DECO-DIP denoises and deconvolves microscopy images. It is based on Deep image prior (DIP) [1] and does not require any training data. This version of DIP provides a new loss function, that includes an additional term to model the forward model of the imaging process. Additionally, the time dependence of image series can be used.
Synthetic test data can be created using the code from TDEntropyDeconvolution, which is a submodule of this repo [2]. See the documentation for more info.
-
Clone the repo and initialize submodules by running
git clone --recurse-submodules https://github.com/lin17a/DECO-DIP
in the command line.
-
Apply patch files for the deep-image-prior repo:
(cd dip && git apply ../dip.patch) -
Set up the python virtual environment:
python -m venv ./path/to/new/virtual/environment source ./path/to/new/virtual/environment/bin/activate pip install -r requirements.txt -
Run the program. With the parameter param_path you can specify a yaml file containing the parameters. Default is ./parameters.yaml.
./main.py --param_path parameters.yamlIf you want to run the program with more than one parameter file, you can specify a folder with parameter files. All yaml files in that folder are processed successively.
Example config files can be found in ./example_configs and default parameters are stored in ./default_parameters.yaml.
For detailed parameter descriptions see ./default_parameters.yaml and ./docs/parameter_description.md.
[1] Ulyanov, D., Vedaldi, A., & Lempitsky, V. 2020. "Deep Image Prior". International Journal of Computer Vision 128 (7): 1867–88. https://doi.org/10.1007/s11263-020-01303-4.
[2] L. Woelk, S. A. Kannabiran, V. Brock, Ch. E. Gee, Ch. Lohr, A. H. Guse, B. Diercks, and R. Werner. 2021. "Time-Dependent Image Restoration of Low-SNR Live Cell Ca2+ Fluorescence Microscopy Data". International Journal of Molecular Sciences 22 (21): 11792. https://doi.org/10.3390/ijms222111792.