We hope this contributes to open & reproducible research in #ComputationalPathology.

Big thanks to Guillaume Balezo, Albert Pla Planas and Etienne Decencière.

Great collaboration between @sanofifr.bsky.social and @minesparis-psl.bsky.social.

#DigitalPathology #FoundationModels #ViT #HuggingFace

🧪 Try it yourself!
We've released an interactive demo on Hugging Face 🤖
👉 huggingface.co/spaces/Estab...
👨‍💻 Code: github.com/Sanofi-Publi...

💡 The core idea of MIPHEI-ViT was to use a ViT foundation encoder (H-optimus-0) for this dense prediction task, thus leveraging the power of pathology foundation models for cross-modality prediction.

🔬 H&E stained tissue slides are cheap, and available for huge retrospective cohorts. mIF (and in particular Orion) are much more informative on particular cell types, but not available for large-scale cohorts.
💡 MIPHEI-ViT bridges the gap, learning to predict mIF from H&E.

#PRAIRIE

@minesparis-psl.bsky.social
@institutcurie.bsky.social
@inserm.fr

👩‍💻 Please checkout our code:

github.com/15bonte/cell...

🔬 We’re also releasing a massive new dataset — over 600,000 annotated nucleus images — now freely available on the BioImage Archive.

www.ebi.ac.uk/biostudies/b...

🧠 We introduce Cell-Cycle Variational Auto-Encoders (CC-VAE) — a deep learning framework to robustly and consistently predict cell cycle phase from microscopy images.

If you are interested in spatial transcriptomics and want to get down to single-cell level, then this is for you! 3/3

@minesparis-psl.bsky.social
@institutcurie.bsky.social
@inserm.fr

#PRAIRIE

Huge thanks to Lucie Gaspard-Boulinc and Luca Gortana for this amazing work and Florence Cavalli and Emmanuel Barillot for this wonderful collaboration at the U1331 - Computational Oncology 2/3

🔗 Resources

- Preprint: doi.org/10.1101/2025...
- RNA2seg package: github.com/fish-quant/r...
- Annotated datasets: zenodo.org/records/1491...

Looking forward to feedback from the community!

Accurate cell segmentation is critical in spatial transcriptomics but often challenged by poor staining and complex tissues. RNA2seg addresses this by integrating all available data types—membrane, nuclear staining, and RNA positions.

RNA2Seg is a deep learning model trained on over 4 million cells across 7 organs, integrating RNA point clouds and multiple stainings for robust and accurate cell segmentation in image based spatial transcriptomics.