Stratified prediction of HER2 status in breast cancer by integrating intratumoral and peritumoral radiomics from DCE-MRI

This is the authors' abstract. We don't add key points for this paper.

This study aimed to investigate the value of Dynamic Contrast-Enhanced Magnetic Resonance Imaging (DCE-MRI) radiomics in predicting HER2 expression status in breast cancer and to develop stratified prediction models. We enrolled 210 breast cancer patients, categorized into HER2-negative and HER2-positive groups, as well as HER2-low and HER2-zero expression groups based on their HER2 status. Radiomic features were extracted from the intratumoral region and various peritumoral regions (1 mm, 3 mm, 5 mm) on DCE-MRI. Predictive models were constructed for distinguishing HER2-negative from HER2-positive cases and for differentiating HER2-low from HER2-zero expression. Feature selection was performed using LASSO regression, followed by classification via logistic regression. The model based on the intratumoral region alone demonstrated robust performance in classifying HER2-negative and HER2-positive status. For distinguishing HER2-low from HER2-zero expression, the combined model incorporating both intratumoral and 3 mm peritumoral features showed superior performance. We introduced the Generalization Decay Index (GDI) as a novel metric for evaluating model generalizability. Analysis using GDI revealed that relying solely on the AUC could be misleading for stability assessment. The model based on the 5 mm peritumoral region exhibited a high GDI, suggesting potential overfitting, whereas the intratumoral model achieved the lowest GDI, indicating the highest stability. The combined intratumoral and 3 mm peritumoral model not only showed good diagnostic efficacy but was also validated by GDI as the most stable and generalizable model among all configurations for the HER2-low vs. HER2-zero classification task. DCE-MRI-based radiomics can effectively predict HER2 expression status and facilitate the construction of stratified prediction models in breast cancer. The peritumoral region-based model demonstrates stability in classifying HER2-negative and HER2-positive status, while the combined intratumoral and 3 mm peritumoral model offers advantages for distinguishing HER2-low from HER2-zero expression. The proposed GDI serves as a valuable new indicator for assessing model generalizability, providing a novel approach for the non-invasive evaluation of HER2 status and offering fresh insights into model performance evaluation.