Predicting dynamic changes of Ki-67 in breast cancer after neoadjuvant therapy based on multi-phase DCE-MRI delta-radiomics.

Background: Ki-67 is a key biomarker of tumor proliferation in breast cancer. A reduction in Ki-67 following neoadjuvant therapy (NAT) reflects chemosensitivity and holds significant prognostic value. Therefore, pre-treatment assessment of Ki-67 dynamics during NAT is crucial for evaluating patient prognosis.This study aims to predict the change in Ki-67 index in breast cancer patients following NAT using radiomic features derived from DCE-MRI. Methods: This retrospective study enrolled 148 breast cancer patients who underwent surgical resection after 6-8 cycles of NAT, randomly divided into training (n=104) and test cohorts (n=44) at a 7:3 ratio. Multivariable logistic regression (P<0.05) identified independent clinical risk factors for Ki-67 downgrading. Radiomics features were extracted from pre-treatment DCE-MRI scans from the early, peak, and delayed phases, along with the corresponding phase differences (delayed-early, delayed-peak, peak-early). Feature selection was performed with Principal Component Analysis (PCA) followed by Recursive Feature Elimination (RFE), and radiomics and delta-radiomics models were built using the LR-Lasso algorithm.The DeLong test compared AUC values to identify the optimal model. Top radiomics features were then combined with clinical factors to construct a hybrid model, evaluated by AUC, calibration, and decision curve analysis (DCA). Results: ROC curve analysis demonstrated that the peak-to-early delta-radiomics model achieved the best diagnostic performance in the testing cohort with an AUC of 0.817(95% CI: 0.685-0.949), significantly outperforming the delayed-to-early delta model [AUC = 0.648(95% CI: 0.484-0.812)]and the standalone peak-phase model [AUC = 0.615(95% CI: 0.444-0.785)]. Logistic regression analysis revealed that HER2 status (p = 0.031) and histological grade (p<0.001) were significant predictors for constructing the clinical-radiological model. Then, integrating the optimal delta-radiomics model with independent clinical-radiological risk factors to form a combined model increased the AUC to 0.851(95% CI: 0.740-0.960), which was significantly superior to both the clinical-radiological model alone [AUC = 0.785(95% CI:0.649-0.922)]and the delta-radiomics model alone. In the external validation cohort, this model [AUC = 0.919(95% CI:0.846-0.992)] also demonstrated superior performance compared to either the standalone clinical-radiological model [AUC = 0.801(95% CI:0.680-0.924)] or the delta-radiomics model [AUC = 0.803(95% CI:0.678-0.928)]. Conclusions: Delta-radiomics based on MRI, combined with clinical parameters, represents a promising non-invasive approach for more accurately predicting Ki-67 downstaging in breast cancer following NAT, outperforming conventional radiomics models. Integrating radiomic features with clinical information holds the potential to further optimize individualized treatment strategies and improve prognostic assessment for breast cancer patients.