This repository contains related code to the paper below. All relevant data, such as the pre-trained deep neural network, is available on zenodo.
To use the code, you need a Python installation together with relevant libraries (imageio, numpy, scipy, tensorflow). In general, no GPU is needed, but especially for training and large scale data inference recommended.
We provide code in training to train a U-Net-like architecture with a reduced latent space.
The default configuration has a single latent space channel, i.e. the latent space image .
For proper usage, you need the BAGLS dataset.
We provide a pre-trained model for retrieving the latent space images at zenodo.
In latent_generation, you will find the respective Jupyter notebook.
In visualize, you find a Jupyter notebook that uses a pre-trained model and its decoder,
as well as endoscopic images to show the respective latent space
and options to investigate the latent space.
Kist et al. "A single latent channel is sufficient for biomedical image segmentation", biorxiv 2021