Max Planck Institute for Dynamics and Self-Organization -- Department for Nonlinear Dynamics and Network Dynamics Group
Personal tools
Log in

Automated Cell-Segmentation with Cycle-Consistent Generative Adversarial Networks

A central problem in biomedical imaging is the preparation of images for further quantitative analysis via automated image segmentation. Especially the segmentation of images with lower quality remains challenging. Recently, deep convolutional neural networks were applied successfully in a vast variety of visual recognition tasks, including automatic biomedical image segmentation. In general, their performance is superior to rule-based methods.


One drawback of these models, however, is the necessity to prepare a well-suited dataset on which the network can be trained. Generating hand-labeled datasets of ground truth - image mask pairs is time-consuming and thus represents an expensive bottleneck. Moreover, even if extensive training data is available, the performance of these systems degrades significantly when they are applied to test data that differ from the training data, for example, due to variations in experimental protocols. Furthermore, pixel by pixel classifiers based on deep convolutional architectures can perform poorly on image data with substantial defects that have been labeled incompletely due to bleaching and label failure.

We propose a new semi-supervised image segmentation method based on generative adversarial networks (GANs) that can be trained even in absence of prepared image - mask pairs. In particular, we use a Cycle-GAN architecture to train on unpaired training data. We could show that this model performs competitively on standard segmentation tasks even when trained on just one target sample. In addition, our model generalises well to test data differing from the training data and successfully performs image segmentation tasks on samples with substantial defects.


Contact:  Stephan Eule 

Members working within this Project:

 Stephan Eule 
 Matthias Häring 
 Fred Wolf