📢 Apply now for the first DL@Janelia Bootcamp (June 4–18, 2026)! 2 weeks of hands-on deep learning for microscopy: train models on your own data, Python needed (no ML experience), housing + meals included, no registration fee

📍 Janelia Research Campus, VA
🗓️ Apply by Jan 15, 2026
🔗 shorturl.at/j4sgY

Don’t miss Elena’s (from @ilastik-team.bsky.social lab) brilliant work @ #ICCV2025 with @anwaiarchit.bsky.social & @cppape.bsky.social poster 292 @ 11:15AM 🔬They tackle segmentation of massive 3D microscopy images & show how BatchRenorm removes tiling artifacts boosting transferability and clarity🌺

Attending BISCUIT workshop at #ICCV2025 🌺 — the first talk by @mariavak.bsky.social on quantifying biases in foundation models was fantastic! Exciting to see interdisciplinary efforts focused not just on prediction but on understanding why — and ViTs seem to be the trend in biomedical vision 👩‍💻

We’ve upgraded ShapeEmbed 🎉 ShapeEmbedLite decodes latent codes via an MLP to guarantee valid EDMs, making it lighter and ideal for small microscopy datasets or limited compute. Hear more at my BIC workshop talk or poster at #ICCV2025! Try it out at github.com/uhlmanngroup...

Why is ShapeEmbed useful in your research? It can extract shape descriptors to distinguish cell phenotypes and reveal drug-condition induced changes through clustering and latent-space exploration, offering a powerful tool for biological shape analysis. (5/N)

On top of all of this, the latent space learned by ShapeEmbed is generative: it can be used to generate shapes and compute meaningful shape averages and trajectories, enabling data exploration & visualization—which are useful for both discovery and hypothesis testing. (4/N)

ShapeEmbed encodes Euclidean distance matrices relying on a novel encoder design and a new indexation-invariant loss. Through this, it learns highly-informative shape features that surpass classical approaches and current SotA learning methods. (2/N)

Happy to share that ShapeEmbed has been accepted at @neuripsconf.bsky.social 🎉 SE is self-supervised framework to encode 2D contours from microscopy & natural images into a latent representation invariant to translation, scaling, rotation, reflection & point indexing
📄 arxiv.org/pdf/2507.01009 (1/N)

ShapeEmbed encodes Euclidean distance matrices relying on a novel encoder design and a new indexation-invariant loss. Through this, it learns highly-informative shape features that surpass classical approaches and current SotA learning methods.

Today at the EMBL-EBI seminars @ebi.embl.org we hosted @alex-krull.bsky.social who talked about self-supervised learning and generative AI for image restoration. It was an absolute blast! Big thanks for sharing your cool work with us! 👩🏻‍💻👨🏻‍💻🔬🎤 #bioimageanalysis #statisticalearning