Feasibility study on deep learning image reconstruction algorithm to improve the quality of low-dose CT images of the brain
10.3760/cma.j.cn112271-20230602-00174
- VernacularTitle:深度学习重建算法改善颅脑低剂量CT图像质量的可行性研究
- Author:
Jinjin CUI
1
;
Guanzhong LIU
;
Xinghe HU
;
Shaojun HAN
;
Hong SUN
;
Xinjiang WANG
;
Hongxiang YAO
Author Information
1. 国家老年疾病临床医学研究中心 解放军总医院第二医学中心放射诊断科，北京 100853
- Keywords:
Deep learning image reconstruction;
Adaptive statistical iterative reconstruction-V;
Low radiation dose;
Image quality;
Lacunar infarction
- From:
Chinese Journal of Radiological Medicine and Protection
2023;43(9):736-740
- CountryChina
- Language:Chinese
-
Abstract:
Objective:To explore the effectiveness of deep learning image reconstruction (DLIR) algorithm compared to adaptive statistical iterative reconstruction (ASIR-V) algorithm in improving the quality of low-dose brain CT images.Methods:Retrospective inclusion of patients who underwent brain CT examination in the People's Liberation Army General Hospital from November 2021 to August 2022. Four different algorithms were used to reconstruct low-dose CT scans of all patients to obtain 30% intensity ASIR-V (ASIR-V-30%) images, low intensity DLIR (DLIR-L) images, medium intensity DLIR (DLIR-M) images, and high intensity DLIR (DLIR-H) images. The regions of interest were selected from four sets of images, including superficial white matter, superficial gray matter, deep white matter, and deep gray matter, and their CT values and standard deviations were measured for calculating signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR).Subjective evaluation of image quality was conducted by three neuroimaging physicians based on the Likert 5-component scale. The objective and subjective scores of the 4 groups of images were analyzed using ANOVA or Kruskal Wallis. If there are overall differences, pairwise comparisons were conducted within the group.Results:A total of 109 patients were enrolled, including 104 males and 5 females, aged 65-110 years (89.16 ± 9.53) years. The radiation exposure of brain CT low-dose scanning was (0.93 ± 0.01)mSv, significantly lower than that of conventional scanning (2.92 ± 0.01) mSv ( t = 56.15, P < 0.05). The differences in objective image quality analysis of ASIR-V-30%, DLIR-L, DLIR-M, and DLIR-H images of low-dose CT in SNR deep gray matter, SNR deep white matter, SNR superficial gray matter, SNR superficial white matter, CNR deep gray white matter, and CNR superficial gray white matter were statistically significant( F =98.23, 72.95, 68.43, 58.24, 241.13, 289.91, P < 0.05). Among them, DLIR-H images had the lowest noise in deep gray matter, deep white matter, superficial gray matter, and superficial white matter, and had statistically significant differences compared to other image groups ( t = 167.43, 275.46, 182.32, 361.54, P < 0.05). The subjective score of DLIR-H image quality was superior to ASIR-V-30%, DLIR-L, and DLIR-M, with the statistically significant difference ( t = 7.25, 8.32, 9.63, P < 0.05). Conclusions:Compared with ASIR-V, DLIR algorithm can effectively reduce image noise and artifacts in low-dose brain CT, and improve SNR and CNR. The subjective and objective image quality evaluation of DLIR-H is the best.
