Cross-cancer knowledge transfer for lymph node metastasis prediction: Applicability of an axillary lymph node-based deep learning model

A DenseNet-121 ensemble trained on axillary lymph nodes (n = 1,128) predicts CT-based lymph node metastasis across cervical, mediastinal, and abdominal stations with site-specific margin tuning, peaking at AUROC 0.939 for cervical nodes.

• Why it matters: Radiologic metastatic patterns are shared across primary cancers, but models trained on one lymph node anatomy may not generalize without accounting for surrounding structures.
• What they did: The study trained a DenseNet-121 ensemble (FSC-ASC) with clinical-implication-applied perinodal cropping and bicubic filtering on 1,128 axillary node CT images from 523 breast cancer patients, then tested cross-anatomical performance on 256 external CT images from four regions.
• The result: Generalizability was above chance in all external regions, with optimal perinodal margins of 40% for cervical (AUROC 0.939, 95% CI 0.875–0.989), 10% for mediastinal (AUROC 0.796, 95% CI 0.672–0.903), and 30% for upper- and lower-abdominal nodes (AUROC 0.814 and 0.917, respectively).

AI summary of the abstract · may contain errors · not medical advice · read the paper.