Application research of deep learning image reconstruction algorithm in CT portal venography
10.3760/cma.j.cn112149-20231106-00351
- VernacularTitle:深度学习图像重建算法在门静脉CT血管成像中的应用研究
- Author:
Yue ZHANG
1
;
Wei JIANG
;
Yuefei GUO
;
Zhuoxin GUO
;
Ke ZHANG
;
Kun MA
;
Zhan'ao MENG
Author Information
1. 中山大学附属第三医院放射科，广州 510630
- Keywords:
Tomography, X-ray computed;
Adaptive statistical iterative reconstruction-Veo;
Deep learning image reconstruction;
CT portal venography;
Radiation dose
- From:
Chinese Journal of Radiology
2024;58(6):653-660
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To explore the feasibility of deep learning image reconstruction (DLIR) in reducing radiation dose and improving image quality in 100 kV portal vein CT angiography (CTPV), and compare the image quality with traditional 120 kV CTPV.Methods:Consecutive 100 patients who underwent upper abdominal CTPV examinations at the Third Affiliated Hospital of Sun Yat-sen University from June 2021 to December 2022 were prospectively enrolled in this study. They were divided into two groups: the standard dose group (S group) using 120 kV and the low dose group (L group) using 100 kV. In the S group, adaptive statistical iterative reconstruction-Veo (ASIR-V) was employed with weights of 40%, 60%, and 80% (S-AV40, S-AV60, and S-AV80, respectively), along with a medium-level DLIR (S-DM). The L group used two weights of 60% and 80% of ASIR-V (L-AV60 and L-AV80) and medium to high-intensity DLIR (L-DM and L-DH) for reconstruction. Objective evaluation measurements, including CT values and standard deviation (SD) of the main portal vein (MPV), right portal vein (RPV), left portal vein (LPV), hepatic parenchyma, and right vertical muscle, were independently measured by two radiologists. The signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) were also calculated. The images of eight groups were blindly evaluated using a 5 points scale by three radiologists. The Volume CT dose index (CTDIvol) and dose length product (DLP) were recorded for both protocol groups, and the effective dose (ED) was calculated. Statistical analysis was performed using Single factor variance analysis for objective indicators and Kruskal-Wallis H test for subjective scores. Results:The effective dose in the L group was reduced by 51% compared to the S group. There were significant statistical differences in SD, SNR and CNR among eight groups for MPV, RPV, and LPV ( P<0.001 for all). The SD of L-DH showed no statistical difference compared to the S-AV80 group ( P>0.05), but it was significantly lower than the other six groups ( P<0.05 for all). Except for CNR in the MPV, which showed no statistically significant difference between the L-DH and S-AV80 groups ( P>0.05), both SNR and CNR were significantly higher in the L-DH group compared to the other groups ( P<0.001 for all). The L-DH (4.61±0.16) achieved the highest subjective image quality score, which was statistically higher than the scores in the other six groups ( P<0.001 for all) except for the S-DM group (4.31±0.19). There was good consistency among the radiologists regarding objective parameter measurements and subjective image quality scores ( ICC=0.584-0.960). Conclusion:Compared to standard-dose CTPV at 120 kV, the combination of 100 kV with the DLIR algorithm significantly reduced the radiation dose by 51% in CTPV while maintaining higher SNR, CNR and subjective scores. The high-level DLIR algorithm produced the best image quality for 100 kV CTPV.
