Stain-Invariant Invasive-Region Detection from a Pathology Foundation Model: Cross-Cohort and Cross-Stain Transfer without Retraining

Delineating invasive tumour regions is a prerequisite for nearly every quantitative immunohistochemistry (IHC) biomarker, including Ki-67, ER/PR, and HER2, yet existing region-detection models are typically trained separately for each stain and cohort and rarely transfer across domains. We investigate whether a single invasive-region detector can be trained once and reused across stains and institutions by combining a frozen pathology foundation model with a stain-invariant image representation. Both H&E and IHC images are first transformed into a shared hematoxylin-only space using stain-independent colour deconvolution, and we demonstrate that the nuclear signal remains physically preserved even under heavy DAB chromogen deposition (approximately threefold above background in the highest-burden decile), providing a common substrate for analysis. Frozen Virchow2 embeddings are then extracted, and only a linear classification head is trained, without any fine-tuning of the foundation model. Trained on the TIGER breast H&E dataset, the model achieves an AUC of 0.95 and Dice score of 0.90 under five-fold cross-case validation, and transfers zero-shot to the independent BACH cohort, collected in a different country using different scanners, with an AUC of 0.80 while maintaining high precision and specificity, with the invasive-versus-in-situ boundary remaining the primary source of error. Across stains, an unsupervised, label-free feature alignment improves invasive-region recall on IHC from near zero to approximately 50%, and with only 12 exhaustively annotated Ki-67 slides, the same linear head achieves a Dice score of 0.89 and precision of 0.96 under leave-one-slide-out evaluation, yielding a strong whole-slide invasive-region detector (pending prospective validation). These results demonstrate that stain-invariant rendering combined with a frozen foundation model enables a single reusable, label-efficient invasive-region detector that transfers across cohorts and stains within the evaluated setting, without requiring stain-specific or cohort-specific model retraining.