Author: Jie Hu, Zhengdong Jiang, Nan Li, Tie Lv, Yuqing Xia, Shouping Xu, Gaolong Zhang, Wei Zhao, Changyou Zhong 👨🔬
Affiliation: School of Physics, Beihang University, National Cancer Center/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Radiotherapy Department of Meizhou People’s Hospital (Huangtang Hospital), UT Health San Antonio, National Cancer Center/ National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, 100021, Peopleʼs Republic of China, Department of Radiation Oncology 🌍
Purpose: Patients usually undergo cone-beam computed tomography (CBCT) scans which are used for patient set-up before radiotherapy. However, the low image quality of CBCT hinders its use in adaptive radiation therapy (ART). A network and deformational registration algorithm were developed for CBCT-to-CT conversion and CT-to-CBCT registration to provide high-resolution synthetic CT (sCT) images and contours. The ultimate goal is to design an online ART process with sCT participation.
Methods: A cycle-consistent generative adversarial network (cycleGAN) with attention gates and specific loss functions was proposed for the sCT generation. Due to the excessive differences between CBCT head-neck and shoulder, the two models were trained separately. Twenty-five patients' data were used for network training, and four patients' full sub-CBCT were used for testing. Subsequently, sCTs were rigidly registered to the planning CT using the clinical set-up registration file. CT was flexibly registered to sCT for migration of contours and generating DICOM files.
Results: Our model greatly enhances the image quality of CBCT while preserving the original anatomical structures. The quantitative values of MAE and PSNR were changed from 122.13±28.40 HU and 37.70±0.47 dB to 76.12±18.72 HU and 38.89±0.73 dB, respectively. In the registration, the MAE between CT and sCT changed from 99.64±22.74 HU to 76.08±18.13 HU (CT registered to CBCT) and 61.20±15.34 HU (CT registered to sCT).
Conclusion: We developed an online sCT generation and registration process for ART. For new patients, sCTs can be generated online and registered to the planning CT using the positioning file. Then, the CT will be registered with the sCT to achieve the contour DICOM files. The excellent image quality of sCT demonstrates its potential to enable online ART.