3695 - Synthetic Contrast-Enhanced CT Generated by Generative Adversarial Networks to Assist Lymph Node Delineation in Lung Cancer Radiotherapy
Presenter(s)
X. Zhong1, X. Sun2, J. Zhang3, B. Lv1, L. Jia4, W. Zhang5, and L. Wang6; 1Department of Radiation Oncology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Jinan, China, 2United Imaging Research Institute of Innovative Medical Equipment, Shenzhen, China, 3Department of Radiation Oncology, Shandong Cancer Hospital Affiliated to Shandong University, Jinan, China, 4Shenzhen United Imaging Healthcare Co., Ltd., Shenzhen, Guangdong, China, 5Shanghai United Imaging Healthcare Co., Ltd., Shanghai, China, 6Department of Radiation Oncology, Shandong Cancer Hospital and Institute, Shandong First Medical University and Shandong Academy of Medical Sciences, Jinan, China, Jinan, Shandong, China
Purpose/Objective(s):
Accurate lymph node delineation in lung cancer radiotherapy relies on contrast-enhanced CT (CECT). However, the use of contrast media is limited by patient-specific factors (e.g., allergic reactions and nephrotoxicity) and clinical scenarios such as online adaptive radiotherapy, while also contributing to increased costs and radiation exposure. This study proposes a multi-task generation model to generate high-quality synthetic CECT from non-contrast CT images, aiming to support lymph node delineation and reduce contrast agent dependence.Materials/Methods:
This study used 196 chest NCCT/CECT scans (2020-2022) of lung and esophageal cancer patients for model training, with independent validation on 20 lung cancer cases (2024-2025). A CycleGAN-based dual-task framework was developed for simultaneous segmentation and generation, incorporating Transformer modules in the generator bottleneck, and focusing on generation quality within delineated regions (e.g., aorta) by up-weighting their loss. The quantitative evaluation of the synthetic CECT images was conducted using image quality assessment metrics such as mean absolute error (MAE), structural similarity (SSIM), and peak signal-to-noise ratio (PSNR). Additionally, a clinical validation was performed by radiation oncologists through a Turing test and an evaluation of IASLC lymph node station selection for radiotherapy target delineation, with real and synthetic CECT images randomly intermixed.Results:
The synthetic CECT images generated by the model exhibited a close similarity to real CECT images, with an MAE of 23.91 ± 6.67 HU, an SSIM of 0.93 ± 0.03, and a PSNR of 30.68 ± 2.80 dB on the independent test set. The Turing test results showed that 55% of the synthetic CECT images perceived as genuine, 35% as pseudo-enhanced, and the remaining 10% were inconclusive by two radiation oncologists. As presented in Table 1, the synthetic CECT images achieved high performance in the target lymph node station delineation, with a sensitivity of 0.903 ± 0.122, a specificity of 0.952 ± 0.088, and an accuracy of 0.928 ± 0.080.Conclusion:
Preliminary experimental results indicate that our model generates synthetic CECT images of relatively high quality, with over 60% of these images being indistinguishable from real ones by radiation oncologists. The high sensitivity and specificity in lymph node evaluation demonstrate the significant potential of synthetic CECT images as a substitute for CECT scans in radiotherapy. In the next phase, we plan to expand the dataset size and diversity to further optimize the generation quality of vascular details and validate the feasibility of synthetic CECT images in clinical radiotherapy.