scdino-v2-base

Self-supervised Vision Transformer (DINOv2) for 5-channel single-cell microscopy crops of human peripheral blood mononuclear cells (PBMCs). It returns one 128-dimensional embedding per crop.

• Input: 50×50 px crops with 5 channels (H, W, C), in this order:

  #	Channel	Markers
  1	647 nm	CD3 (APC), CD14 (Alexa Fluor 647)
  2	Brightfield	–
  3	DAPI	–
  4	488 nm	CD4, CD19 (FITC)
  5	594 nm	CD8, CD16, CD56 (PE), CD20 (Alexa Fluor 594)

• Data: A labelled set of more than 1.3 million single-cell images of seven immune cell types is openly available: Morphologically annotated single-cell images of human PBMCs.

Usage

See the README of the code repository, https://github.com/JonasMeirer/scDINOv2, for how to load the model, preprocess crops and evaluate embeddings.

Usage on images from other microscopes, stains, cells or crop sizes might need new normalization statistics and retraining.

Not a diagnostic tool, and not for clinical use.

Citation

The publication describing this model is currently under review. This section will be updated with the reference once it is published.

License

MIT, like the scDINO code. Training used timm (Apache-2.0) and code vendored from Lightly (MIT).