redundancy analysis and the least absolute shrinkage and selection operator (LASSO) regression. Data were randomly split into an 8:2 ratio for training (n=542) and testing (n=136) sets. Eight ML models were then constructed: logistic regression (LR), support vector machine (SVM), K-nearest neighbors (KNN), random forest, extreme random trees (ExtraTrees), extreme gradient boosting (XGBoost), light gradient boosting machine (LightGBM), and multi-layer perceptron (MLP). Univariate and multivariate LR analyses were performed to screen clinical and radiological features for establishing clinical models. Concurrently, a combined model integrating radiomics features with clinical characteristics was developed. The discriminatory power of each model was evaluated using the area under the curve (AUC). AUC values for the radiological model, clinical model, and combined model underwent statistical comparison via Delong's test. Decision curve analysis (DCA) was employed to assess their clinical utility. Results: A total of 1,197 radiomics features were extracted, and after dimensionality reduction, 23 features with the highest predictive value were selected. The clinical prediction model constructed based on multifactorial analysis results indicated that LVI positivity was more likely to occur in postmenopausal patients [odds ratio (OR) =1.690; 95% confidence interval (CI): 1.174-2.433], those with higher histological grade (OR =1.527; 95% CI: 1.107-2.107), sentinel lymph node metastasis (OR =0.198; 95% CI: 0.137-0.285), distinct molecular subtypes (OR =0.740; 95% CI: 0.567-0.965), and MRI maximum diameter ≥2 cm (OR =2.059; 95% CI: 1.362-3.113). Among radiomics models, the XGBoost model demonstrated optimal performance with a training set AUC of 0.912 and a validation set AUC of 0.706. The combined model exhibited the highest discriminatory ability in the training set (AUC =0.956) and a validation set AUC of 0.778. DCA indicated the combined model provided higher clinical net benefit. Conclusions: A combined model incorporating MRI radiomics features and clinical factors demonstrates predictive value for the presence or absence of LVI in IDC of the breast, serving as a reference for individualized treatment decisions.