Cytopathology vision foundation model, a ViT-Large pretrained with DINOv2 self-supervision on over 10 million unannotated cytology image patches.
No providers recorded yet. Browse all providers
A cytology specimen is not a tissue section. A Pap smear, a thyroid fine-needle aspirate or a pleural effusion spreads individual cells and small clusters across a slide, stripped of the architecture that histopathology relies on, so the diagnosis rests almost entirely on nuclear size, chromatin texture and the shape of single cells. Encoders pretrained on histology tiles learn glands, stroma and tissue organization, features that are largely absent from a cytology slide, while task-specific cytology classifiers each need their own labelled dataset for every organ and every question.
CROWN (Cytology visual foundation netwoRk Optimized With self-supervised learNing) answers this with a single vision encoder pretrained only on cytology. It was developed by the Department of Pathology and the Artificial Intelligence Laboratory at Sun Yat-sen University Cancer Center, with collaborators at Sun Yat-sen Memorial Hospital and the School of Computer Science and Engineering of Sun Yat-sen University, the First Affiliated Hospital of Guangzhou Medical University and Zhongshan People's Hospital. It was published as a Technical Report in Nature Cancer in September 2026.
The model is pretrained with DINOv2 self-distillation, in which a student network learns to match a slowly updated teacher's representation of different crops and augmentations of the same image. No diagnostic labels are involved, so the encoder learns what distinguishes one cell from another from more than 10 million unannotated patches, and the same frozen features are then reused across organs and tasks. It follows CytoFM, an earlier cytology foundation model from a different group, and sits alongside hematology-specific DINOv2 encoders such as DinoBloom and general histopathology encoders such as UNI, CONCH and Virchow.
CROWN is a DINOv2 ViT-Large with a 16-pixel patch size, 224-pixel input, LayerScale and a SwiGLU feed-forward layer; the released CROWN.pth checkpoint (about 304 million parameters) loads into the DINOv2 vit_large definition shipped in the repository, and the class-token output serves as the image embedding. The pretraining images are patient-derived cytology collected at the authors' institutions and are not public. Evaluation covers 202 task-setting combinations drawn from two institutional cohorts, one from Sun Yat-sen Memorial Hospital (20 binary and 14 multiclass patch tasks) and one from Sun Yat-sen University Cancer Center (17 binary and 5 multiclass), plus public cervical, effusion, thyroid and blood datasets.
The protocols are zero-shot patch recognition, linear probing with five-fold cross-validation, image retrieval scored by majority-vote accuracy over the top five neighbours, weakly supervised multiple-instance learning on whole slides, and few-shot slide classification across a range of shot counts. Dense tasks use the public Thyroid Seg. and Cervix Seg. datasets for cell segmentation and two blood-smear datasets for detection. Across these settings CROWN achieved the best overall performance among the established pretrained encoders it was compared with, and exceeded 95% accuracy in 48 patch-level classification evaluations, 22 of them above 98%.
CROWN is aimed at computational cytopathology groups that need a feature extractor for screening and diagnostic workflows: triaging cervical Pap smears, grading thyroid aspirates, finding malignant cells in serous effusions and classifying blood cells. Because the backbone stays frozen, a lab with a new cohort can train a linear probe or a slide-level aggregator on top of precomputed embeddings, and the retrieval setting supports case-based lookup, where a cytopathologist queries an archive for morphologically similar cells. The code is public on GitHub but ships without a license file, and the weights on Hugging Face are tagged Apache 2.0 behind a contact-sharing access request, so the terms of reuse are not the same for both. The model has not been validated as a clinical device.
Foundation models for pathology have concentrated on histology whole-slide images, and cytology, which underpins much of population cancer screening, has had far fewer shared encoders. CROWN adds a large, cytology-only backbone with released weights and a public benchmark for comparing new cytology encoders. The group describes it as part of a cytopathology foundation-model series alongside COIN, their earlier generative model for cytology. The evaluation is retrospective and in silico, drawing on two Guangzhou institutions and public datasets, so prospective and multi-national validation is the natural next step before the model supports screening decisions.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.