Background and Objectives: The prediction of pathological response outcome following neoadjuvant chemotherapy (NAC) in breast cancer is an important clinical task that aids treatment planning and personalized therapeutic strategies. Traditional prediction strategies require expert knowledge and are sometimes influenced by False Positive and False Negative outcomes. Methods: In this study, we propose a new multi-modal machine learning framework that combines the structural, textural and functional imaging biomarkers obtained from Dynamic Contrast-Enhanced MRI (DCE-MRI) for non-invasive classification of complete response (CR) and partial response (PR). The clinical dataset used in our study was collected at Annunziata Hospital, Cosenza, Italy (33 breast cancer patients total with confirmed response labels). The method investigates three complementary feature spaces: (i) quantitative perfusion and geometry parameters obtained from time–intensity curves capturing tumor vascular dynamics, (ii) radiomic features encoding intra-tumoral structural and textural heterogeneity, and finally (iii) a modal-aware late fusion using a Bayesian meta-learner integrating both descriptors. Results: To prevent overfitting on medical imaging datasets, a stringent validation approach based on Nested Leave-One-Out Cross-Validation (Nested LOOCV) in conjunction with repeated Stratified 5 × 20 cross-validation was implemented. Accuracy was as high as 0.8955 with perfusion-based models, and a radiomic model yielded an accuracy score of 0.8982 using Mann–Whitney feature selection under Nested LOOCV. Overall, the multi-modal fusion approach achieved superior performance (0.9115 accuracy, 0.9054 F1-score, and 0.9488 ROC-AUC) using Random Forest classifier, which indicates that combining perfusion dynamics with radiomics enables a more clinically-relevant prediction of NAC response. Conclusions: This shows that the combination of perfusion dynamics with radiomic descriptors from DCE-MRI allows for better and more reliable prediction of NAC response. This proposed multimodal framework also emphasizes the promising role of imaging biomarkers in facilitating tailored treatment approaches for breast cancer.
Integrating radiomic and perfusion kinetics from DCE-MRI for accurate prediction of pathological response to neoadjuvant chemotherapy
Haddadi Y. R.;Hazarika R. A.;Raza A.;Buonanno G.;Maggiolini M.;Costanzo S.;Guzzo A.;Fortino G.
2026-01-01
Abstract
Background and Objectives: The prediction of pathological response outcome following neoadjuvant chemotherapy (NAC) in breast cancer is an important clinical task that aids treatment planning and personalized therapeutic strategies. Traditional prediction strategies require expert knowledge and are sometimes influenced by False Positive and False Negative outcomes. Methods: In this study, we propose a new multi-modal machine learning framework that combines the structural, textural and functional imaging biomarkers obtained from Dynamic Contrast-Enhanced MRI (DCE-MRI) for non-invasive classification of complete response (CR) and partial response (PR). The clinical dataset used in our study was collected at Annunziata Hospital, Cosenza, Italy (33 breast cancer patients total with confirmed response labels). The method investigates three complementary feature spaces: (i) quantitative perfusion and geometry parameters obtained from time–intensity curves capturing tumor vascular dynamics, (ii) radiomic features encoding intra-tumoral structural and textural heterogeneity, and finally (iii) a modal-aware late fusion using a Bayesian meta-learner integrating both descriptors. Results: To prevent overfitting on medical imaging datasets, a stringent validation approach based on Nested Leave-One-Out Cross-Validation (Nested LOOCV) in conjunction with repeated Stratified 5 × 20 cross-validation was implemented. Accuracy was as high as 0.8955 with perfusion-based models, and a radiomic model yielded an accuracy score of 0.8982 using Mann–Whitney feature selection under Nested LOOCV. Overall, the multi-modal fusion approach achieved superior performance (0.9115 accuracy, 0.9054 F1-score, and 0.9488 ROC-AUC) using Random Forest classifier, which indicates that combining perfusion dynamics with radiomics enables a more clinically-relevant prediction of NAC response. Conclusions: This shows that the combination of perfusion dynamics with radiomic descriptors from DCE-MRI allows for better and more reliable prediction of NAC response. This proposed multimodal framework also emphasizes the promising role of imaging biomarkers in facilitating tailored treatment approaches for breast cancer.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
