Abstract
PURPOSE: This study aims to predict human epidermal growth factor receptor-2 (HER-2) expression in breast cancer based on radiomics of magnetic resonance imaging (MRI) habitat and ultrasound (US). PATIENTS AND METHODS: This retrospective study included 182 breast cancer patients confirmed by pathology from May 25, 2019 to April 15, 2025. The data set was randomly divided into a training set (n=145) and a testing set (n=37) with an 8:2 ratio. All patients underwent MRI and US before surgery. Volumes of interest were delineated on the second phase of dynamic contrast-enhanced T1-weighted imaging, which were clustered into different habitat regions via K-means clustering. Feature selection was using Spearman correlation, greedy recursive elimination strategy, least absolute shrinkage and selection operator regression. Models based on extremely randomized trees were developed using radiomics features extracted from MRI habitats, or from regions of interest on US. A clinical model was developed based on baseline data, followed by stacking the best habitat model and US model, as well as a combination of the best habitat, US, and clinical models. Model performance was evaluated by areas under the curve (AUCs) and integrated discrimination improvement (IDI). The interpretability of the best habitat model and US model was using Shapley Additive exPlanations analysis. RESULTS: Model_H1_(multi-parametric) was selected as the best habitat model (AUC was 0.880 and 0.801 in the training set and testing set). Model_(H1+US+Cli) (AUC was 0.945 and 0.835 in the training set and testing set) outperformed Model_H1_(multi-parametric), the US model and the clinical model. The IDI analysis demonstrated further improvement by Model_(H1+US+Cli). CONCLUSION: A combined model based on multi-parametric MRI habitat radiomics, US imaging radiomics, and clinical features can effectively predict HER-2 expression status in breast cancer.