A experimental study of applying deep learning image reconstruction algorithm to reduce radiation dose of dynamic CT myocardial perfusion
10.3760/cma.j.cn112149-20220309-00214
- VernacularTitle:应用深度学习图像重建算法降低心肌动态CT灌注辐射剂量的实验研究
- Author:
Wenlei GENG
1
;
Yang GAO
;
Na ZHAO
;
Hankun YAN
;
Yunqiang AN
;
Liujun JIA
;
Bin LYU
Author Information
1. 中国医学科学院 北京协和医学院 国家心血管病中心 阜外医院放射影像科，北京 100037
- Keywords:
Tomography, X-ray computed;
Myocardial perfusion;
Deep learning
- From:
Chinese Journal of Radiology
2022;56(11):1182-1187
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To investigate the impact on image quality of a new deep learning image reconstruction (DLIR) algorithm in dynamic CT myocardial perfusion imaging (CTP) and to explore whether the algorithm affects the quantification of myocardial blood flow (MBF) in swine.Methods:Dynamic CTP imaging was performed in five anesthetized domestic swine [body weight (58.6±1.9) kg], at both rest and stress state. The tube voltages were fixed at 100 kV, and the low-dose and high-dose scanning tube currents were set as 150 mA and 300 mA, respectively. The low-dose (LD) scan data were reconstructed with filtered back projection (FBP) and three different DLIR strengths (low, medium, and high). High-dose (HD) scan data were reconstructed with filtered back projection (FBP) only. Subjective (5-point scale) image quality was evaluated, and objective evaluations included image noise, signal-to-noise ratio (SNR), and contrast-to-noise ratio (CNR) was performed. Linear regression was used to test the linear trend between DLIR algorithm strength and image quality. Data sets normality was determined by the Shapiro-Wilk test. Comparisons between groups were performed using Student′s t test for normally distributed data or the Wilcoxon rank-sum test for non-normally distributed data. Results:The mean effective radiation dose was 7.2 and 3.8 mSv for the HD protocol and the LD protocol, respectively, with statistically significant difference found between two protocols ( t=282.50, P<0.001). The image noise of the images obtained at LD protocol gradually decreased and the image SNR and CNR gradually increased with DLIR algorithm strength increased ( F=60.10,35.87,41.41; P for trend were all<0.001). As for DLIR-high strength (LD) and FBP (HD) images, the image noise values were (31.7±3.1) and (38.2±1.2) HU; SNR were 16.6±2.0 and 13.8±0.8; CNR were 14.5±1.7, 11.6±0.9, respectively, with significant differences found between two groups ( t=5.70, 4.15, 5.68; all P<0.05). The subjective scores of DLIR-high strength (LD) and FBP (HD) images were significantly different (4.8±0.4 and 4.2±0.6, Z=2.12, P<0.05). No significant differences were found between the MBF calculated from FBP (LD) and from DLIR-high strength (LD), with the values as (81.3±17.3) ml·100 ml -1·min -1 vs. (79.9±18.3)ml·100 ml -1·min -1 at rest state; and (99.4±24.9)ml·100 ml -1·min -1 vs. (100.7±27.3) ml·100 ml -1·min -1 at stress state ( t=1.10, 0.89; P>0.05). Conclusion:DLIR-high strength can improve image quality of myocardial CTP in swine, and can reduce radiation dose without influencing the MBF calculation.
