Machine learning models based on magnetic resonance imaging for predicting Lymphovascular Invasion in Invasive Breast Cancer

Objectives: Treatment strategies for invasive breast cancer require accurate lymphovascular invasion (LVI) predictions. This study aimed to investigate the feasibility and effectiveness of delta radiomics signature based on dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) and radiomics signature based on T2-weighted fat suppressed imaging(T2FS) for assessing LVI in invasive breast cancer. Materials and methods: A total of 166 patients with resectable invasive breast cancer who underwent preoperative DCE-MRI and T2FS from July 10, 2020 to December 31, 2023 were enrolled. Radiomics features were extracted from pre-contrast phase (RFpre), second post-contrast phase (RFpost), and Delta radiomics features (RFDelta) calculated as RFpost minus RFpre. Then four radiomics signatures (RST2, RSpre, RSpost, RSDelta) were further developed based on the Random Forest model for RFT2, RFpre, RFpost and RFDelta, respectively. The predictive performance of all signatures was evaluated by receiver operating characteristic (ROC) analysis, with accuracy and area under the curve (AUC) as the main quantitative metrics. Results: In the test set, RSDelta (10 features) achieved the highest accuracy of 0.717 and an AUC of 0.764; RSpost (8 features) had an accuracy of 0.565 and an AUC of 0.610; RSpre (7 features) and RST2 (6 features) both showed an accuracy of 0.565 with AUCs of 0.535 and 0.662, respectively. Statistical differences were observed in predictive performance between RSDelta and RSpre, RSpost (both p