Multi-StyleGAN: Towards Image-Based Simulation of Time-Lapse Live-Cell Microscopy

Christoph Reich*, Tim Prangemeier*, Christian Wildner & Heinz Koeppl
*Christoph Reich and Tim Prangemeier - both authors contributed equally

| Project Page | Paper | Dataset | Poster | Slides |

This repository includes the official and maintained PyTorch implementation of the paper Multi-StyleGAN: Towards Image-Based Simulation of Time-Lapse Live-Cell Microscopy.

Abstract

Time-lapse fluorescent microscopy (TLFM) combined with predictive mathematical modelling is a powerful tool to study the inherently dynamic processes of life on the single-cell level. Such experiments are costly, complex and labour intensive. A complimentary approach and a step towards completely in silico experiments, is to synthesise the imagery itself. Here, we propose Multi-StyleGAN as a descriptive approach to simulate time-lapse fluorescence microscopy imagery of living cells, based on a past experiment. This novel generative adversarial network synthesises a multi-domain sequence of consecutive timesteps. We showcase Multi-StyleGAN on imagery of multiple live yeast cells in microstructured environments and train on a dataset recorded in our laboratory. The simulation captures underlying biophysical factors and time dependencies, such as cell morphology, growth, physical interactions, as well as the intensity of a fluorescent reporter protein. An immediate application is to generate additional training and validation data for feature extraction algorithms or to aid and expedite development of advanced experimental techniques such as online monitoring or control of cells.

If you find this research useful in your work, please cite our paper:

@inproceedings{Reich2021,
    title={{Multi-StyleGAN: Towards Image-Based Simulation of Time-Lapse Live-Cell Microscopy}},
    author={Reich, Christoph and Prangemeier, Tim and Wildner, Christian and Koeppl, Heinz},
    booktitle={{International Conference on Medical image computing and computer-assisted intervention}},
    year={2021},
    organization={Springer}
}

Method

Figure 1. Architecture of Multi-StyleGAN. The style mapping network $f$ (in purple) transforms the input noise vector $z\sim \mathcal{N}_{512}(0, 1)$ into a latent vector $w\in\mathcal{W}$ , which in turn is incorporated into each stage of the generator by three dual-style-convolutional blocks. The generator predicts a sequence of three consecutive images for both the brightfield and green fluorescent protein channels. The U-Net discriminator [2] distinguishes between real and a fake input sequences by making both a scalar and a pixel-wise real/fake prediction. Standard residual discriminator blocks in gray and non-local blocks in blue.

Figure 2. Dual-styled-convolutional block of the Multi-StyleGAN. The incoming latent vector w is transformed into the style vector s by a linear layer. This style vector modulates (mod) the convolutional weights $\theta_{b}$ and $\theta_{g}$ , which are optionally demodulated (demod) before convolving the (optionally bilinearly upsampled) incoming features of the previous block. Learnable biasses ( $b_{b}$ and $b_{g}$ ) and channel-wise Gaussian noise ( $\mathcal{N}$ ) scaled by a learnable constant (cb and cg), are added to the features. The final output features are obtained by applying a leaky ReLU activation.

Results

Figure 3. Samples generated by Multi-StyleGAN. Brightfield channel on the top and green fluorescent protein on the bottom.

Table 1. Evaluation metrics for Multi-StyleGAN and baselines.

Model	FID (BF) $\downarrow$	FVD (BF) $\downarrow$	FID (GFP) $\downarrow$	FVD (GFP) $\downarrow$
Multi-StyleGAN	33.3687	4.4632	207.8409	30.1650
StyleGAN 2 3d + ADA + U-Net dis.	200.5408	45.6296	224.7860	35.2169
StyleGAN 2 + ADA + U-Net dis.	76.0344	14.7509	298.7545	31.4771

Dependencies

All required Python packages can be installed by:

pip install -r requirements.txt

To install the necessary custom CUDA extensions adapted from StyleGAN 2 [1] run:

cd multi_stylegan/op_static
python setup.py install

The code is tested with PyTorch 1.8.1 and CUDA 11.1 on Ubuntu with Python 3.6! Using other PyTorch and CUDA versions newer than PyTorch 1.7.0 and CUDA 10.1 should also be possible. Please note using a different PyTorch version eventually requires a different version of Kornia or Torchvision.

Data

Our proposed time-lapse fluorescent microscopy is available at this url.

The dataset includes 9696 images structured in sequences of both brightfield and green fluorescent protein (GFP) channels at a resolution of 256 × 256. Data loader classes can be found in the python package dataset.

Trained Model

The checkpoint of our trained Multi-StyleGAN is available at this url.

The checkpoint (PyTorch state dict) includes the EMA generator weights ("generator_ema"), the generator weights ("generator"), the generator optimizer state ("generator_optimizer"), the discriminator weights ("discriminator"), the discriminator optimizer state ("discriminator_optimizer"), and the path-length regularization states ("path_length_regularization")

Usage

To train Multi-StyleGAN in the proposed setting run the following command:

 python -W ingore train_gan.py --cuda_devices "0, 1, 2, 3" --data_parallel --path_to_data "60x_10BF_200GFP_200RFP20_3Z_10min"

Dataset path and cuda devices may differ on other systems! To perform training runs with different settings use the command line arguments of the train_gan.py file. The train_gan.py takes the following command line arguments:

Argument	Default value	Info
--cuda_devices (str)	`"0, 1, 2, 3"`	String of cuda device indexes to be used.
--batch_size (int)	`24`	Batch size to be utilized while training.
--data_parallel (binary flag)	False	Binary flag. If set data parallel is utilized.
--epochs (int)	100	Number of epochs to perform while training.
--lr_generator (float)	`2e-04`	Learning rate of the generator network.
--lr_discriminator (float)	`6e-04`	Learning rate of the discriminator network.
--path_to_data (str)	`"./60x_10BF_200GFP_200RFP20_3Z_10min"`	Path to dataset.
--load_checkpoint (str)	`""`	Path to checkpoint to be loaded. If `""` no loading is performed.
--resume_training (binary flag)	False	Binary flag. If set training is resumed and so cut mix aug/reg and wrong order aug is used.
--no_top_k (binary flag)	False	Binary flag. If set no top-k is utilized.
--no_ada (binary flag)	False	Binary flag. If set no adaptive discriminator augmentation is utilized.

To generate samples of the trained Multi-StyleGAN use the get_gan_samples.py script.

 python -W ingore scripts/get_gan_samples.py --cuda_devices "0" --load_checkpoint "checkpoint_100.pt"

This script takes the following command line arguments:

Argument	Default value	Info
--cuda_devices (str)	`"0"`	String of cuda device indexes to be used.
--samples (int)	`100`	Number of samples to be generated.
--load_checkpoint (str)	`"checkpoint_100.pt"`	Path to checkpoint to be loaded.

To generate a latent space interpolation use the gan_latent_space_interpolation.py script. For producing the final .mp4 video ffmpeg is required.

 python -W ingore scripts/gan_latent_space_interpolation.py --cuda_devices "0" --load_checkpoint "checkpoint_100.pt"