Purpose: To evaluate whether radiomic features extracted from first post-contrast DCE-MRI improve the prediction of breast cancer molecular subtypes beyond conventional MRI features, and to compare the performance of conventional MRI-based, radiomics-based, and combined models. Materials and Methods: In this retrospective, single-center cross-sectional study with prediction modelling, 206 consecutive patients with pathologically confirmed primary breast cancer who underwent pretreatment 3.0-T breast MRI between January 2018 and January 2026 were included. Molecular subtypes were assigned using immunohistochemistry, with HER2-equivocal cases confirmed by FISH/dual ISH. Conventional MRI features were assessed according to the BI-RADS MRI lexicon. Radiomic features were extracted from a manually segmented 2D ROI on the first post-contrast DCE-MRI image, harmonized with ComBat, filtered for reproducibility, and selected using LASSO. One-vs-rest logistic regression models based on MRI features alone, radiomics alone, and combined features were constructed and compared. Results: The cohort comprised 81 Luminal A (LA), 70 Luminal B (LB), 31 HER2-enriched (HER2), and 24 triple-negative breast cancers (TNBC). On multivariable analysis, LA tumors were less likely to have positive lymph nodes (OR, 0.31; p = 0.001), LB tumors were associated with multifocality (OR, 2.22; p = 0.004), HER2 tumors were less likely to present as a mass lesion (OR, 0.21; p = 0.020), and TNBC was strongly associated with rim enhancement (OR, 12.42; p < 0.001). The combined model yielded the highest AUCs for LA (0.788), LB (0.732), HER2 (0.858), and TNBC (0.890), compared with MRI models (AUC range, 0.616-0.765) and radiomics models (AUC range, 0.692-0.853). Pairwise DeLong comparisons showed subtype-dependent incremental value, with the clearest added value for HER2 and additional improvement over MRI models for LB and TNBC. Conclusion: DCE-MRI radiomics adds incremental value to conventional MRI for predicting breast cancer molecular subtypes. Integrating radiomic and morphologic MRI features provides the best discrimination, particularly for HER2 and TNBC.
Hue et al. (Wed,) studied this question.