生成对抗网络在医学图像计算上的进展与展望

Review and Prospects for Generative Adversarial Networks on Medical Image Computation

生成对抗网络中生成器和判别器进行博弈的对抗训练方法在计算机视觉任务中引起了大量关注。生成器的参数更新不是直接来自真实数据样本,而是依据判别器的真伪和类别判别,从而具有生成媲美真实图像的能力。此外,生成对抗网络的对抗训练方式具有半监督/无监督训练特性,非常适合应用于医学图像计算领域,用以解决医学图像数据量少、质量低的缺陷。本文从不同角度对基于生成对抗网络的医学图像计算(医学图像合成、超分辨率重建和辅助诊断)的研究进展进行了回顾,并从模型设计、性能表现等方面对相关工作进行了概述和分析。最后,对生成对抗网络在该领域面临的挑战及潜在应用进行了展望。

In generator adversarial networks (GANs), the adversarial training mechanism between the generator and discriminator has attracted lots of attention in computer vision community. The update of the parameters in generator depends on the decision of the discriminator rather than the ground truth images, so the generator is able to synthesize more plausible images. In addition, the adversarial training mechanism makes GANs suitable for the semi-supervised/unsupervised training. This characteristic has been proven useful to address the problem that the medical images are limited and the quality is poor in the field of medical image computation. This paper reviews the researches of medical image computation (medical image synthesis, super-resolution and computer aided diagnosis) based on generative adversarial networks from different perspectives, and analyzes the related works from the aspects of model architecture and performance. Finally, the challenges and potential applications of generative adversarial networks in this field are presented.