DISCo: Deep learning, Instance Segmentation and Correlations for cell segmentation in calcium imaging

This is a method to perform the cell segmentaiton step in caclium imaging analysis, which uses the temporal information from caclium imaging videos in form of correlations, and combines a deep learning model with an instance segmentation algorithm.

Publication

"DISCo: Deep learning, Instance Segmentation, and Correlations for cell segmentation in calcium imaging", E. Kirschbaum, A. Bailoni, F. A. Hamprecht, arXiv preprint arXiv:1908.07957, 2019. [pdf]

Requirements:

• Python 3.6 (or later): we recommend installing it with Anaconda
• PyTorch 1.0 (or later)
• GASP
• inferno 0.3 (or later)

Preparations

1. Download or clone this repository
2. Install GASP as described here
3. Get inferno as described here
4. Download the neurofinder training and test data from here
5. Extract the neurofinder data into HDF5 files:
   • create for each neurofinder video a HDF5 file with a dataset named 'video' containing the video with shape (time x X x Y)
   • unzip the file named BF_labels.h5.zip containing the foreground-background labels for each video
   • unzip the file summary_images.h5.zip containing the mean intensity projection for each video
   • unzip the file gt_segmentations.h5.zip containing the instance labels for each video

Usage

Option	Name	Description
-p	path	Path to the folder containing the .h5 video files and the ground truth segmentations
-m	mode	Decide whether a single network is trained on all videos ('disco') or individual networks on the five dataset series ('discos')
-gpu	gpu ID	Select the GPU to train on.
-a	additional ending	Additional ending to the output filename.

Example:
python run.py -p ../neurofinder_videos/ -m disco -gpu 1