Author: Jie Hu, Nan Li, Chuanbin Xie, Shouping Xu, Xinlei Xu, Gaolong Zhang, Zhilei Zhang ๐จโ๐ฌ
Affiliation: School of Physics, Beihang University, National Cancer Center/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Department of Radiation Oncology, the First Medical Center of the People's Liberation Army General Hospital, National Cancer Center/ National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, 100021, Peopleสผs Republic of China, Department of Radiation Oncology, School of Physics, Beihang University, Beijing, 102206, Peopleสผs Republic of China ๐
Purpose: This study aims to synthesize CT images for MRI-only radiation therapy using a deep learning approach that integrates information from the T1- and T2-weighted MRI sequence.
Methods: 97 head and neck cancer patients were enrolled in this study. T1-,T2-weighted MRI (3.0T, UIH, 512x440-pixel, 0.6mm thickness), and no-contrast enhancement CT (SIEMENS, 512x440-pixel, 0.6mm thickness) were collected for each patient,. We proposed a dual-channel cycle generative adversarial network (DUC-CycleGAN) comprising a PatchGAN discriminator and a ResNet_9blocks generator tailored for the T1 and T2 MRI sequences. A joint loss combining structural similarity index measure (SSIM) and perceptual loss was employed. The SSIM loss enforced anatomical structure preservation. In contrast, the perceptual loss based on VGG19 ensured high-level feature regularization and preservation of fine-grained details. The dataset was divided into 90 training cases(31,496 slices) and 7 testing cases(2,464 slices). All training and testing procedures were conducted on a 3090 GPU server. The original kVCT-based treatment plans were transferred to the synthesized CT for dose recalculation to evaluate the clinical feasibility. Image quality and clinically relevant dosimetric metrics, including mean absolute error (MAE), SSIM, peak signal-to-noise ratio (PSNR) and gamma passing rates (GPRs), were assessed for evaluation.
Results: For the 7 test cases, the synthesized CT images achieve an average PSNR of 36.678 dB, an average SSIM of 0.982, an average MSE of 0.701HU, and an MAE of 16.109HU. The GPRs were 99.99%ยฑ0.01% (3mm/3%), 99.72%ยฑ0.06% (2mm/2%), and 95.56%ยฑ0.15%(1mm/1%).
Conclusion: We successfully synthesized CT images from unpaired MR-to-CT datasets using a deep-learning approach, demonstrating excellent image quality and dose accuracy. This work represents a significant step forward in enabling the MRI-only radiation therapy.