A universal visual foundation model for computational cytopathology.

CROWN achieves over 95% accuracy in 48 cytology classification tasks, establishing a universal visual backbone for computational cytopathology.

• Why it matters: Cytopathology is vital for cancer diagnosis but is labor-intensive, highlighting the need for scalable, automated computational tools to improve efficiency and consistency.
• What they did: The team pretrained CROWN on more than 10 million cytology image patches using a self-supervised DINOv2-based approach, avoiding manual annotations, and tested it across 202 diverse tasks and datasets.
• The result: CROWN outperformed existing pretrained encoders, with accuracy exceeding 95% in many tasks, enabling broad application of automated cytopathology analysis across various datasets and diagnostic tasks.