Cell-DINO: Self-supervised image-based embeddings for cell fluorescent microscopy
Théo Moutakanni,Camille Couprie,Seungeun Yi,Michael Doron,Zitong S. Chen,Nikita Moshkov,Elouan Gardes,Mathilde Caron,Hugo Touvron,Armand Joulin,Piotr Bojanowski,Wolfgang M. Pernice,Juan C. Caicedo
Microscopy is a powerful way to study what’s happening inside cells, but researchers often collect hundreds or thousands of images. This large volume of data makes it incredibly difficult to discern patterns between these images. Tired of straining their eyes, researchers trained a self-supervised algorithm. This means that you don’t need to manually label your data to analyze the images, the software can learn to do it by itself!
Scientists fed this algorithm large volumes of cell images with little direction, and allowed it to learn patterns on its own. The model was able to distinguish meaningful and subtle differences between cells that would be difficult to detect normally.
When tested in a competition against many other methods designed for the same task, the approach performed strongly, ranking in the top 8.8% and 11.6% across two evaluations with only a small gap to first place.
The hands-off approach of training models to perform a specific task is an easy way to teach algorithms. This work identifies how you can leverage self-supervised algorithms to accelerate the analysis of complex biological image data, (and give your eyes a break!)
