Objective:To create radiomics models based on abbreviated multimodal magnetic resonance imaging (MRI) for the diagnosis of breast cancer.Methods:All breast MR imaging data between Jun 2014 and Mar 2019 were retrospectively collected. Patients with pathological results of puncture or surgical resection were involved in this study. One thousand three hundred and six patients (416 benign and 890 breast cancer) were divided into training cohort ( n=702), internal validation cohort ( n=302), and external validation cohort ( n=302). All images were reduced to: the joint model group [including T2 weighted imaging (T2WI), DWI (diffusion-weighted imaging) and first contrast-enhanced sequences], non-enhanced group (T2WI and DWI) and single-phase enhanced group (first contrast-enhanced sequences). Analysis of variance (ANOVA) and least absolute shrinkage and selection operator (LASSO) were used to reduce the dimension of texture features. Three supervised machine learning algorithms (Bagging decision tree, Gaussian process, support vector machine) were used to predict benign and malignant breast lesions, and the best classifier was selected to construct breast cancer diagnosis model. Models were validated by internal and external validation cohorts. Results:The Gaussian process algorithm was chosen. The area under the curve (AUC) of the joint model and the non-enhanced model for predicting breast cancer were 0.903 and 0.893 for the training cohort, 0.893 and 0.863 for the internal validation cohort, and 0.878 and 0.864 for the external validation cohort.Conclusions:The radiomics model based on abbreviated multimodal MRI can accurately diagnose breast cancer. And the non-enhanced model can accurately diagnose breast cancer without contrast enhancement, which provides feasibility for simplifying the diagnosis process.

Objective:To investigate the feasibility of developing a radiomics model based on MRI and clinical features to predict the PD-L1 level in breast cancer.Methods:A total of 139 consecutive patients with breast cancer confirmed by pathology were enrolled retrospectively, including 79 PD-L1 negative patients and 60 PD-L1 positive patients. All patients were randomly assigned to a training dataset( n=97) and a validation dataset( n=42). Radiomics features were extracted from dynamic contrast-enhanced MRI. Radiomics feature selection was generated through the analysis of variance(ANOVA), least absolute shrinkage and selection operator(LASSO). Radiomics model and comprehensive model were developed for predicting the level of PD-L1. The receiver operating characteristic curve(ROC) was used to evaluate the predictive capacity of the models. Results:The radiomics model exhibited good performance in the training and validation datasets, with an area under the curve(AUC) of 0.847(95% confidence interval CI: 0.770-0.924) and 0.826(95% CI: 0.699-0.954), respectively. Compared with the radiomics model, the clinical feature combined prediction model showed better results, with AUC of 0.919(95% CI: 0.868-0.970) and 0.882(95% CI: 0.782-0.982), respectively, but without statistically significant difference( Z=1.32, P=0.19), respectively, but without statistically significant difference. Conclusions:The radiomi.Conclusions:The radiomics model has a certain value in preoperative prediction of PD-L1 expression level in breast cancer, which may be used as a supplement and improvement to the pathological gold standard to provide support for clinical decision-making.