Deep learning in breast cancer histopathology: predicting and detecting axillary lymph node metastasis

Deep learning (DL) methods are increasingly applied to digitized hematoxylin and eosin (H&E) slides in breast cancer to assist in staging and prognosis. One emerging application is the assessment of axillary lymph node (ALN) status, either by detecting metastases on lymph node slides or by predicting nodal involvement from primary tumor histology. This narrative review summarizes recent progress in these two complementary areas. Literature was identified through a focused, non-systematic search of PubMed, Scopus, and Web of Science using terms related to breast cancer, lymph node metastasis, and deep learning on H&E slides. Prediction studies showed variable discrimination. In independent or external evaluations, reported AUCs ranged from approximately 0.50 to 0.83, with multimodal models generally outperforming image-only models. However, performance fell across institutions and favorable luminal cohorts. Endpoint definitions were inconsistent, and several studies did not state whether isolated tumor cells (ITCs), micrometastases, or only macrometastases constituted a positive node. In contrast, DL models for detecting metastases on lymph node slides achieved near-perfect sensitivity and are now entering clinical use, reducing immunohistochemistry workload and review time without compromising diagnostic accuracy. Overall, deep learning on histopathology slides holds promise for improving axillary staging efficiency and precision, with metastasis detection already clinically viable and predictive modeling from primary tumor morphology representing the next frontier for research and validation.