Sea Urchin EM Segmentation

This repository contains scripts for (semi-manual) segmentation of neurons in a large EM-Volume of a sea urchin. There are different scripts for the parts of the project:

Training

The folder training contains the scripts for training neural networks for segmenting the neuron boundaries in EM. Here, we make use of the method published in Matskevych et al.. It segemnts the neuron boundaries in three steps:

1. Initial segmentation of the boundaries with ilastik pixel classification.
2. Improving the boundary segmentation with an Enhancer, a network that was trained to improve the pixel classification predictions on data with segmentation groundtruth from www.cremi.org.
3. Training a final segmentation network that learns to segment boundaries from the EM data directly using the predictions from the Enhancer as target signal.

The scripts train_precomputed.py and train_pseudolabels.py are used to train different versions of the final semgentation network, the other scripts are used for data preprocessing and for checking intermediate results.

Segmentation

The folder segmentation contains the scripts to segment neuron fragments in the full EM volume. It uses the method from Pape et al. based on the boundary predictions of the segmentation network. This segmentation is then converted into a paintera project, which enables manually assembling full neurons out of the neuron fragments.

The script segment_full_volume.py runs the segmentation workflow, to_paintera.py converts the result to the paintera format.

Data extraction

The folder tif_extraction contains scripts to extract data to other data formats that can be loaded with Amira:

• extract_highres_cutout.py to extract a cutout from the full EM raw data to tif (at full resolution).
• extract_lowres_volume.py to extract the full EM raw data to tif (at a low resolution).
• extract_meshes_painera.py to extract reconstructed neurons from paintera as meshes in the ply format.

Installation

You can set-up a conda environment with all necessary dependencies to run these scripts via:

conda env create -f environment.yaml

Note that you will need a powerful workstation with a gpu or access to a computer cluster with gpu nodes to run the network training and the segmentation workflow.