Synthetic CT generation from CBCT images using self-attention generative adversarial network(PDF)
《中国医学物理学杂志》[ISSN:1005-202X/CN:44-1351/R]
- Issue:
- 2023年第11期
- Page:
- 1356-1361
- Research Field:
- 医学影像物理
- Publishing date:
Info
- Title:
- Synthetic CT generation from CBCT images using self-attention generative adversarial network
- Author(s):
- WU Shuyu1; WANG Wei1; YANG Lu1; YU Hui1; ZHOU Cheng2; MEI Yingjie3
- 1. Department of Radiotherapy, Cancer Center of Guangzhou Medical University, Guangzhou 510095, China 2. Department of Equipment, Liuzhou Peoples Hospital, Liuzhou 545006, China 3. Department of Radiology, Guangdong Provincial Peoples Hospital, Guangzhou 510080, China
- Keywords:
- Keywords: cone-beam CT synthetic CT self-attention mechanism generative adversarial network image-guided radiotherapy
- PACS:
- R318；R811.1
- DOI:
- DOI:10.3969/j.issn.1005-202X.2023.11.007
- Abstract:
- Abstract: Objective To generate synthetic CT (sCT) from cone-beam computed tomography (CBCT) images using the self-attention generative adversarial network (SAGAN) for improving CBCT image quality and enhancing the accuracy of image-guided radiotherapy. Methods The CBCT and planning CT (pCT) were collected from 58 cases of head and neck squamous cell carcinoma. SAGAN was constructed using U-Net architecture and a Markovian discriminator, incorporating a self-attention mechanism. WGAN-GP loss function was utilized for adversarial training to learn the spatial and density mapping characteristics of CBCT modality, which enhances feature expression in CBCT images and improves sCT generation accuracy. Qualitative and quantitative analyses were conducted to compare the sCT generated using SAGAN or Res-Unet (sCTSAGAN or sCTRes-Unet) with the original pCT for verifying model performance. Results The qualitative evaluation showed that both SAGAN and Res-Unet effectively suppressed artifacts, while sCTSAGAN closely resembled pCT, with smaller errors. The quantitative analysis demonstrated that SAGAN outperformed Res-Unet in mean absolute error, structural similarity index and peak signal-to-noise ratio, particularly in soft tissue areas ([P]<0.05). sCTSAGAN had linear regression slopes of 0.956, 0.959 and 0.839 for body, bone tissue and soft tissues, respectively, indicating superior CT number calibration capability as compared with the other methods. Conclusion SAGAN demonstrated superior CT number calibration capability and image generation quality in CBCT-based synthetic CT generation, providing a reliable basis for image-guided radiotherapy and improving treatment accuracy.
Last Update: 2023-11-24