Application of deep learning image reconstruction algorithm in low-dose abdominal CT / 西安交通大学学报(医学版)

ABSTRACT

【Objective】 To investigate the value of deep learning image reconstruction (DLIR) in improving image quality and reducing beam-hardening artifacts of low-dose abdominal CT. 【Methods】 For this study we prospectively enrolled 26 patients (14 males and 12 females, mean age of 60.35±10.89 years old) who underwent CT urography between October 2019 and June 2020. All the patients underwent conventional-dose unenhanced CT and contrast-enhanced CT in the portal venous phase (noise index of 10; volume computed tomographic dose index 9.61 mGy) and low-dose CT in the excretory phase(noise index of 23; volume computed tomographic dose index 2.95 mGy). CT images in the excretory phase were reconstructed using four algorithms ASiR-V 50%, DLIR-L, DLIR-M, and DLIR-H. Repeated measures ANOVA and Kruskal-Wallis H test were used to compare the quantitative (skewness, noise, SNR, CNR) and qualitative (image quality, noise, beam-hardening artifacts) values among the four image groups. Post hoc comparisons were performed using Bonferroni test. 【Results】 In either quantitative or qualitative evaluation, the SNR, CNR, overall image quality score, and noise of DLIR images were similar or better than ASiR-V 50%. In addition, the SNR, CNR, and overall image quality scores increased as the DLIR weight increased, while the noise decreased. There was no statistically significant difference in the distortion artifacts (P=0.776) and contrast-induced beam-hardening artifacts (P=0.881) scores among these groups. 【Conclusion】 Compared with the ASiR-V 50% algorithm, DLIR algorithm, especially DLIR-M and DLIR-H, can significantly improve the image quality of low-dose abdominal CT, but has limitations in reducing contrast-induced beam-hardening artifacts.