HCF-Net: a Hierarchical Cascade Fusion Network Aligned with Clinical Treatment Pathways for Preoperative Molecular Subtyping of Breast Cancer on Dynamic Contrast-Enhanced MRI.

Preoperative molecular subtyping of breast cancer is essential for treatment selection and prognosis assessment, yet most deep learning approaches formulate subtype prediction as a conventional flat multi-class task without reflecting the sequential clinical decision process. We developed HCF-Net, a hierarchical cascade fusion network that decomposes subtype prediction into three clinically aligned binary decision nodes and integrates multiphase DCE-MRI with clinicoradiologic information. The model was retrospectively developed and validated using a dual-center cohort of 373 women who underwent preoperative 3.0 T multiphase DCE-MRI. On an independent test set (n = 75), HCF-Net achieved a Macro-AUC of 0.868 (95% CI, 0.806-0.923), accuracy of 73.3%, Macro-F1 of 0.732, and Cohen's κ of 0.631. The fusion model consistently achieved the highest AUC across all three hierarchical decision nodes, reaching 0.890 for luminal versus non-luminal tumors, 0.865 for Luminal A versus Luminal B tumors, and 0.974 for HER2-enriched versus triple-negative tumors. The same ordering of classification strategies was observed with a ConvNeXt-Tiny backbone. Compared with conventional flat classification strategies, hierarchical cascade decomposition and decision-level fusion provided substantial performance gains. These findings suggest that a hierarchical fusion framework aligned with clinical decision pathways may provide an effective and interpretable approach for preoperative breast cancer molecular subtype prediction.