ESTRO 2025 - Abstract Book

S2984

Physics - Image acquisition and processing

ESTRO 2025

1758

Digital Poster Deep Learning-Based Synthetic CT Generation for MR-Only Radiotherapy in Nasopharyngeal Carcinoma Guanqun Zhou 1,2 , Hua Li 3 , Yuxian Yang 1,2 , Yuxi Xiong 1,2 , Ziquan Wei 3 , Lecheng Jia 3 , Guangyu Wang 1,2 , Xiaobo Jiang 1,2 , Feng Chi 1,2 , Ying Sun 1,2 1 Department of Radiation Oncology, Sun Yat-sen University Cancer Center, Guangzhou, China. 2 Collaborative Innovation Center for Cancer Medicine, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, State Key Laboratory of Oncology in South China, Guangzhou, China. 3 Research cooperation department, Shenzhen United Imaging Research Institute of Innovative Medical Equipment, Shenzhen, China Purpose/Objective: The accuracy of synthetic computed tomography (sCT) generated from Magnetic resonance (MR) images and the precision of treatment planning using sCT are central to the success of MR-only radiotherapy. This study aims to evaluate the accuracy of sCT images generated via contrastive learning generative adversarial networks (CLGANs) in terms of image quality and treatment planning effectiveness for nasopharyngeal carcinoma (NPC) radiotherapy. Material/Methods: We collected T1-weighted MR images and pair CT images of 320 NPC patients in the same position. This study introduces a generative adversarial network model that incorporates multi-window width normalization loss and contrastive learning loss to enhance the accuracy of sCT images from MR images for NPC. We evaluated the differences between the sCT and the pair CT using mean absolute error (MAE), peak signal-to-noise ratio (PSNR), and structural similarity (SSIM). Additionally, T1-weighted images and planning CT images (rCT) from the first treatment session of 9 NPC patients undergoing All-in-one treatment on a CT-Linac were collected. We delineated targets and organs at risk (OARs) on the sCT, assessing its image accuracy by comparing the HU values and Dice between the delineations on sCT and rCT. Subsequently, we created a treatment plan on the sCT and mapped it onto the rCT as an IGRT plan, comparing target coverage and OAR doses between the IGRT plan and the original plan on the rCT. We assessed the feasibility of using sCT instead of planning CT in radiotherapy based on the rate of meeting criteria (number of patients meeting criteria/total number of patients). Results: Our model outperformed the plain CycleGAN in terms of image accuracy on 30 test datasets, where MAE decreased from 15.738 to 10.397, PSNR increased from 37.297 to 38.180, and SSIM rose from 0.981 to 0.996. The synthesis results can be shown in Figure 1. In the 9 All-in-one NPC datasets, the MAE of sCT in body, bone, and soft tissue based on HU values were 75.48, 212.25, and 36.19, with dice coefficients of 0.97, 0.79, and 0.93, respectively. On rCT, there were no significant differences between the IGRT plan and the original plan in terms of target coverage and OAR doses on average, with a compliance rate of 100% for all metrics, as detailed in Table 1.