Rigid registration of multimodal images based on CycleGAN(PDF)
《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]
- Issue:
- 2021年第2期
- Page:
- 198-203
- Research Field:
- 医学影像物理
- Publishing date:
Info
- Title:
- Rigid registration of multimodal images based on CycleGAN
- Author(s):
- GUO Yi1; WU Xiangyi1; WU Qian2; CHEN Zhi1; XU Xie1; PEI Xi1
- 1. Department of Engineering and Applied Physics, University of Science and Technology of China, Hefei 230026, China 2. Anhui Medical University, Hefei 230032, China
- Keywords:
- Keywords: medical images rigid registration multimodal cycle-consistent generative adversarial networks
- PACS:
- R318
- DOI:
- DOI:10.3969/j.issn.1005-202X.2021.02.013
- Abstract:
- Abstract: Objective To reduce the time of image registration while ensuring the rigid registration accuracy of medical images. Methods Firstly, the training data were standardized and normalized. At the same time, the images were re-sampled and clipped to remove the redundant air. Secondly, threshold method and scanning line method were used to obtain the outline information of the images. Based on cycle-consistent generative adversarial networks (CycleGAN), two generator networks and two discriminator networks were established. The generator networks received multimodal image pairs and output the registration results. The discriminator networks received registration results and output registration accuracy. To improve the convergence speed of CycleGAN, the contour loss was added to restrict the training direction of the network. Results 75 abdominal cases were selected, of which 65 cases were used as training data set and 10 cases were used as test data set. Results of registration were compared with those of the registration software Elastix. The average Dice coefficient of the image pairs before registration was 0.858, that after Elastix registration was 0.926, and that after CycleGAN registration was 0.925. The average registration time of Elastix is 12.1 s. However, CycleGAN took 0.04 s to finish image registration, and the acceleration ratio was 302. Conclusion The proposed method not only ensures the accuracy of image registration, but also greatly reduces the time required for image registration and improves the efficiency of the registration process. In addition, compared with other deep learning methods, this method does not need real registration results and similarity measures.
Last Update: 2021-02-04