Accurate detection of lymph node micrometastases is essential for thyroid cancer management but remains challenging due to their tiny size and limited annotated data. We propose MTFuse, a meta‐learning‐based dual‐scale transformer framework that integrates paired 10× and 4× patches to simultaneously capture cellular details and tissue context. Evaluated on the thyroid cancer micro‐metastasis dataset dataset, MTFuse achieved 98. 46% AUC, 93. 63% weighted recall, and 93. 77% weighted precision for micrometastasis detection, outperforming all state‐of‐the‐art convolutional neural network and transformer baselines. It also demonstrated strong robustness on macrometastases (AUC 99. 51%) and generalized effectively in 0‐shot settings to isolated tumor cells (AUC 93. 36%) and isolated psammoma bodies (AUC 91. 06%). These results show that MTFuse provides a highly accurate and interpretable solution for small‐lesion detection, offering strong potential for clinical deployment in data‐limited digital pathology workflows.
Liu et al. (2026) studied this question.