Toward a Breast-Cancer-Specific Pathology Foundation Model: Preliminary Analysis

Digital pathology foundation models trained with self-supervised learning are increasingly used as general-purpose feature extractors. Most current models are trained across many tissue and cancer types at once, which may dilute tissue-specific signals. This work-in-progress report describes an early breast-cancer-specific pathology foundation model: a 1.1-billion-parameter, DINOv3-based Vision Transformer trained on a balanced, curated image corpus from a large breast cancer registry. A key methodological choice is adapting explicitly at native 20× field of view during a high-resolution adaptation phase intended to better preserve cellular- and nuclear-level detail. We evaluate the resulting representations qualitatively via patch-level cosine-similarity and principal component analysis visualizations that reveal coherent tissue structure and an emergent tissue-vs-background segmentation with no dedicated supervision. We also evaluate quantitative metrics against an open pathology foundation model, including cell classification and rotation-equivariance. Benchmarked against H-optimus-0, a size-matched public general-purpose pathology foundation model, our final model shows a modest but consistent advantage on cell classification and lower rotation drift. This paper is an early look at foundation-model training for a breast-cancer-specific cohort — further benchmarks, downstream task-specific modeling, and scaling to a ~7-billion-parameter model are in progress.