Mutual information-based contrastive learning for the generation of pseudo-CT images of the head from magnetic resonance imaging
10.3760/cma.j.cn112271-20210810-00318
- VernacularTitle:一种基于互信息对比学习由磁共振成像生成头部伪CT图像的方法
- Author:
Jiangtao WANG
1
;
Xinhong WU
;
Bing YAN
;
Lei ZHU
;
Yidong YANG
Author Information
1. 中国科学技术大学工程与应用物理系,合肥 230026
- Keywords:
Magnetic resonance imaging;
Contrastive learning;
Pseudo-CT
- From:
Chinese Journal of Radiological Medicine and Protection
2022;42(2):95-102
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To compare the abilities of different neural networks to generate pseudo-computed tomography (CT) images from magnetic resonance imaging (MRI) images and to explore the feasibility of pseudo-CT for clinical radiotherapy planning.Methods:A total of 29 brain cancer patients with planning CT and diagnostic MRI were selected. 23 of these patients were used for training neural networks and 6 for testing pseudo-CT images. Cycle-consistent generative adversarial network (cycleGAN), contrastive learning for unpaired image-to-image translation (CUT), and improved network denseCUT proposed in this study were applied to generate pseudo-CT images from MRI images. The pseudo-CT images were imported into a clinical treatment planning system to verify the feasibility of applying this method to radiotherapy planning.Results:The comparison between the generated pseudo-CT images and real CT images showed that the mean absolute errors were (72.0±6.9), (72.5±8.0), and (64.6±7.3) HU for the cycleGAN, CUT, and denseCUT, respectively. Meanwhile, the structure similarity indices were 0.91±0.01, 0.91±0.01, and 0.93±0.01, respectively. The peak signal-to-noise ratios were (28.5±0.7), (28.5±0.7), and (29.5±0.7) dB, respectively. The 2%/2 mm γ passing rates were 98.05%, 97.92%, and 98.31% for the cycleGAN, CUT, and denseCUT, respectively.Conclusions:DenseCUT can generate more accurate pseudo-CT images and the pseudo-CT can meet the demand for the dose calculation of IMRT plan.