The method of pelvic synthetic CT generation based on the cycle-consistent generative adversarial networks(PDF)
《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]
- Issue:
- 2021年第1期
- Page:
- 21-29
- Research Field:
- 医学影像物理
- Publishing date:
Info
- Title:
- The method of pelvic synthetic CT generation based on the cycle-consistent generative adversarial networks
- Author(s):
- WU Xiangyi1; CAO Feng2; CAO Ruifen3; WU Qian4; DONG Jiangning2; X. George Xu1; PEI Xi1; 5
- 1. School of Physical Sciences, University of Science and Technology of China, Hefei, 230025, China 2. Department of Medical Imaging, the First Affiliated Hospital of University of Science and Technology of China, Hefei 230001, China 3. College of Computer Science and Technology, Anhui University, Hefei 230601, China 4. School of Humanistic Medicine, Anhui Medical University, Hefei 230032, China 5. Anhui Wisdom Technology Co. Ltd, Hefei 230088, China
- Keywords:
- Keywords: synthetic CT unpaired pelvis MRI CycleGAN
- PACS:
- R318；TP391.41
- DOI:
- DOI:10.3969/j.issn.1005-202X.2021.01.005
- Abstract:
- Abstract: Objective Based on the Cycle-Consistent Generative Adversarial Networks (CycleGAN), pelvic data from unpaired patients are applied to achieve interconversion between MRI images and CT images, and the accuracy and dose performance of synthetic CT (sCT) images generated based on this model are assessed. Methods This CycleGAN network includes two generators and two discriminators. Based on fully convolutional networks (FCNs), two generators were constructed first. One converted 2D pelvic MRI images to 2D pelvic sCT images, and the other converted CT images to synthetic MRI (sMRI) images. Based on FCNs, two discriminators were then constructed for discriminating real images and generated synthetic images to facilitate the sCT/sMRI quality improvement. To ensure the consistency between sCT images and MRI images, normalized mutual information was applied as a similarity-constraint loss term to improve the model. The training set consists of T1-weighted pelvic MRI images from 35 patients and pelvic CT images from an additional 36 patients, and the test set consists of pelvic MRI images and CT images from 10 patients. Assessment methods included errors between sCT images and CT images and the radiation dose gamma pass rates. Results For all the cases in the test set, the mean absolute error (MAE) between the generated sCT images and the true CT images was 35.537 (±4.537) HU. The maximum voxel-based mean dose difference was 0.49%. The mean gamma pass rates were above 99%, 98% and 95% with 3%/3 mm, 2%/2 mm, and 1%/1 mm criteria, respectively. Conclusion Accurate pelvic sCT images can be generated by using CycleGAN network and unpaired training data from different patients, and the dose accuracy calculated based on the sCT images meets the clinical requirements.
Last Update: 2021-01-29