CellCycleGAN

Spatiotemporal Microscopy Image Synthesis of Cell Populations using Statistical Shape Models and Conditional GANs.

Installing Dependencies

• CellCycleGAN requires a working MATLAB installation including the image processing toolbox. We tested the software on R2019b (Ubuntu) and R2020a (macOS).
• Moreover, a Python environment needs to be installed to run the pretrained model files. The repository contains the file environment.yml where all the dependencies of the Python scripts are contained. We recommend to use an environment management system such as Anaconda which can be obtained, e.g., from https://docs.conda.io/en/latest/miniconda.html .
• Start the Anaconda prompt and call the following command to setup a new Python environment called ccg with the above-mentioned dependencies: conda env create -p /path/to/my/environments/ccg -f /path/to/the/file/environment.yml.
• After all dependencies have been installed successfully, switch to the environment by calling conda activate /path/to/my/environments/ccg.
• Finally, start MATLAB by invoking the command matlab from the terminal. If the command is not found (e.g., if MATLAB was not added to the PATH environment variable before), navigate to the bin folder of your MATLAB installation and run MATLAB by starting ./matlab.

Ground Truth Preparation

1. Download the supplementary material of Zhong et al., 2012 from https://static-content.springer.com/esm/art%3A10.1038%2Fnmeth.2046/MediaObjects/41592_2012_BFnmeth2046_MOESM257_ESM.zip and extract it to a destination of your choice.
2. Copy the files ExtractImageData.m and SegmentCenterNucleus.m to the extracted %SI_ZIP_CONTENT%/code/ folder, create a new directory for the training images and execute the file ExtractImageData.m. You'll be asked for an output folder, where you can specify the previously created one.
3. The images from the archive will be processed and written to a *.h5 format that will be used for CNN training.

Generation of Synthetic Image Data

• The data generation can be started in MATLAB by executing the script CellCycleGAN.m. To make it run properly on your system, you have to adapt at least the output folder path in line 38 (settings.outputFolder).
• After adjusting the output path properly, the script should run and should produce result images. Note that the processing can take a few minutes depending on the number of frames you simulate and on the performance of your machine.
• The Parameters for the data generation are highlighted in the CellCycleGAN.m script and can be adjusted according to the desired image quality and to the difficulty level to be generated.
• Note that the script sets the seed for the random generator right in the beginning of the script (settings.randomSeed). Make sure to change this seed if you want to generate a different image sequence.
• Results are stored in the Cell Tracking Challenge format (http://celltrackingchallenge.net/) and should thus be usable with the performance evaluation tools of the challenge.
• A demonstration video of a created data set can be seen below and we created five similar demo data sets with the default settings and random initializations that can be obtained here: https://bit.ly/35nODtr

<p align="center"><img src="Data/Screenshots/01_Compressed.gif" width="70%"></p>

Citation

If you find this work useful, please cite the following publication:

D. Bähr, D. Eschweiler, A. Bhattacharyya, D. Moreno-Andrés, W. Antonin, J. Stegmaier, "CellCycleGAN: Spatiotemporal Microscopy Image Synthesis of Cell Populations using Statistical Shape Models and Conditional GANs", IEEE International Symposium on Biomedical Imaging (ISBI), 2021.

References

[1] Zhong, Q., Busetto, A. G., Fededa, J. P., Buhmann, J. M., & Gerlich, D. W. (2012). "Unsupervised Modeling of Cell Morphology Dynamics for Time-Lapse Microscopy". Nature Methods, 9(7), 711-713.

[2] Ulman, V., Maška, M., Magnusson, K. E., Ronneberger, O., Haubold, C., Harder, N., ... & Smal, I. (2017). "An Objective Comparison of Cell-Tracking Algorithms". Nature Methods, 14(12), 1141-1152.