Deep learning image reconstruction for 50-keV virtual monoenergetic dual-energy CT of the thyroid: a prospective "dual-low" dose study.
Authors
Affiliations (8)
Affiliations (8)
- Department of Radiology, The Affiliated Hospital of Xuzhou Medical University, Xuzhou, China.
- School of Medical Imaging, Xuzhou Medical University, Xuzhou, China.
- Jiangsu Provincial Engineering Research Center for Medical Imaging and Digital Medicine, Xuzhou, China.
- Department of Breast and Thyroid Surgery, The Affiliated Hospital of Xuzhou Medical University, Xuzhou, China.
- CT Imaging Research Center, GE HealthCare China, Shanghai, China.
- Department of Radiology, The Affiliated Hospital of Xuzhou Medical University, Xuzhou, China. [email protected].
- School of Medical Imaging, Xuzhou Medical University, Xuzhou, China. [email protected].
- Jiangsu Provincial Engineering Research Center for Medical Imaging and Digital Medicine, Xuzhou, China. [email protected].
Abstract
Dual-energy CT (DECT) at 50 keV increases iodine attenuation but exponentially amplifies image noise. The feasibility of using deep learning image reconstruction (DLIR) to counteract this noise under a "dual-low" (low-radiation and low-contrast medium) thyroid CT protocol remains underexplored. To investigate the performance of a dual-low DECT protocol combined with DLIR in contrast-enhanced thyroid CT compared with a standard-dose protocol using adaptive statistical iterative reconstruction-Veo (ASIR-V). In this prospective study (August-December 2025), patients were randomly assigned to a standard-dose group (120 kVp, 1.0 mL/kg iodine, ASIR-V 50%) or a dual-low dose group (DECT, 0.6 mL/kg iodine). Dual-low spectral data were reconstructed into 50-keV virtual monoenergetic images using ASIR-V 50%, low-strength DLIR (DLIR-L), and high-strength DLIR (DLIR-H). Objective metrics (CT attenuation, image noise, contrast-to-noise ratio, edge rise slope (ERS), noise power spectrum) and subjective 5-point Likert scores were compared using independent-samples t-tests, paired t-tests, Mann-Whitney U tests, and Wilcoxon signed-rank tests. Sixty-four patients (mean age, 48.6 years ± 12.2; 49 women) were evaluated (32 per group). The dual-low group achieved a 61% reduction in effective radiation dose (0.36 mSv ± 0.05 vs 0.93 mSv ± 0.22; p < 0.001) and a 20% reduction in iodine intake (11.7 g ± 1.8 vs 14.6 g ± 4.0; p < 0.001). Despite these reductions, DLIR-H demonstrated significantly lower image noise (10.9 HU ± 2.0 vs 15.3 HU ± 2.5; p < 0.001) and steeper ERS than the standard-dose protocol. DLIR-H preserved a natural noise texture comparable to ASIR-V. Subjectively, DLIR-H scored near-perfectly in overall image quality (5.0 ± 0.2), significantly outperforming the standard protocol (4.6 ± 0.5; p < 0.001). Combining 50-keV virtual monoenergetic imaging and DLIR-H facilitates a dual-low scanning strategy for thyroid CT, yielding superior objective and subjective image quality while substantially reducing radiation and iodine burdens. Question How can we counteract the severe noise of 50-keV dual-energy CT to enable a low-radiation, low-contrast scanning protocol for thyroid imaging? Findings Deep learning reconstruction at 50-keV significantly suppressed noise, maintaining edge sharpness and diagnostic confidence while reducing radiation by 61% and iodine by 20%. Clinical relevance This dual-low strategy provides a safer imaging alternative that preserves diagnostic quality. It is highly beneficial for vulnerable patients requiring lifelong CT surveillance or those with borderline renal function amid global contrast media shortages.