CytoFM

CytoFM is a self-supervised foundation model for digital cytology — the microscopic examination of individual cells collected by fine-needle aspiration, Pap smear, or fluid sampling, as opposed to the intact tissue sections histopathology studies. Cytology is a mainstay of cancer screening precisely because collection is fast, cheap, and minimally invasive, but building reliable deep learning models for it has been hard: staining and preparation protocols vary widely between laboratories, cell appearance differs across organs, and public annotated datasets remain small and fragmented. CytoFM, developed by UCLA with a collaborator at UT Southwestern Medical Center and presented at the CVPR 2025 Workshop on Computer Vision for Microscopy Image Analysis, is the first foundation model pretrained specifically on cytology imagery.

The gap it addresses is concrete. Histopathology foundation models such as UNI learn from tissue tiles, where diagnostic signal lives in glandular and stromal architecture. Cytology slides instead hold dispersed cells and small clusters scattered across preparation debris, and what a cytopathologist reads is nuclear morphology, chromatin texture, and mitotic activity. Encoders pretrained on histology transfer poorly to that regime, and the usual alternative — fine-tuning a general-purpose encoder separately for each dataset — must be repeated for every new task and institution.

CytoFM takes the self-supervised route instead, learning from unlabeled cytology patches so that one frozen backbone serves many downstream tasks. The authors are explicit that the pretraining corpus is small by foundation-model standards and that the model is cervix-weighted; they position CytoFM as evidence that domain-specific pretraining pays off in cytology, not as a finished clinical-grade encoder.

#Key Features

• Cytology-specific self-supervision: Pretrained with iBOT, which combines masked image modeling with teacher-student self-distillation, so the backbone captures both fine-grained detail such as mitotic figures and higher-level properties such as cellularity.
• Frozen-feature transfer: The trained Vision Transformer is frozen and used purely as a patch encoder; downstream tasks add only an attention-based multiple instance learning (ABMIL) head plus a linear layer, with no backbone retraining per dataset.
• Multi-organ, multi-institution corpus: Pretraining spans breast, cervical, and thyroid cytology drawn from seven public datasets and one private hospital cohort, covering several scanner and camera types.
• Higher-magnification tiling: Patches are extracted at 40x rather than the 20x convention of histopathology encoders, because cytology reads individual cells and small clusters rather than tissue architecture. No stain normalization is applied; the self-distillation objective is left to absorb staining and preparation variation.
• Generalization to unseen data: On HiCervix, a large public dataset excluded entirely from pretraining, CytoFM outperforms both comparison encoders.

#Technical Details

CytoFM is a ViT-Base backbone initialized from ImageNet-1K iBOT weights and pretrained on 1,393,282 non-overlapping 256x256 patches at 40x magnification. The corpus combines seven public cytology datasets — FNAC2019 for breast, and MLBC, SiPaKMeD, BMT, APACS23, CCEDD, and the Bialystok collection for cervix — with a private UCLA thyroid cohort of 496 Pap-stained whole-slide images from 160 patients. Because those slides are gigapixel WSIs rather than curated cell images, a fine-tuned VGG-16 classifier scores each patch for diagnostic content and only the 1,500 highest-scoring patches per slide are kept.

Evaluation uses ABMIL over frozen patch features on three classification tasks, compared against iBOT-ImageNet and UNI at their released weights. On four-class cervical cell typing in MLBC, CytoFM reaches 0.930 accuracy and 0.993 AUROC against 0.895/0.986 for UNI and 0.879/0.983 for iBOT-ImageNet. On the fully unseen HiCervix dataset it reaches 0.844 accuracy and 0.968 AUROC, ahead of iBOT-ImageNet (0.803/0.956) and UNI (0.800/0.952), though below the 0.921 accuracy of a HierSwin model trained directly on HiCervix. On breast malignancy classification in FNAC2019 it trails both baselines at 0.908 accuracy, a shortfall the authors attribute to breast contributing only 2,544 images against roughly a million cervical ones.

#Applications

CytoFM is intended as a general-purpose patch encoder for cytopathology pipelines: cervical cancer screening triage, thyroid nodule assessment from fine-needle aspirates, breast FNAC malignancy classification, and slide- or patient-level prediction under weak labels. Because the backbone stays frozen, groups with modest labeled cohorts can train a lightweight aggregation head rather than a full encoder — which matters in a specialty where expert annotation is scarce.

#Impact

CytoFM establishes cytology as a domain that warrants its own pretrained encoder rather than borrowing one from histopathology, and its attention maps and UMAP embeddings show the model attending to nuclei, nuclear boundaries, and mitotic activity. The evidence is deliberately bounded: three downstream tasks, a pretraining corpus two orders of magnitude smaller than the histopathology encoder it is compared against, and a skew toward cervical material that costs it performance on breast. Weights are distributed through the first author's GitHub repository with a loading notebook, but the repository carries no license file and the private thyroid cohort is not released, so reuse terms and full reproduction stay constrained.