Multi-task Model with​ Peritumoral MRI Radiomics Learning for Breast Cancer Segmentation-Guided Subtype Classification

Accurate breast cancer molecular subtype classification is important for treatment planning, but immunohistochemical assessment may be affected by intratumoral heterogeneity and requires invasive tissue sampling. We hypothesize that segmentation-guided learning of tumor-core and peritumoral microenvironment (PTM) MRI features would improve molecular subtype discrimination over tumor-core-only analysis. A transformer-based segmentation network integrates cross-modal and multi-scale features to delineate tumors, after which the predicted masks are used to construct tumor-core and peritumoral inputs for an energy-regularized classification network. On the test cohort, the retrospective single-center study includes 381 patients with histopathologically confirmed breast cancer and corresponding DCE, DWI, T1WI, and T2WI images. The proposed method achieves a Dice coefficient of 88.29%, a Jaccard index of 79.80%, a precision of 91.98%, and a 95HD of 4.03 pixels for tumor segmentation. For molecular subtype classification, it achieves an accuracy of 84.0%, a macro-averaged sensitivity of 85.7%, a macro-averaged specificity of 94.8%, and a macro-averaged F1 score of 83.9%, providing the best overall performance among the evaluated methods. These findings support further evaluation of segmentation-guided multimodal MRI analysis for preoperative molecular subtype assessment.