Value of multi-parameter MRI radiomics features in the preoperative prediction of triple-negative and non-triple-negative breast cancer / 中华放射学杂志

ABSTRACT

Objective:To explore the value of radiomics features extracted from multi-parameter MRI (mp-MRI) in preoperative prediction of triple negative breast cancer (TNBC) and non triple negative breast cancer (NTNBC).Methods:The clinical and preoperative-MRI data of 371 patients with breast cancer confirmed by surgical pathology from January 2017 to July 2019 in Henan Provincial People′s Hospital were retrospectively analyzed. Based on the results of immunohistochemistry (IHC) and fluorescence in situ hybridization (FISH) from postoperative pathological specimens, the cancer was classified as TNBC and NTNBC. Patients were randomly assigned to a training set ( n=250) and a validation set ( n=121). Quantitative radiomics features were extracted from three-dimensional lesions based on dynamic contrast enhanced-T 1WI (DCE-T 1WI) and fat-suppressed T 2WI sequences, and 32 quantitative radiomics features were selected by Mann-Whitney U test, elastic network, and support vector machine recursive feature elimination. Three radiomics signatures were constructed by using the algorithm of support vector machine based on the quantitative radiomics features extracted from fat-suppressed T 2WI, DCE-T 1WI and the mp-MRI of their combination. The prediction performances were evaluated by receiver operating characteristic (ROC) curve and the area under the ROC curve, accuracy, sensitivity, and specificity were calculated. Results:There were 61 patients with TNBC and 310 patients with NTNBC. The clinicopathological characteristics between NTNBC and TNBC were statistically different in the pathological grade (χ2=24.544, P<0.001). Other clinicopathological characteristics (age, maximum diameter of mass, vascular tumor thrombus, axillary lymph nodes) were not statistically differences between NTNBC and TNBC ( P>0.05). The radiomics signature presenting the best performance for predictive TNBC and NTNBC were based on mp-MRI radiomics features. The area under the ROC curve, accuracy, sensitivity, and specificity were 0.91[95% confidence interval (CI) 0.881-0.932], 86.0%, 84.4% and 86.3% in training set, and 0.84 (95%CI 0.807-0.868), 75.2%, 68.7% and 76.1%, in validation set, respectively. Conclusion:Radiomics based on mp-MRI features can be a effectively potential tool for predictive TNBC and NTNBC breast cancer and provide scientific basis for clinicians to make treatment decisions.