Lymph node whole-slide images provide incremental value beyond primary tumors for AI-based pCR prediction in node-positive breast cancer

To investigate whether lymph node histopathology provides incremental value for predicting pathological complete response (pCR) beyond the primary tumor in node-positive breast cancer.

In this retrospective study of 1812 node-positive breast cancer patients receiving neoadjuvant therapy, we collected paired whole-slide images from primary tumor and lymph node biopsies. We developed a dual-lesion AI framework (DP-Net) integrating predictions from separate primary-tumor (P-Net) and lymph-node (N-Net) models via decision-level fusion, and evaluated its incremental predictive value over P-Net alone.

P-Net predicted pCR with moderate accuracy (AUROC 0.793), while N-Net showed similar performance (AUROC 0.781). DP-Net achieved an AUROC of 0.814, significantly outperforming P-Net in internal (P = 0.003) and preliminary external validation (P = 0.016). Multivariable regression confirmed that lymph node information provided independent predictive value beyond P-Net in both cohorts (internal: χ² = 34.749, P < 0.001; external: χ² = 6.428, P = 0.011). An IHC-Age model (age, ER, PR, HER2, Ki-67) achieved an AUROC of 0.855; adding DP-Net improved it to 0.869, significantly outperforming the IHC-Age model alone (P = 0.014) and the model combining IHC-Age features with P-Net (P = 0.034). Attention heatmaps and transcriptomic analysis revealed distinct profiles between primary tumors and lymph node metastases.

Lymph node histopathology provides incremental pCR-predictive value beyond the primary tumor model, a benefit that persists when the IHC-Age model is included. Routinely acquired lymph node biopsies may serve as a cost-effective predictive resource for guiding neoadjuvant therapy decisions.