RadiolGAN: Multicenter Feasibility Study of Synthetic CT From 3D Ultra-Short Echo Time MRI for Enhanced Pulmonary Radiologic Sign Visualization.
Authors
Affiliations (9)
Affiliations (9)
- Department of Radiology, Northern Jiangsu People's Hospital, Yangzhou, Jiangsu, China.
- College of Biological Sciences and Medical Engineering, Southeast University, Nanjing, Jiangsu, China.
- Department of Radiology, The Affiliated Hospital of Yangzhou University, Yangzhou, Jiangsu, China.
- CT/MRI, Second Affiliated Hospital of Fujian Medical University Shishi Hospital, Quanzhou, Fujian, China.
- Department of Technology, Northern Jiangsu People's Hospital, Yangzhou, Jiangsu, China.
- Department of Radiology, Henan Provincial Chest Hospital, Zhengzhou, Henan, China.
- Department of Emergency, Northern Jiangsu People's Hospital, Yangzhou, Jiangsu, China.
- MR Research, GE Healthcare, Beijing, China.
- College of Automation Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing, Jiangsu, China.
Abstract
Chest CT requires breath-holding and ionizing radiation. 3D ultrashort echo time (UTE) MRI allows radiation-free imaging, but the image quality is suboptimal. To develop RadiolGAN and evaluate synthetic CT (sCT) from 3D UTE MRI for enhanced pulmonary visualization. Prospective multicenter study. Three hundred and fifty-nine subjects (167 women, 192 men; 52 ± 19 years) from four centers: 244 training, 61 internal test, and 54 external test. 3 T, 3D UTE gradient-echo sequence. Signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), noise, peak signal-to-noise ratio (PSNR), mean structural similarity index (MS-SSIM), universal quality index (UQI), and learned perceptual image patch similarity (LPIPS). Three radiologists rated pulmonary structures (bronchi, vessels, fissures, artifacts, diagnostic confidence) and radiologic signs (nodules/masses, ground-glass opacities, patchy shadows/consolidation, emphysema/bullae, bronchiectasis) on a 5-point Likert scale. Repeated-measures ANOVA, paired t-tests, and Friedman tests; p < 0.05 significant. In the external test set, RadiolGAN-CT showed higher SNR (32.63 ± 1.21 vs. 26.07 ± 1.53) and CNR (25.36 ± 1.06 vs. 21.64 ± 1.32), and lower noise (15.74 ± 0.85 vs. 19.66 ± 1.01) than 3D UTE. Versus CycleGAN-CT, RadiolGAN-CT achieved higher PSNR (65.32 ± 0.19 vs. 64.68 ± 0.21), MS-SSIM (0.912 ± 0.004 vs. 0.892 ± 0.004), FSIM (0.808 ± 0.007 vs. 0.783 ± 0.006), and UQI (0.854 ± 0.007 vs. 0.843 ± 0.007), and lower LPIPS (0.221 ± 0.010 vs. 0.236 ± 0.009). No differences were found between RadiolGAN-CT and CycleGAN-CT in SNR (p = 0.612), CNR (p = 0.547), or noise (p = 0.595). Diagnostic confidence was higher for RadiolGAN-CT (3.98 ± 1.09) than CycleGAN-CT (3.59 ± 1.06) and 3D UTE (2.84 ± 1.30). Ground-glass opacity depiction did not differ between RadiolGAN-CT and CycleGAN-CT (p = 0.903). RadiolGAN enables high-fidelity sCT from 3D UTE, improving structural depiction and perceptual similarity. 1. 2.