基于多尺寸卷积与残差单元的快速收敛 GAN 胸部 X 射线图像数据增强

针对深度学习中数据增强的方法， 改进生成式对抗网络 (GAN，Generative adversarial networks)模型，形成一种快速收敛生成式对抗网络，能够克服 GAN 训练过程不稳定、收敛速度缓慢容易发生模式崩溃等问题。采用在判别器中使用多尺寸卷积，加强判别器的特征提取能力；在生成器中添加残差单元的方法，使得生成器可以快速拟合真实数据的分布；同时对判别器进行预训练的策略，有利于提高生成器前期训练稳定性和加快训练过程。运用 CIFAR-10 标准数据集进行实验，与几种基于 GAN 的模型对比，证实本文的改进算法效果较好，图像质量和多样性更优。利用本文提出的改进算法用于美国 NIH 临床数据库的胸部 X 射线数据集，生成扩充样本，经图灵测试证实了算法的有效性。 

Enhancement of Chest X-ray Image Data by Using Fast convergence GAN Based on Multi-Dimensional Convolution and Residual Unit

Aiming at the data enhancement method in deep learning, the Generative adversarial networks (GAN) model is improved to form a fast convergence generation-type confrontation network, which can overcome the problems of unstable GAN training process, slow convergence speed and easy pattern collapse. . The multi-size convolution is used in the discriminator to enhance the feature extraction ability of the discriminator; the method of adding the residual unit in the generator enables the generator to quickly fit the distribution of the real data; and pre-training the discriminator The strategy is beneficial to improve the stability of the pre-build training and speed up the training process. Experiments with CIFAR- 10 standard datasets, compared with several GAN-based models, confirm that the improved algorithm is better and the image quality and diversity are better. The improved algorithm proposed in this paper was used to generate the extended samples from the chest X-ray dataset of the NIH clinical database in the United States. The validity of the algorithm was confirmed by Turing test.