Zebrafish Microglia

The goal of this project is to accurately segment microglia in 3D from Z-stack images of zebrafish heads collected on a spinning disk confocal. Accurate segmentation of the microglia in 3D is critical to classifying the different states they present in. Some challenges with this data set included a very large z step, high background and inconsistent autofluorescence from other tissues in the samples.

I used transfer learning in CellPose's architecture to tune their cyto3 model to detect the fine protrusions and heterogeneous shapes of the microglia. Because of the large Z-step, training a 3D model in CellPose would be challenging for this dataset. To circumvent this, I leveraged training a 2D CellPose model with post-segmentation stitching to generate the final 3D objects.

A combination of imaris_ims_file_reader, pyclesperanto (GPU accelerated image processing), CellPose, numpy, scikit-image, and pandas was used to created a reproducible pipeline to analyze this data set in batch.