Boundary-Guided Mammographic Radiomics for Molecular Subtype Risk Stratification in Breast Cancer

This work proposes a multi-stage interpretable framework for mammography-based molecular subtype-oriented risk stratification in breast cancer, focusing on visible mass lesions with segmentable tumor boundaries. The framework integrates a Boundary-Enhanced module within an attention-based U-Net architecture to improve tumor boundary delineation. Morphological and radiomics descriptors capturing shape, margin-related geometry, intensity, and texture heterogeneity are then extracted from the segmented tumor region and used for machine-learning-based subtype-oriented classification. For this task, the best-performing CatBoost model achieved an ROC AUC of 94.6% on Dataset A and 93.7% on Dataset B, used as an external test cohort for downstream subtype classification. These findings suggest that boundary-guided mammographic radiomics may provide a non-invasive adjunct for pre-biopsy risk stratification and subtype-oriented clinical assessment. However, tissue-based pathological evaluation remains the reference standard for definitive molecular subtype determination.