ProtoDS-MHIM: A Prototype-Enhanced DSMIL Framework with Teacher-Guided MHIM for Metastatic Breast Cancer Detection in Lymph-Node Whole-Slide Images

We propose ProtoDS-MHIM, a prototype-enhanced dual-stream MIL framework that combines teacher-guided hard-instance mining with class-conditional prototype learning and learnable branch fusion for weakly supervised WSI classification. The framework achieves 96.30% accuracy and 99.72% ROC-AUC on the held-out CAMELYON16 validation subset, while prototype analysis provides insight into the learned class-conditional representations.