The impact of a novel deep learning reconstruction algorithm on image quality in ultralow-dose CT: a quantitative phantom study.
Authors
Affiliations (3)
Affiliations (3)
- Computed Tomography Business Unit, Neusoft Medical Systems Co., Ltd., Shenyang, China.
- Department of Radiology, Affiliated Hospital of Shaanxi University of Chinese Medicine, Xianyang City, China.
- Department of Radiology, General Hospital of Ningxia Medical University, Yinchuan, China. [email protected].
Abstract
The aim of this study is to evaluate the performance of a novel deep learning image reconstruction (DLIR) algorithm in noise reduction, contrast-to-noise ratio (CNR), and low iodine concentration detection for ultralow-dose computed tomography (CT) imaging. A nine-hole phantom with iodine concentrations (0-40 mg/mL) was scanned at various tube voltages (60-120 kVp). Images were reconstructed using filtered back projection (FBP), iterative reconstruction (IR), and DLIR at different weight levels (10%-90%). Objective metrics (noise, CNR, CT value accuracy via Bland-Altman analysis) and subjective image quality were assessed. At all tube voltages (60-120 kVp), DLIR with medium-to-high weight levels (50%-90%) reduced background noise and increased CNR compared with FBP and IR (p < 0.001). The CNR at a low iodine concentration (1.25 mg/mL) was enhanced, and the DLIR algorithm (weight levels 30%-90%) was able to continuously detect an iodine concentration of 1.25 mg/mL (CNR ≥ 3) at all tube voltages. Under fixed ultralow-dose conditions, DLIR preserved image quality and low-contrast detectability. DLIR (weight levels 90%) reduced background noise by 84.7% compared with FBP and improved CNR (p < 0.001). Bland-Altman analysis confirmed excellent quantitative accuracy for DLIR. The exploratory subjective evaluation was consistent with objective metrics. The DLIR algorithm can enhance image quality in low-dose CT imaging and improve the ability to detect low concentrations of iodine. These findings demonstrate that DLIR maintains image quality and CNR at low iodine concentrations in phantom studies. Clinical implications require further validation. This phantom study shows that the deep learning reconstruction algorithm can still maintain the diagnostic image quality and low-contrast detectability even under ultralow-dose CT (94% dose reduction). These findings support further clinical research to optimize the dosage regimens and potentially reduce the use of iodine contrast agents. Under ultralow-dose conditions (60 kV), DLIR preserved image quality metrics and detectability thresholds in a phantom under ultralow-dose conditions. It significantly suppressed image noise and improved the CNR. The algorithm reliably detected low iodine concentrations (1.25 mg/mL) at all dose levels.