eduzhai > Applied Sciences > Engineering >

Hierarchical Amortized Training for Memory-efficient High Resolution 3D GAN

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 12 pages

Abstract: Generative Adversarial Networks (GAN) have many potential medical imagingapplications, including data augmentation, domain adaptation, and modelexplanation. Due to the limited embedded memory of Graphical Processing Units(GPUs), most current 3D GAN models are trained on low-resolution medicalimages. In this work, we propose a novel end-to-end GAN architecture that cangenerate high-resolution 3D images. We achieve this goal by separating trainingand inference. During training, we adopt a hierarchical structure thatsimultaneously generates a low-resolution version of the image and a randomlyselected sub-volume of the high-resolution image. The hierarchical design hastwo advantages: First, the memory demand for training on high-resolution imagesis amortized among subvolumes. Furthermore, anchoring the high-resolutionsubvolumes to a single low-resolution image ensures anatomical consistencybetween subvolumes. During inference, our model can directly generate fullhigh-resolution images. We also incorporate an encoder with a similarhierarchical structure into the model to extract features from the images.Experiments on 3D thorax CT and brain MRI demonstrate that our approachoutperforms state of the art in image generation and clinical-relevant featureextraction.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×