Context-Gan: Controllable Context Image Generation Using Gans

We propose an enhancement to label-to-image GANs. Based on a Pix2Pix architecture, ConText-GAN allows generating images in a controlled way. Given a feature map as input, ConText-GAN can generate images with a specified layout and label content. As an application, ConText-GAN is used to perform a more realistic than usual data augmentation from an MRI dataset. We show the validity of the generated images with respect to the input feature maps. The relevance of the approach is demonstrated by the improvement of the segmentation result following a data augmentation performed with ConText-GAN compared to classical methods. A practical application is presented in the challenging context of U-Net segmentation of MRI of fat infiltrated muscles.