Preoperative breast cancer subtyping using multiparametric magnetic resonance imaging and explainable machine learning.

Breast cancer exhibits substantial molecular heterogeneity with distinct molecular subtypes that differ markedly in prognosis, treatment response, and clinical management. To enable imaging-based subtype differentiation, this study retrospectively analyzed suspected breast cancer patients who underwent breast MRI between September 2020 and September 2024, finally including 292 breast cancer lesion records from 283 patients with postoperative IHC-confirmed subtypes. All patients received SyMRI, DCE-MRI, and DWI. Quantitative metrics were extracted-pre- and post-contrast T1, T2, and PD values, T1-Ratio, and PD-Ratio, ADC, TIC patterns, BI-RADS categories, and morphological features. Elastic-net logistic regression was embedded within the training folds for feature selection, and five machine-learning classifiers-LR, SVM, KNN, RF, and DT-were evaluated using the historical internal test set and repeated patient-grouped nested cross-validation. After false-discovery-rate adjustment, ADC-Tumor, post-contrast T1, post-contrast T2, T1-Ratio, and Burr sign differed significantly across molecular subtypes. Repeated patient-grouped nested validation selected the random-forest model (macro-average one-vs-rest AUC, 0.946 ± 0.035), which achieved a macro-average AUC of 0.862 (95% CI, 0.789-0.927) in the historical internal test set; class-specific results and confusion matrices were reported with uncertainty intervals. TreeSHAP analysis of the selected random-forest model identified post-contrast T1 (overall mean absolute SHAP, 0.053), post-contrast T2 (0.053), and T1-Ratio (0.050) as the leading overall contributors, with distinct subtype-specific patterns. These findings support further investigation of multiparametric breast MRI as a preliminary adjunct for preoperative subtype assessment; the model cannot replace histopathological and IHC confirmation and requires independent external validation.