Registration-guided GAN Synthesis of Virtual Contrast-enhanced Thoracic CT for Hilar and Mediastinal Lymph Node Detection.

Abstract

Purpose To develop and evaluate a registration-guided generative adversarial network (Reg-GAN) that synthesizes virtual contrast-enhanced CT (vCECT) from noncontrast CT (NCCT) images and to assess its utility in detecting enlarged hilar and mediastinal lymph nodes. Materials and Methods In this retrospective study, data from 700 patients for training, 100 patients for internal testing, and 63 patients for external testing were used to construct Reg-GAN, which aligns synthesized vCECT with true CECT (tCECT) images during training. Image quality was compared among Reg-GAN, Non-Reg-GAN, and a publicly available GAN (Public GAN). Enlarged lymph nodes were defined by a 10-mm short-axis diameter threshold, and their detectability was evaluated. Wilcoxon signed rank and DeLong tests were used. Results In both the internal and external test sets, Reg-GAN-generated vCECT images were most similar to tCECT, as indicated by lower median values of the mean absolute error than Non-Reg-GAN and Public GAN (internal test, 13.60 vs 17.84 and 15.52, respectively, both P < .001; external test, 14.91 vs 19.49 and 16.37, respectively, both P < .001). In the reader study using the external test set, adding vCECT to NCCT improved area under the receiver operating characteristic curve values for detecting enlarged lymph nodes, particularly for the two most experienced radiologists, from 0.81 to 0.90 and from 0.74 to 0.88 (both P < .05). Conclusion Reg-GAN synthesized higher-quality vCECT images from NCCT images than Non-Reg-GAN and Public GAN. Adding vCECT images improved the detectability of enlarged lymph nodes, supporting the potential utility of vCECT in patients contraindicated for contrast agent administration or screening settings. Keywords: Generative Adversarial Network, Virtual Contrast-enhanced CT, Registration-guided GAN, Hilar and Mediastinal Lymph Nodes, Noncontrast CT, Deep Learning Supplemental material is available for this article. © RSNA, 2026.