Setup

git clone https://github.com/fratim/detectSomata  

cd detectSomata  

conda create -n SomataDetection --file requirements.txt  

conda activate SomataDetection  

Parameter Specification

All parameter must be specified in parameters.py. Most importantly, the correct filepaths for the training segmentation, the corresponding soma mask and the segmentation for which the somata shall be predicted.

Example Data and trained Model

We provide examples for both training and prediction. The training data is taken from the JWR dataset [1]. The segmentation as well as the soma mask can be found in the folder Training_Data. The soma mask is a binary tensor that was manually created by us. The segmentation for which somata are is taken from the Zebrafinch dataset [2] and can be found in the folder Prediction_Data. We further provide the trained network in Trained_Network, which was achieved by training on the JWR data. This network has been used for the somata predictions on the Zebrafinch data that is presented in our paper.

Running Training and Prediction

To train a new network, run python run_training.py. To predict the somata of an input segmentation with a pretrained network, run python run_prediction.py.

[1] Lin, Zudi, et al. "Two Stream Active Query Suggestion for Active Learning in Connectomics." Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XVIII 16. Springer International Publishing, 2020.
[2] Kornfeld, Jörgen, et al. "EM connectomics reveals axonal target variation in a sequence-generating network." Elife 6 (2017): e24364.