Chest CT Super-resolution and Domain-adaptation using Memory-efficient 3D Reversible GANs

Tycho F.A. van der Ouderaa^{[1, 2]}    Daniel E. Worrall^[1]    Bram van Ginneken^[2]

^[1] University of Amsterdam, Amsterdam, The Netherlands
^[2] Diagnostic Image Analysis Group, Radboud UMC, Nijmegen, The Netherlands

In MIDL 2019

Paper | Code | Poster

Abstract

Medical imaging data are typically large in size. As a result, it can be difficult to train deep neural models on them. The activations of invertible neural networks do not have to be stored to perform backpropagation, therefore such networks can be used to save memory when handling large data volumes. We use a technique called additive coupling to obtain a memory-efficient partially-reversible image-to-image translation model. With this model, we perform a 3D super-resolution and 3D domain-adaptation task, on both paired and unpaired CT scan data. Additionally, we demonstrate experimentally that the model requires significantly less GPU memory than models without reversibility.

Paper

Link to Paper, MIDL 2019

Code

A PyTorch implementation has been made available in this Github repository.
For questions about the code, please create a ticket on the github project or contact Tycho van der Ouderaa.

Related Work

• Tycho F.A. van der Ouderaa, Daniel E. Worrall. "Reversible GANs for Memory-efficient Image-to-Image Translation", in CVPR 2019.
• Phillip Isola, Jun-Yan Zhu, Tinghui Zhou, and Alexei A. Efros. "Image-to-Image Translation with Conditional Adversarial Networks", in CVPR 2017.
• Jun-Yan Zhu*, Taesung Park*, Phillip Isola, and Alexei A. Efros. "Unpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks" in ICCV 2017.
• Aidan N. Gomez, Mengye Ren, Raquel Urtasun, Roger B. Grosse. "The Reversible Residual Network: Backpropagation Without Storing Activations", in NIPS 2017.

Acknowledgements

We grateful to the Diagnostic Image Analysis Group (DIAG) of the Radboud University Medical Center, and in particular Prof. Dr. Bram van Ginneken for his collaboration on this project. We also thank the Netherlands Organisation for Scientific Research (NWO) for supporting this research and providing computational resources.