Most pathology foundation models learn by encoding thousands of image patches from a single representative slide, then aggregating those features. That approach limits how well a model can integrate the multiscale and multi-slide context that pathologists actually use when making a diagnosis. The WILSON model, described in a recent arXiv preprint, instead represents whole-slide images and multi-slide cases as multi-magnification composites, allowing it to reason at the patient level.

WILSON was trained on roughly 189,000 slides covering 42 organs and 838 diagnostic entities. In zero-shot evaluation, it achieved a macro-F1 of 0.52 on a case-level classification task, compared with 0.38 for a dedicated case-level model, and it matched substantially larger slide-level models while requiring far fewer computational operations. End-to-end fine-tuning further improved performance on breast cancer subtyping and stromal lymphocyte grading.

The model also showed strengths in multimodal retrieval and text generation. It surpassed PRISM on image-text retrieval, with recall@1 of 75.6% versus 58.1%, and its generated diagnostic descriptions received higher ratings from five board-certified pathologists than those from competing models. The authors argue that compact, multiscale representations can support clinically grounded and computationally efficient pathology foundation models.