Modeling of Cell Shape Using 3D GANs

The ongoing advancement of deep-learning generative models, showing great interest of the scientific community since the introduction of the generative adversarial networks (GAN), paved the way for generation of realistic data. The utilization of deep learning for the generation of realistic biomedical images allows to alleviate the constraints of the parametric models, limited by the employed mathematical approximations. Building further upon the foundation laid by the original GAN, the 3D GAN added another dimension, allowing generation of fully 3D volumetric data. We present an approach to generating fully 3D volumetric cell masks using 3D GANs. Proposed model is able to generate high-quality cell masks with variability matching the real data.

Requirements

The available source codes were tested on a machine with CUDA-capable GPU (Quadro P6000) running Ubuntu 18.04 with following versions of software:

• GENERATIVE MODEL
  • Python 3.6.8
  • TensorFlow 1.12.0
  • TensorBoard 1.12.0 (optional, used for training visualization)
  • SciPy 1.1.0
  • NumPy 1.15.4
• DATA PROCESSING
  • Matlab R2019a
  • DIPimage 2.9

Quick Start

• Generating data
  Download archive with the trained model, open it in terminal and run:
  $> python main.py generate ./tf_checkpoints/e1000.cptk
• Training the model
  Download archive with the sample dataset and the source code. Set "b_training_data_augmented" as the training data (TRAIN_DATA_DIR) for the model in "02_gan-3d-v10-ICIAP/configuration.py". Then, open the model directory "02_gan-3d-v10-ICIAP" in terminal and run:
  $> python main.py train
• Training visualization
  Open the model directory "02_gan-3d-v10-ICIAP" in terminal and run:
  $> tensorboard --logdir=./tf_logs --port=7088
  The TensorBoard visualization will be subsequently accessible from the web browser at "<ip_address>:7088".

Download

• sample dataset (33 MB)
• trained model (301 MB)
• source code (18 kB)

License

The source code is freely available under the MIT license. Copyright belongs to the Centre for Biomedical Image Analysis (CBIA) at Masaryk University.

References

• Wiesner D, Nečasová T, Svoboda D. On Generative Modeling of Cell Shape Using 3D GANs. In Ricci Elisa, Rota Buló Samuel, Snoek Cees, Lanz Oswald, Messelodi Stefano, Sebe Nicu. Image Analysis and Processing – ICIAP 2019. LNCS 11752., Trento: Springer, 2019. p. 672-682, 11 pp., ISBN 978-3-030-30645-8. doi:10.1007/978-3-030-30645-8_61.

Acknowledgement

This project was funded by Czech Science Foundation, grant No. GA17-05048S.