Segment2P: Parameter-free automated segmentation of cellular fluorescent signals

Abstract

The availability of genetically modified calcium indicators has made calcium imaging of neural signaling accessible and widespread whereby recording hundreds or even thousands of cells simultaneously is commonplace. Immunocytochemistry also produces large images with a great number of antibody labeled cells. A major bottleneck towards fully harnessing these techniques is the delineation of the neural cell bodies. We designed an online robust cell segmentation algorithm based on deep learning which does not require installation or expertise. The robust segmentation is achieved by pre-processing images submitted to the site and running them through DeepLabv3 networks trained on human segmented micrographs. The algorithm does not entail any parameter tuning; can be further trained if necessary; is robust to cell types and microscopy techniques (from immunocytochemistry to single and multi-photon microscopy) and does not require image pre-processing.