Huge update to Convpaint, and preprint out on bioRxiv! We develop a pixel classifier similar to Ilastik, but built for napari & with technical improvements. Now even works on animals! 1/10

USER FRIENDLY: Easy to install and use, even with multi-dimensional data (3D / timeseries / multichannel). Video shows annotation, iterative training and segmentation in ~30 seconds. All interactive. 2/10

PIPELINES: Can be combined with other @napari_imaging plugins into complex workflows without coding! We show different use cases in the paper, and step-by-step instructions in the documentation: 3/10

TECHNICAL DETAILS: We use pretrained neural networks (CNNs/ViTs) to extract local image features, and use those to train a small random forest classifier with labels scribbled by user. Modular framework to easily add other feature extraction algorithms. 4/10

This allows to "repurpose" pretrained neural nets for specific use cases without retraining, saving time and energy. We implement lots of tricks to improve performance (tiling, parallel processing, lazy loading for large files) -- runs on cheap laptops! 5/10

PREPRINT: For implementation details and performance quantification, check the CONVPAINT paper! We tested diverse datasets and cover many use cases. 6/10

PROGRAMMATIC CONTROL: In @Pertzl we use Convpaint for feedback microscopy, segmenting cells in data live-streamed from the microscope, allowing to automate optogenetic stimulation. Excited to see what other labs will do with it! 7/10

CODE is all open source on github. Feel free to drop feature requests and file issues! 8/10

The project has come a looooong way since first sharing on twitter. Makes me feel nostalgic to look at this prototype! Huge thanks to the community for feedback, CZI for funding, and everyone who contributed to the code and paper <3 9/10

10/10 THANKS: @guiwitz @RomanSchwob @VetAna_Sto @macdobry @benigraedel @PertzL @czi @IZB_unibern Data Science Lab UNIBE, Joël & Mykhailo not on twitter, and first external contributor to the repo @psobolewskiPhD :D

Congrats, great to see it out! I’ve been using convpaint for a while and it’s working really well for me!