3605 - Synthetic Contrast-Enhanced CT Generation Based on Generative Adversarial Networks for Lymph Node Delineation in Nasopharyngeal Carcinoma
Presenter(s)
J. Deng1,2, X. Sun3, G. Q. Zhou2, S. Wu2,4, G. Y. Wang2, X. Jiang2, L. Jia3, W. C. Diao2, W. Zhang5, S. Wenzhao2, and Y. Sun2; 1School of Biomedical Engineering, Southern Medical University, Guangzhou, China, 2Department of Radiation Oncology, Sun Yat-sen University Cancer Center, State Key Laboratory of Oncology in South China, Guangdong Provincial Clinical Research Center for Cancer, Guangdong Key Laboratory of Nasopharyngeal Carcinoma Diagnosis and Therapy, Guangzhou, China, 3United Imaging Research Institute of Innovative Medical Equipment, Shenzhen, China, 4Zhongshan School of Medicine, Sun Yat-sen University, Guangzhou, China, 5Shanghai United Imaging Healthcare Co., Ltd., Shanghai, China
Purpose/Objective(s): Compared with non-enhanced CT (NECT), contrast-enhanced CT (CECT) can more accurately delineate the lymph nodes gross tumor volume (GTVnd) of nasopharyngeal carcinoma (NPC) due to the obvious contrast enhancement of blood vessels. However, contrast agents carry risks of allergic reactions in susceptible patients and additional radiation dose, and only NECT scans are typically performed in adaptive radiotherapy (ART) which pose challenges to the efficiency and accuracy of delineation. This study investigated the feasibility of synthetic contrast-enhanced CT (sCECT) from NECT using generative adversarial networks (GANs) for GTVnd delineation of NPC. Meanwhile, the vision transformer and the perceptual loss were introduced to improve the sCECT.
Materials/Methods: A dataset of NECT/CECT images from 150 patients with NPC was retrospectively collected, and randomly divided into training and testing sets in a 4:1 ratio. We introduced the vision transformer and perceptual loss in the plain CycleGAN to strengthen the learning of global features and emphasize the perceptual quality. The normalized mean absolute error (MAE), peak signal-to-noise ratio (PSNR), and structural similarity (SSIM) were used to evaluate the sCECT-to-CECT image similarity. The radiation oncologists successively referred to NECT only, sCECT and real CECT to delineate the left and right GTVnd on the NECT images, and evaluated the feasibility of applying sCECT in GTVnd delineation using dice similarity coefficient (DSC), 95% Harsdorff distance (HD95) and average surface distance (ASD).
Results: The sCECT images generated by our model showed high similarity to the real CECT images (normalized MAE: 0.008±0.001, SSIM: 0.987±0.007, PSNR: 42.16±2.22 dB). The visual effect of blood vessels contrast enhancement in sCECT was fairly good. The GTVnd delineations with reference to real CECT were used as the reference contours. Compared with reference to NECT only, the DSC, HD95, and ASD of the left and right GTVnd delineations with reference to sCECT were significantly higher (DSC: 0.63±0.09/0.65±0.09 to 0.81±0.04/0.83±0.03, HD95: 10.68±4.11/9.03±4.59 mm to 2.60±0.71/2.14±0.42 mm, ASD: 1.79±0.71/3.11±2.63 mm to 0.60±0.26/0.54±0.15 mm.
Conclusion: The preliminary research results indicate that our proposed model is able to generate high-quality sCECT images and has the potential to assist in the GTVnd delineation for NPC. We found that it is more difficult to differentiate between blood vessels and lymph nodes on NECT with advanced lymph node staging. In the subsequent work, we will supplement the data to further optimize the model and conduct further feasibility studies based on lymph node staging.