MRI-based radiomics analysis: comparative application of different machine learning models in predicting pathological complete response after neoadjuvant chemotherapy for breast cancer

To develop and compare machine-learning (ML) models based on dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) radiomics, clinical variables, and their combination for predicting pathological complete response (pCR) after neoadjuvant chemotherapy (NAC) in breast cancer, and to critically evaluate the stability of each model in an independent test cohort. This retrospective single-center study included 272 patients treated between May 2020 and May 2025 who were retrospectively analyzed. pCR was determined using the Miller-Payne grading system, corresponding to ypT0/Tis ypN0 (no residual tumor in sampled axillary nodes). In total, 1,834 features were extracted from pretreatment phase-3 DCE-MRI with PyRadiomics. All feature selection (Mann–Whitney U test, Spearman redundancy pruning, LASSO) and hyperparameter tuning were performed exclusively in the training cohort ( n = 217); the held-out test cohort ( n = 55) was used once for final evaluation. LR, a radial-basis-function (RBF) SVM, and an MLP were compared, together with a clinical signature (HER2 and axillary nodal status) and a radiomics–clinical nomogram; discrimination (AUC) and decision curve analysis (DCA) were assessed. Of 272 patients, 98 (36.03%) achieved pCR. Mann–Whitney screening and Spearman pruning retained 34 features, from which LASSO selected 17 with non-zero coefficients. The SVM-based radiomics model performed best (training AUC: 0.873; test AUC: 0.782). The clinical signature was the most stable model (training AUC 0.885; test AUC 0.806), whereas the nomogram, despite the highest training AUC (0.901), showed a marked decline in the test cohort (AUC 0.686, 95% CI 0.350–1.000), and its DCA advantage was not reproduced in the small test cohort. HER2 positivity and node-negative disease robustly identified patients likely to achieve pCR, and MRI radiomics provided moderate complementary discrimination; however, the incremental benefit of the combined nomogram was not confirmed in the independent test cohort, indicating possible overfitting. The radiomics-based models require stratified or external validation before clinical deployment.