Radiomics Study Of Molecular Subtypes Of Breast Cancer Based On Dynamic Contrast Enhanced-Magnetic Resonance Imaging

Objective: Breast cancer is the most common malignant tumor threatening women’s health worldwide.In order to discover the potential relationship between medical images and biological information of tumors,this study intends to construct radiomics models based on dynamic contrast enhanced magnetic resonance imaging(DCE-MRI)and evaluate its performance in predicting the molecular subtypes of breast cancer,so as to realize the analysis of breast tumor heterogeneity.Methods: This study was divided into two parts.The first part included 186 female breast cancer patients who received DCE-MRI examination,including 111 cases with high Ki-67 expression status and 75 cases with low Ki-67 expression status.Three functional parameter maps of wash in,wash out and signal enhancement ratio and apparent diffusion coefficient(ADC)maps were obtained by using the pre-and post-enhancement images and the diffuse-weighted imaging(ADC)images with different b-values obtained from the image storage system.Then,946 radiomics features were extracted from three DCE-MRI functional parameter maps and ADC maps respectively,and then feature screening was performed using the least absolute shrinkage and selection operator(LASSO)method.Finally,six support vector machine classifiers were trained by combining features obtained from different parameter graphs to predict the expression level of Ki-67.In the second part of the study,153 women with an average age of 50.08 years who underwent DCE-MRI examination and were pathologically diagnosed with breast cancer were retrospectively included.According to the time required for the contrast agent to reach its peak,the entire tumor region was divided into three intratumoral subregions,and 540 texture features were extracted from the three subregions and the entire tumor region.LASSO method was carried out for feature selection and radiomics score(Rad-score)construction.Two machine learning classifiers were developed using the optimal features obtained by filtering,and multivariate logistic regression analysis was applied to select independent predictors from Rad-score and clinical factors to establish the nomogram.In both studies,indicators such as area under receiver operating characteristic curve(AUC),sensitivity and specificity were used to assess the predictive performance of models.Results: Of the 3,792 features originally obtained in the first part of this study,a total of2,622(69.1%)had good inter-observer agreement(ICC > 0.8).After feature selection,3,8,3 and 11 best features were finally obtained from signal enhancement ratio,wash in,wash out function parameter maps and ADC maps respectively.Combined model 2,which was obtained by adding the features of ADC maps to the functional parameter maps model(Combined model 1),improved the performance of Ki-67 status prediction.The best prediction effect was obtained in the training cohort and the validation cohort,with the AUC values of 0.875 and 0.785,respectively.In addition,compared with the single parameter maps model,the predictive performance of the model(Combined model 2)obtained by fusing the features of each image was significantly improved in the training cohort(P < 0.05).In the second part of this study,5,1,5 and 1 features were selected from the three intratumoral subregions of rapid,medium and slow and the whole tumor region,respectively.The predictive performance of texture features from the rapid subregion was the best among the three intratumoral subregions,and the AUC values in the training cohort and validation cohort were 0.805 and 0.737,respectively.Therefore,the optimal features of the rapid subregion were used to calculate the Rad-score.Independent predictors from univariate and multivariate analyses were then used to construct nomogram.Nomogram including Rad-score,HER2 status,and histological grading showed good discrimination of breast cancer luminal types,with AUC values of 0.830 and 0.879 in the training and validation cohorts,respectively.Calibration curves and decision curve analysis reflected good calibration performance and clinical usefulness of nomogram.Conclusion: DCE-MRI based radiomics model could effectively distinguish Ki-67 status and luminal type of breast cancer.The nomogram combined with medical imaging and clinical factors had potential clinical value and could help doctors predict the molecular subtypes of breast cancer.