Improved virtual non-contrast images in liver dual-energy CT with deep learning algorithm for image reconstruction.
Authors
Affiliations (3)
Affiliations (3)
- Affiliated Hospital of Shaanxi University of Chinese Medicine, 712000, Xianyang, Shaanxi, China.
- Shaanxi University of Chinese Medicine, 712000, Xianyang, Shaanxi, China.
- Affiliated Hospital of Shaanxi University of Chinese Medicine, 712000, Xianyang, Shaanxi, China. [email protected].
Abstract
We aimed to explore the use of deep learning image reconstruction (DLIR) to improve virtual non-contrast (VNC) images from enhanced dual-energy computed tomography (DECT). A total of 91 patients undergoing treatment for liver lesions from September to December 2024 were analyzed. All patients underwent true non-contrast (TNC) and dual-phase enhanced DECT with medium-level DLIR (DLIR-M). The VNC images were generated from arterial phase (VNC<sub>A</sub>) and venous phase (VNC<sub>V</sub>) scans. The CT values and standard deviations of liver, spleen, bilateral erector spinae, and liver lesions were measured across the three image groups, and the signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) were calculated. Two radiologists subjectively scored the image quality. Measurements and quality scores were statistically compared among the groups. There was no difference in CT values among the three groups (P > 0.05); VNC<sub>A</sub> and VNC<sub>V</sub> images did not alter the primary imaging characteristics of hypervascular liver lesions. The VNC<sub>A</sub> and VNC<sub>V</sub> images had lower image noise than the TNC images, with VNC<sub>V</sub> images having the lowest noise. Except for the spleen, VNC<sub>A</sub> and VNC<sub>V</sub> images had higher SNR than TNC images, with VNC<sub>V</sub> images having the best SNR. The subjective scores of the TNC, VNC<sub>A</sub>, and VNC<sub>V</sub> images were all greater than 3, showing good consistency between the two observers (kappa = 0.78, 0.76, and 0.83, respectively).There was no difference between the subjective evaluation of VNC<sub>A</sub> and TNC (P > 0.05), while VNC<sub>V</sub> had better image quality than TNC (P < 0.05). The total effective radiation dose was 8.55 ± 0.38 mSv; thus, replacing TNC with VNC would reduce the dose by 33.4%. Virtual non-contrast images generated in liver DECT with DLIR provide similar CT values to TNC images. The VNC<sub>V</sub> images provide lower image noise as well as higher SNR and subjective scores than TNC and are recommended for clinical use. Virtual non-contrast images with DLIR generated from DECT can improve image quality and reduce radiation dose.