Abstracts

RETURN TO ABSTRACT LISTING


E2293. Generative Adversarial Networks: Definitions, Applications, Pitfalls, and Potential Applications
Authors
  1. Tushar Garg; Seth GS Medical College & KEM Hospital
  2. Apurva Shrigiriwar; Seth GS Medical College & KEM Hospital
Background
In recent years the use of neural network-based deep learning in image analysis has increased tremendously. Generative adversarial networks (GANs) are a newer deep learning environment invented by Ian Goodfellow et al. The main aim of these networks is to generate new images from raw data. GANs consist of two neural networks that are trained simultaneously: one network tries to synthesize samples that resemble real data points, while a second network is trained to differentiate synthesized samples from real samples. This concept can be used to generate synthetic diagnostic images but also for image annotation. GANs usually generate impressive results, but it is difficult to grasp their real value in image processing due to their apparent complexity. Radiologists must be well-versed with the potential of this technique and its pitfalls.

Educational Goals / Teaching Points
The goals of this exhibit are to: discuss the basic concepts in deep learning, to provide a basic introduction to GANs, discuss the applications of GANs in radiology: image reconstruction and denoising, data augmentation, transfer between different radiology modalities, synthesis, segmentation, discuss the limitations and pitfalls of GANs and highlight potential applications of GANs.

Key Anatomic/Physiologic Issues and Imaging Findings/Techniques
GANs are classified into three structures: network, loss function, and input noise. Different networks perform different functions like Deep Convolutional GAN (DCGAN) improves the quality of generated images, and 3DGAN generates 3D images from 2D images. Loss of function structures involves Least Squares (LSGAN) GAN and Wasserstein GAN (WGAN), which stabilize the training using least square loss function and Wasserstein distance and mode collapse, respectively. Conditional GAN (CGAN) and Information maximizing GAN (InfoGAN) are input noise structures that help in generation imaging by adding inputting noise and label together. Some examples of applications of GANs in radiology are the generation of high-resolution mammography using Progressive GAN, transforming magnetic resonance images to computerized tomography (CT) images or vice versa using cycleGAN, denoising of low dose CT images using WGAN, segmentation, and bone suppression in chest x-ray using pix2pix and anomaly detection in optical coherence tomography of the retina using GANs.

Conclusion
GANs are state-of-the-art networks for the unsupervised medical images synthesis, which may have applications in many radiology tasks. They enable creating new data and can be used for clinical care, education, and research.