Ultrasound-guided deep learning framework with whole slide images for molecular subtyping of breast cancer: a multicentre, retrospective study

Accurate molecular subtyping of breast cancer is essential for personalized therapy, yet inconsistencies between biopsy and postoperative immunohistochemistry (IHC) due to tumor heterogeneity and sampling limitations pose clinical challenges. We developed Ultrasound-Guided Pathomic Subtyping (US-GPS), a cross-modal framework enabling subtype prediction using ultrasound (US) alone, guided by IHC-stained whole-slide images (WSIs) during training. A total of 1283 patients from four centers were retrospectively included, with 454 cases (paired US and WSIs) for model training/validation and 829 US-only cases for external testing. US-GPS aligned WSI-derived patch- and region-level features (from HER2, ER, PR, Ki-67 IHC stains) with handcrafted and deep US features via co-attention and contrastive learning. Using US alone, the model achieved an AUROC of 0. 966 (95% CI 0.960–0.991) in internal validation, comparable to pathology-only (0. 971) and joint models (0.980), and superior to US-only (0.861). Performance remained robust across both internal (AUROC range: 0.930–0.942) and all external cohorts (AUROC range: 0.915–0.963). US-GPS enables scalable, non-invasive, and interpretable subtype classification, with potential for both pre-treatment decision-making and longitudinal monitoring of molecular phenotypic shifts.