Utility of vision-language pre-training for histopathology had been limited by small datasets and insufficient grounding in diagnostic reasoning. We scaled PRISM to 700k specimen-report pairs and expanded supervision signal to clinical-dialogue, aligning histomorphologic features with diagnostic language. The method matched and even exceeded cancer-detection performance of clinical-grade products without additional training.

Clinical pathology reports are at the specimen level, yet foundation models are limited to small patches within specimens. We developed PRISM, a vision-language model which aggregates representations across 100k patches to match detailed clinical diagnosis extracted from the reports, thus creating a denser supervision signal. The method enables 1) zero-shot cancer detection via text prompts and 2) data-efficient transfer of specimen-level representations to biomarker prediction, which outperforms supervised baselines with as little as 10% of the training data.

We investigate the most effective way to scale foundation models for histopathology. Prior work scaled along individual axes of data and model size, however an exploration of domain-specific changes to the algorithm was lacking. We introduced new data augmentations and a regularisation objective tuned to pathology data, and demonstrate their advantages over the vanilla DINOv2 method in a large ablation study. We then successfully scaled Virchow to 3.1M whole slide images and 1.9B parameters, achieving state-of-the-art on 12 tile-level tasks.

Computational pathology models must generalise across diverse tissue types and rare cancers, which requires large-scale diverse datasets. However, prior foundation models had not been scaled to the data volumes needed for clinical-grade performance on rare tissue types. We trained Virchow, a ViT-H DINOv2 model, on 1.5M whole slide images, the first histopathology foundation model at this scale, enabling clinical-grade tile-level feature extraction with strong generalisation to rare tissues and cancers.