One AI model reads lymph nodes across many cancers — including the hardest cases: isolated tumour cells

Finding metastasis in a lymph node decides staging and treatment, and it is slow, tiring work in which pathologists disagree most over the smallest deposits. A team has built a weakly supervised model (MambaMIL+HiLA-MIL) that sorts whole-slide lymph-node images into four classes — negative, isolated tumour cells, micrometastasis and macrometastasis — and, in ten-fold cross-validation across several centres, outperformed six competing approaches under four different image-feature extractors. Its advantage was largest exactly where humans struggle: isolated tumour cells and micrometastases. This is a research tool evaluated retrospectively on archived slides, not a diagnostic device in clinical use.