Pronounced scientist. A project for the automated detection of synaptic partners in Electron Microscopy brain data.
This is the original implementation of the paper "Synaptic partner prediction from point annotations in insect brains".
Preliminary, not well documented version. At the moment, this repos mainly serves as a reference for the paper to see the original U-Net architecture in tensorflow
here and the training pipeline in gunpowder
.
No installation is required, but following dependencies:
- The gunpowder training pipeline currently runs with this hacky (old) gunpowder version:
- tensorflow = 1.3.0
- augment
- h5py
Note: We internally use docker, so this has not fully been tested.
Sorry :( not yet available, but coming soon.
Start the training script
Inference scripts and trained networks coming soon.
We provide a small test dataset cropped from CREMI dataset Sample C
, which includes 1) our network predictions, 2) ground truth segmentation, 3) raw data that can be used to run the synaptic partner extraction script.
Donwload the dataset from here. You can have a look at the data with this notebook.
In order to obtain synaptic partners from the U-Net output and validate with CREMI stats, run this extraction and evaluation script. As a record, we created the cropped dataset with this script.
- we provide network predictions (inference) for CREMI samples a, b, c and a+, b+, c+ here
- trained networks in tensorflow.
- for extracting synaptic partners, we used automatically generated segmentation for a+,b+,c+: segmentation download.
- you can find the synaptic partners used for submission in cremi format here
- original code for training, prediction, and extraction here.