MRI-based Radiomics Signature for Screening Lymphovascular Invasion in Breast Cancer Patients

Abstract Objective The early detection of lymphovascular invasion (LVI) is critical for the effective treatment of breast cancer (BC). This study aimed to investigate a non-invasive radiomics model based on MRI sequences for LVI screening in BC patients. Methods A total of 454 BC patients were enrolled in our study, with 150 in the LVI group and 304 in the non-LVI group. Radiomics features were extracted from MRI scans, including T2WI and DCE sequences, using LASSO analysis. Common machine learning algorithms (including LR, RF, KNN, SVM, GBDT, XGBoost, and LightGBM) were employed to construct radiomics signatures for assessing LVI status in BC patients. Results Eighteen radiomics features, 10 from DCE and 8 from T2WI, were retained to construct the radiomics signature. Among all the machine learning algorithms, the RF classifier model demonstrated superior performance in assessing the LVI status of BC patients, with an accuracy, sensitivity, and specificity of 63.32%, 74.47%, and 43.68%, respectively. The decision curve demonstrated significant clinical benefit of this model. Conclusion The radiomics-based RF model derived from MRI serves as a reliable indicator for identifying LVI status in BC, and holds great clinical utility for prompt intervention in invasive BC to improve the survival rate of BC patients.