欢迎访问《图学学报》 分享到:

图学学报 ›› 2022, Vol. 43 ›› Issue (4): 570-578.DOI: 10.11996/JG.j.2095-302X.2022040570

• 图像处理与计算机视觉 • 上一篇    下一篇

一种基于 CycleGAN 改进的低剂量 CT 图像增强网络

  

  1. 1. 徐州医科大学医学影像学院,江苏 徐州 221004;
    2. 中国科学院苏州生物医学工程技术研究所医学影像技术研究室,江苏 苏州 215163;
    3. 苏州高新区人民医院放射科,江苏 苏州 215129
  • 出版日期:2022-08-31 发布日期:2022-08-15
  • 通讯作者: 高欣(1975),男,研究员,博士。主要研究方向为低剂量锥束 CT、基于智能计算的精准医疗、手术导航及机器人
  • 作者简介:廖仕敏(1997),男,硕士研究生。主要研究方向为医学图像处理
  • 基金资助:
    国家自然科学基金项目(81871439,61801475);中科院苏州医工所自主部署项目(Y95K091K05)

An improved low-dose CT image enhancement network based on CycleGAN

  1. 1. School of Medical Imaging, Xuzhou Medical University, Xuzhou Jiangsu 221004, China;
    2. Medical Imaging Department, Suzhou Institute of Biomedical Engineering and Technology, Chinese Academy of Sciences, Suzhou Jiangsu 215163, China;
    3. Department of Radiology, the People’s Hospital of Suzhou New District, Suzhou Jiangsu 215129, China
  • Online:2022-08-31 Published:2022-08-15
  • Contact: GAO Xin (1975), researcher, Ph.D. His main research interests cover low-dose cone-beam CT, precision medicine based on intelligent computing, surgical navigation and robot
  • About author:LIAO Shi-min (1997), master student. His main research interest covers medical image processing
  • Supported by:
    National Natural Science Foundation of China (81871439, 61801475); Suzhou Institute of Biomedical Engineering and Technology (CAS) Planned Project (Y95K091K05)

摘要:

低剂量 CT 是一种有效且相对安全的胸腹部疾病筛查手段,但图像中的伪影和噪声会严重影响医生的诊断。基于深度学习的图像增强方法中网络训练大多依赖于难以获取的配对数据,即同一患者相同部位像素级匹配的低剂量和常规剂量 CT 图像。针对非配对数据,提出了一种基于循环一致性生成对抗网络(CycleGAN)改进的低剂量 CT 图像增强网络,在生成器前添加浅层特征预提取模块,增强对 CT 图像特征的提取能力;并利用深度可分离卷积替换生成器中的部分普通卷积,减少网络参数和显存占用。该网络使用 3 275 张低剂量 CT 图像和 2 790 张非配对常规剂量 CT 图像进行训练,另外 1 716 张低剂量 CT 图像进行测试。结果表明,该网络生成的CT 图像的平均感知图像质量评价指标(PIQE)为 45.53,比 CycleGAN 的结果降低了 8.3%,更远低于三维块匹配滤波(BM3D) 31.9%、无监督图像转换网络(UNIT) 20.9%,且在结构细节保持、噪声和伪影抑制方面均获得了更好的主观视觉效果,是一种具有潜在临床应用前景的低剂量 CT 图像增强方法。

关键词: 低剂量 CT, 图像增强, 深度学习, 非配对数据, 循环一致性生成对抗网络

Abstract:

Low-dose CT is an effective and relatively safe screening method for thoracic and abdominal diseases, but the artifacts and noise in the image will seriously affect the doctor’s diagnosis. Network training in image enhancement methods based on deep learning mostly relies on paired data that is pixel-level matched low-dose and conventional-dose CT images at the same site of the same patient. An improved low-dose CT image enhancement network based on the cycle-consistent generative adversarial network (CycleGAN) was proposed for unpaired data. A shallow feature pre-extraction module was added in front of the generator to enhance the capability to extract CT images features. In addition, the depthwise separable convolution was used to replace some common convolutions in the generator to decrease network parameters and reduce GPU memory usage. In the proposed network, a total of 3 275 two-dimensional low-dose CT slices and a total of 2 790 two-dimensional unpaired conventional-dose CT slices were used for training, and a total of 1 716 two-dimensional low-dose CT slices were employed for testing. The results show that the averaged perception-based image quality evaluator (PIQE) of CT images generated by the network is 45.53, which is 8.3% lower than that of CycleGAN, 31.9% lower than that of Block-Matching and 3D filtering (BM3D), and 20.9% lower than that of unsupervised image translation network (UNIT). Moreover, the proposed network can produce better subjective visual effects in terms of structural detail preservation, noise and artifact suppression. This shows that the network is a low-dose CT image enhancement method with potential clinical prospects.

Key words: low-dose CT, image enhancement, deep learning, unpaired data, cycle-consistent generative adversarial network

中图分类号: