南方医科大学学报

南方医科大学学报 ›› 2023, Vol. 43 ›› Issue (7): 1214-1223.doi: 10.12122/j.issn.1673-4254.2023.07.18

• • 上一篇    下一篇

弦图插值结合UNIT网络图像转换的CT金属伪影校正

于佳弘，张昆鹏，靳 爽，苏 哲，徐晓桐，张 华   

  1. 南方医科大学生物医学工程学院，广东省医学图像处理重点实验室，广东省医学影像与诊断技术工程实验室，广东 广州 510515
  • 出版日期:2023-07-20 发布日期:2023-07-20

Sinogram interpolation combined with unsupervised image-to-image translation network for CT metal artifact correction

YU Jiahong, ZHANG Kunpeng, JIN Shuang, SU Zhe, XU Xiaotong, ZHANG Hua   

  1. School of Biomedical Engineering, Guangdong Provincial Key Laboratory of Medical Image Processing, Guangdong Provincial Engineering Laboratory for Medical Imaging and Diagnostic Technology, Southern Medical University, Guangzhou 510515, China
  • Online:2023-07-20 Published:2023-07-20

RichHTML

9

PDF

158

摘要： 目的 本文提出一种结合弦图域插值和基于非监督图像转换网络（UNIT）的伪影校正模型，提升图像的质量。方法 将初步校正的图像与无伪影先验图像视为两个不同图像域内的元素，通过图像转换网络后可得到去除伪影的图像。采用仿真数据进行验证实验。采用PSNR与SSIM作为定量指标评估校正结果。结果 在仿真数据的实验中，与对比的去除伪影的方法相比，本文提出的方法获得了更好的效果，本文提出的方案保留了图像更多的组织结构的细节。与ADN算法相比，在金属较小时本文算法的PSNR提高了2.4449，SSIM提高了0.0023，在金属较大时，本文算法的PSNR提高了5.942，SSIM提高了0.0139。在金属有两个时，本文算法PSNR提高8.8388，SSIM提高0.0130。结论 本文提出的金属伪影校正的方法，可以有效去除金属伪影，并且提高图像质量，保留更多的组织结构细节。

关键词: CT金属伪影；投影插值；深度学习；图像转换

Abstract: Objective To propose a framework that combines sinogram interpolation with unsupervised image-to-image translation (UNIT) network to correct metal artifacts in CT images. Methods The initially corrected CT image and the prior image without artifacts, which were considered as different elements in two different domains, were input into the image transformation network to obtain the corrected image. Verification experiments were carried out to assess the effectiveness of the proposed method using the simulation data, and PSNR and SSIM were calculated for quantitative evaluation of the performance of the method. Results The experiment using the simulation data showed that the proposed method achieved better results for improving image quality as compared with other methods, and the corrected images preserved more details and structures. Compared with ADN algorithm, the proposed algorithm improved the PSNR and SSIM by 2.4449 and 0.0023 when the metal was small, by 5.9942 and 8.8388 for images with large metals, and by 8.8388 and 0.0130 when both small and large metals were present, respectively. Conclusion The proposed method for metal artifact correction can effectively remove metal artifacts, improve image quality, and preserve more details and structures on CT images.

Key words: CT metal artifacts; sinogram interpolation; deep learning; image transformation