Code for the paper "Deep learning-based bias transfer for overcoming laboratory differences of microscopic images". Included is the implementation of color transfer, cycleGAN, U-Net cycleGAN, Fixed-Point GAN and all additional losses in Tensorflow 2. Training, validation and test images are defined via csv files (see data folder) and training runs are tracked via sacred (+ mongoDB and omniboard). All requirements are defined in the docker_context folder, including a Dockerfile that can be used to set up Docker images and containers for execution.
Contains an example for the required data structure.
An installable python module containing all approaches (pip install --user ./debiasmedimg)
Contains a Dockerfile for setting up a docker environment including all package requirements.
Output directory. Tensorflow checkpoints, generated images, and evaluation metrics get saved here.
Scripts using sacred to track experiments. An example script for each approach is located here. The subfolder 'configs' contains .yaml files which define the hyperparameters for the experiments.