东北大学学报:自然科学版 ›› 2020, Vol. 41 ›› Issue (3): 332-336.DOI: 10.12068/j.issn.1005-3026.2020.03.006

• 信息与控制 • 上一篇    下一篇

基于GAN的医学图像仿真数据集生成算法

孟琭1, 钟健平1, 李楠2   

  1. (1.东北大学 信息科学与工程学院, 辽宁 沈阳110819; 2.沈阳产品质量监督检验院, 辽宁 沈阳110000)
  • 收稿日期:2019-06-17 修回日期:2019-06-17 出版日期:2020-03-15 发布日期:2020-04-10
  • 通讯作者: 孟琭
  • 作者简介:孟琭(1982-),男,辽宁沈阳人,东北大学副教授.
  • 基金资助:
    国家自然科学基金资助项目(61973058).

Generating Algorithm of Medical Image Simulation Data Sets Based on GAN

MENG Lu1, ZHONG Jian-ping1, LI Nan2   

  1. 1. School of Information Science & Engineering, Northeastern University, Shenyang 110819, China;2. Shenyang Product Quality Supervision and Inspection Institute, Shenyang 110000, China.
  • Received:2019-06-17 Revised:2019-06-17 Online:2020-03-15 Published:2020-04-10
  • Contact: MENG Lu
  • About author:-
  • Supported by:
    -

摘要: 基于生成对抗网络(generative adversarial networks,GAN),提出了面向肝脏肿瘤CT图像仿真数据集生成深度学习算法.首先,将CT图像数据文件进行格式解析,单独保存为PNG格式的图像文件;然后,将肝脏病变区域统一标注为白色,并结合肝脏CT原图组成配对图片;最后,用生成对抗网络的pix2pix架构仿真生成病变肝脏图像.为将生成图像与目标图像进行定量分析、比较,本文采用了峰值信噪比和结构相似性作为模型的评价指标.实验结果表明,本文算法所生成的肝脏肿瘤CT仿真数据集的平均峰值信噪比为64.72dB,平均结构相似性为0.9973,证明了所生成的仿真图像数据有着非常高的真实度.

关键词: 生成对抗网络, 图像处理, 肝脏图像仿真, 参数调整, 数据增强

Abstract: Based on generative adversarial networks (GAN), a deep learning algorithm for generating diseased liver CT image data sets was proposed. Firstly, the CT image data file was formatted and saved as an image file in PNG format. Then the liver lesion area was uniformly marked as white, and the liver CT original image was combined to form a paired picture. Finally, diseased liver image was generated using a pix2pix architecture that created an anti-network. In order to quantitatively analyze and compare the generated image with the target image, the peak signal-to-noise ratio and structural similarity were used to evaluate the model. The results showed that the average peak signal-to-noise ratio of the simulated CT diseased liver image generated by the proposed algorithm is 64.72dB and the average structural similarity is 0.9973, thus proving these simulated image data have very high trueness.

Key words: generative adversarial networks(GAN), image processing, liver image simulation, parameter adjustment, data augmentation

中图分类号: