HER2 expression status reflects the heterogeneity of breast cancer and is closely associated with variations in the tumor microenvironment. Noninvasive imaging approaches capable of capturing this spatial heterogeneity may improve subtype stratification in HER2-negative breast cancer. The aim of this study was to develop a global tumor habitat model combining habitat signatures derived from multiparametric MRI (mpMRI) with intratumoral and peritumoral radiomic features, and to evaluate its feasibility for predicting subtypes of HER2-negative breast cancer. In this multicenter retrospective analysis, 432 patients diagnosed with breast cancer were divided into training (n = 259), validation (n = 112), and test (n = 61) cohorts. Each voxel within the annotated region of interest (ROI) from both dynamic contrast-enhanced (DCE) and T2-weighted image (T2WI) sequences was characterized using a set of localized features. Voxel-wise feature vectors were subsequently clustered via the K-means algorithm to partition tumor ROIs into morphologically distinct subregions. Peritumoral regions were generated by radial expansion of the original ROI by 3 and 5 mm. Independent machine learning models were developed for intratumoral radiomics, peritumoral (PeriXmm), habitat (Habitat, HabitatT2, HabitatDCE), and clinical signatures. A combined predictive model integrating the optimal peritumoral features, habitat-derived signatures, and clinical parameters was constructed. Compared with the intratumoral radiomics model, the habitat model demonstrated superior predictive performance across all cohorts, with area under the ROC curve (AUC) values of 0.890, 0.841, and 0.820 in the training, validation, and test cohorts, respectively, versus 0.839, 0.723, and 0.639 for the intratumoral model. The Peri3mm model provided a more reliable representation of the peritumoral microenvironment than the Peri5mm model across external cohorts (AUC: 0.749 vs. 0.735). The combined model achieved the highest predictive overall performance, with AUCs of 0.906, 0.899, and 0.824. The combined intratumoral-peritumoral habitat-based model demonstrated the most robust and generalizable performance in the accurate and noninvasive prediction of HER2-negative breast cancer subtypes across multicenter cohorts.
Pan et al. (Thu,) studied this question.