Cross-cancer knowledge transfer for lymph node metastasis prediction: Applicability of an axillary lymph node-based deep learning model
Evaluation of lymph node (LN) metastasis is critical for cancer staging and treatment planning. Since radiologic hallmarks of metastatic LNs are shared across primary cancers, a computed tomography (CT)-based deep learning (DL) classifier developed from one anatomy may transfer to others. This study built a DenseNet-121 ensemble model using clinical-implication-applied preprocessing and evaluated its cross-anatomical generalizability to LNs from diverse primary cancers. A DenseNet-121 ensemble model combining fixed-size cropping (FSC) and adjustable square cropping (ASC) was trained from ImageNet initialization on an axillary lymph node (ALN) dataset (n = 1,128 from 523 breast cancer patients). The model utilized clinical-implication-applied perinodal cropping with bicubic filtering and standardized resizing. To test cross-anatomical generalizability, an external dataset of 256 CT images across four anatomical regions was evaluated: cervical, mediastinal, upper-abdominal, and lower-abdominal LNs. Perinodal free margins were systematically varied from 10% to 40% to optimize the balance between context preservation and surround interference. Bicubic filtering and standardized resizing yielded the highest internal performance on the original ALN dataset (FSC-ASC ensemble area under the receiver operating characteristic curve [AUROC] 0.958, 95% confidence interval [CI] 0.920-0.987 at 40% margin). The model demonstrated site-dependent, above-chance generalizability across all external regions. Cervical LNs achieved the highest performance at a 40% margin (AUROC 0.939, 95% CI 0.875-0.989). Mediastinal LNs peaked at a 10% margin (AUROC 0.796, 95% CI 0.672-0.903), with wider margins degrading performance due to adjacent structural interference. Upper-abdominal and lower-abdominal LNs reached optimal performance at a 30% margin (AUROC 0.814 for upper; AUROC 0.917 for lower). A single ALN-trained DL model can accurately predict LN metastasis across cervical, mediastinal, and abdominal stations through site-specific perinodal margin tuning. These findings support the feasibility of cross-cancer knowledge transfer and the clinical utility of generalizable backbones for pan-cancer nodal assessment.