Multiparametric MRI-Based Radiomics For Pretreatment Prediction Of Pathological Complete Response To Neoadjuvant Chemotherapy In Breast Cancer

Purpose:To assess the potential of radiomics analysis of pretreatment multiparametric MRI in breast cancer to predict pathological complete response(p CR)after neoadjuvant chemotherapy(NAC).Methods:This retrospective study included 204 women with stage II–III unilateral invasive breast ductal carcinoma between January 2015 and December 2021.All patients received the standard 6-8 cycles of NAC,followed by breast surgery at our institution.The patients were randomly assigned to a cross-validation(CV)set and a test set(7:3 ratio).Quantitative imaging features were extracted from preoperative breast MRI for each patient,and the dimensionality was reduced using the student t-test and Elastic Net regression.Then,we utilized the Mann-Whitney U test,chi-square test,or Fisher’s exact test to evaluate clinical predictors of p CR.Finally,we constructed three logistic regression classifier models for p CR prediction based on the radiomics features and predictive clinical risk factors,namely,the LR-R model with only radiomics features,the LR-C model with only clinical risk factors,and the LR-RC model with the fusion of both.To compare the performance of different models,receiver operating characteristic(ROC)curves and precision-recall(P-R)curves were further plotted,and the area under the ROC curve(AUC)and average precision(AP),among others,were calculated to evaluate the performance of the prediction models.Results:In univariate analysis,five clinical characteristics showed significant differences(P < 0.05)between the p CR and non-p CR groups.20 radiomics features were selected to construct the radiomics signature.The AUC(95%CI)of the LR-C and LR-R models in the CV set were 0.768(0.689-0.847)and 0.821(0.747-0.894),respectively,while in the test set were 0.805(0.692-0.919)and 0.812(0.679-0.945),respectively.The LR-RC model that combined radiomics features and clinical information showed the optimal diagnostic performance,with the AUC(95% CI)of 0.880(0.824-0.935)in the CV set and 0.888(0.800-0.975)in the test set,respectively.Conclusions:This study demonstrated that the fusion model combining the radiomics features of multiparametric MRI with clinical risk factors could provide a potential tool to predict p CR in patients with breast cancer receiving NAC before treatment.