A universal visual foundation model for computational cytopathology

Cytopathology is central to cancer screening and diagnosis but remains labor-intensive, motivating the development of scalable computational solutions. Here, we present CROWN (Cytology visual foundation netwoRk Optimized With self-supervised learNing), a universal visual foundation model for computational cytopathology. CROWN was pretrained on more than 10 million cytology image patches using a DINOv2-based self-supervised framework, without requiring manual annotations during pretraining. We evaluated CROWN in 202 task-setting combinations spanning classification, segmentation, detection, retrieval and slide-level prediction across private institutional cohorts and public datasets. CROWN achieved the best overall performance among established pretrained encoders across diverse benchmark settings, with accuracy exceeding 95% in 48 patch-level classification evaluations and 98% in 22 of them. These findings support CROWN as a scalable visual backbone for general-purpose computational cytopathology across heterogeneous datasets and task settings.