CellSynthesis

3D Fluorescence Microscopy Data Synthesis for Segmentation and Benchmarking
D. Eschweiler, M. Rethwisch, M. Jarchow, S. Koppers, J. Stegmaier, PLoS ONE, 2021

This repository contains code used for our 3D cell simulation and synthesis pipeline. We are actively working on improving synthetic data quality, computation performance and user-friendliness.

Data Preparation

The data needs to be in a hdf5 format containing image data for the network input and positional + shape information as output. The data is assumed to be in a structure similar to the following schematic.

-|data_root
----|experiment1
--------|images
--------|masks
----|experiment2
--------|images
--------|masks

To prepare your own TIF data, proceed as explained in the following steps:

1. Convert the data using utils.h5_converter.prepare_images and utils.h5_converter.prepare_masks to prepare image and mask data, respectively.
2. Create a .csv filelist using utils.csv_generator.create_csv, while the input is assumed to be a list of tuples containing image-mask pairs ->
   [('experiment1/images_converted/im_1.h5', 'experiment1/masks_converted/mask_1.h5'),
...,
('experiment2/images_converted/im_n.h5', 'experiment2/masks_converted/mask_n.h5')]
   

Training and Application

This pipeline was tested on Ubuntu 18 and 20 using python 3.7. A minimally required list of python packages and versions can be found in requirements.txt. For training and application use the provided train_script.py and apply_script.py and make sure to adjust the data paths in the models.GAN accordingly.