## Cycle Consistent Embedding of 3D Brains with Auto-Encoding Generative Adversarial Networks

07 Apr 2021, 15:24 (modified: 22 Jul 2022, 20:09)MIDL Short 2021 PosterReaders: Everyone
Keywords: Auto-Encoder, Latent Space, Generative Adversarial Network, Cycle consistency, 3D MRI
TL;DR: We show that a better embedding also leads to a better 3D brain image generation using a cycle consistent embedding GAN.
Abstract: Modern generative adversarial networks (GANs) have been enabling the realistic generation of full 3D brain images by sampling from a latent space prior $\mathcal{Z}$ (i.e., random vectors) and mapping it to realistic images in $\mathcal{X}$ (e.g., 3D MRIs). To address the ubiquitous mode collapse issue, recent works have strongly imposed certain characteristics such as Gaussianness to the prior by also explicitly mapping $\mathcal{X}$ to $\mathcal{Z}$ via encoder. These efforts, however, fail to accurately map 3D brain images to the desirable prior, which the generator assumes to be sampling the random vectors from. On the other hand, Variational Auto-Encoding GAN (VAE-GAN) solves mode collapse by enforcing Gaussianness by two learned parameter, yet causes blurriness in images. In this work, we show how our \textit{cycle consistent embedding} GAN (CCE-GAN) both accurately encodes 3D MRIs to the standard normal prior, and maintains the quality of the generated images. We achieve this without a network-based code discriminator via the Wasserstein measure. We quantitatively and qualitatively assess the embeddings and the generated 3D MRIs using healthy T1-weighted MRIs from ADNI.
Paper Type: methodological development
Primary Subject Area: Image Synthesis
Secondary Subject Area: Unsupervised Learning and Representation Learning
Paper Status: original work, not submitted yet
Source Code Url: https://github.com/ShiboXing/Light-3dbraingen