Towards generative adversarial networks as a new paradigm for radiology education

  • 2018-12-04 17:32:16
  • Samuel G. Finlayson, Hyunkwang Lee, Isaac S. Kohane, Luke Oakden-Rayner
  • 31

Abstract

Medical students and radiology trainees typically view thousands of images inorder to "train their eye" to detect the subtle visual patterns necessary fordiagnosis. Nevertheless, infrastructural and legal constraints often make itdifficult to access and quickly query an abundance of images with auser-specified feature set. In this paper, we use a conditional generativeadversarial network (GAN) to synthesize $1024\times1024$ pixel pelvicradiographs that can be queried with conditioning on fracture status. Wedemonstrate that the conditional GAN learns features that distinguish fracturesfrom non-fractures by training a convolutional neural network exclusively onimages sampled from the GAN and achieving an AUC of $>0.95$ on a held-out setof real images. We conduct additional analysis of the images sampled from theGAN and describe ongoing work to validate educational efficacy.

 

Quick Read (beta)

loading the full paper ...