Weakly supervised artificial intelligence for multi-cancer detection of lymph node metastasis on whole slide images

Abstract The accurate identification of lymph node metastasis is critical for cancer diagnosis/treatment but remains time-consuming and error-prone for pathologists. We develop MambaMIL+HiLA-MIL, a multiple instance learning model combining Vision Mamba with high-low attention separation mechanism. This model is compared against six baselines under four feature extractors (ResNet, UNI, Virchow, and GigaPath). Ten-fold cross-validation is employed for model evaluation. Our model significantly outperforms all baselines. It exhibits strong performance in detection of isolated tumor cells and micro-metastasis, while maintaining high performance for negative and macro-metastasis cases. The model still demonstrates good stability in detecting lymph node metastasis across multi-cancer and various individual cancer types. UNI and GigaPath yield significantly better performance than ResNet and Virchow. Here, we show that MambaMIL+HiLA-MIL is a multi-cancer four-class lymph node metastasis model, demonstrating high efficiency and robustness across multiple centers and various feature extractors, offering a reliable tool for clinical lymph node metastasis classification.