Denoising Diffusion Probabilistic Models for the Generation of Realistic Fully-Annotated Microscopy Image Data Sets

An approach for deep learning-based generation of realistic 2D and 3D microscopy image data based on sketches via denoising diffusion probabilistic models (Publication). Sketches indicating position, shape and course details of cellular structures are used as guidelines for generating corresponding realistic microscopy image data, which allows for the automated generation of fully-annoated data sets. Ultimately, we envision that this approach supports in diminishing the need for human annotation efforts for training deep learning-based segmentation in microscopy image data.

Code is available on github and generated data sets are available at OSF. If you are using code or data, please cite our work:
@article{eschweiler2022celldiffusion,
  title={Denoising diffusion probabilistic models for generation of realistic fully-annotated microscopy image data sets},
  author={Dennis Eschweiler and R{\"u}veyda Yilmaz and Matisse Baumann and Ina Laube and Rijo Roi and Abin Jose and Daniel Br{\"u}ckner and Johannes Stegmaier},
  journal={PLOS Computational Biology},
  volume={20},
  number={2},
  pages={e1011890},
  year={2023}
}


Use this demo by choosing a model and a simulated annotation mask to preview the correspondingly generated synthetic microscopy image. Feel free to download models, example masks and example image data.
Contact Dennis Eschweiler or Johannes Stegmaier if you have any questions. Additional website contribution by Matisse Baumann and Daniel Brückner.

Models

			Download Model

Masks

			Download Mask

Output

			Download Output