基于生成对抗网络的肺结节数据增扩技术研究

Research on Lung Nodules Data Augmentation Technology Based on Generative Adversarial Networks

摘要: 针对医疗影像中特定种类数据不足的问题，主要探索了基于生成对抗网络(GAN)的肺结节数据增扩技术。采用了改进的生成对抗网络技术对原始肺结节影像数据集进行数据增扩。并在LIDC-IDRI数据集上进行了验证。实验结果表明，与DCGAN相比，基于WGAN-GP网络结构的GAN技术在生成肺结节影像方面FID指标均值达到137.85，表现更佳。此外，经过其他三种生成图像质量评估指标综合测试，WGAN-GP网络生成的肺结节数据更接近真实数据分布，生成图像质量较高。

Abstract: In response to the issue of insufficient data of specific types in medical imaging, this study primarily explores lung nodules data augmentation techniques based on Generative Adversarial Networks (GAN). An enhanced generative adversarial network technique was employed to augment the original lung nodule image dataset, which was subsequently validated using the LIDC-IDRI dataset. Experimental findings demonstrate that, compared to DCGAN, GAN technology utilizing the WGAN-GP network architecture excels in generating lung nodules images, yielding an average FID index of 137.85. Furthermore, following comprehensive evaluation of three additional image quality assessment metrics, it was found that lung nodules data generated by the WGAN-GP network closely approximates the distribution of real data, resulting in higher image quality.