Background: The accurate prediction of response to neoadjuvant therapy (NAT) is crucial for optimizing breast cancer management. Conventional breast Dynamic Contrast-Enhanced Magnetic Resonance Imaging (DCE-MRI) radiomics typically relies on single post-contrast phases and may not fully capture temporal enhancement patterns related to tumor heterogeneity. This study evaluated a machine learning model based on time-dependent radiomic features extracted from pre-treatment DCE-MRI for predicting NAT response in breast cancer patients. Methods: Breast DCE-MRI examinations of women scheduled for NAT, acquired on 1.5 T scanners from three different vendors, were retrospectively collected from two centers. Tumors were automatically segmented on the third post-contrast DCE image using a 3D nnUNet model trained on 30 lesions. All DCE phases were registered to the reference image, and radiomic features were extracted from a consistent tumor region of interest across all phases. Time-dependent radiomic features were computed using linear regression modeling of feature evolution over time. A random forest classifier integrating static and time-dependent radiomic features was developed to predict pathological complete response (pCR), partial response (pPR), and non-response (pNR). Model performance was evaluated using internal validation (Center 1) and an independent external test cohort (Center 2). Results: A total of 212 patients were included (173 from Center 1 and 39 from Center 2), comprising 103 pCR, 103 pPR and 6 pNR cases. Among 759 extracted features, 30 showed significant differences across response groups. Several time-dependent texture features related to intratumoral heterogeneity were significantly associated with pNR. The model achieved AUC values of 0.80, 0.81, and 0.95 in the internal validation cohort and 0.75, 0.74, and 0.86 in the external test cohort for predicting pCR, pPR, and pNR, respectively. Conclusions: Time-dependent radiomic features derived from pre-treatment breast DCE-MRI enable the accurate prediction of response to NAT, with particularly strong performance in identifying non-responders. This approach may support imaging-based risk stratification and contribute to more personalized treatment.
Vatteroni et al. (Thu,) studied this question.