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

Recently, paired (e.g. Pix2pix) and unpaired (e.g. CycleGAN) image-to-image translation methods have shown effective in medical imaging tasks. In practice, however, it can be difficult to apply these deep models on medical data volumes, such as from MR and CT scans, since such data volumes tend to be 3-dimensional and of a high spatial resolution pushing the limits of the memory constraints of GPU hardware that is typically used to train these models. Recent studies in the field of invertible neural networks have shown that reversible neural networks do not require to store intermediate activations for training. We use the RevGAN model that makes use of this property to perform memory-efficient partially-reversible image-to-image translation. We demonstrate this by performing a 3D super-resolution and a 3D domain-adaptation task on chest CT data volumes.