Effect of different reconstruction algorithms for ultra-low dose chest CT on quantitative detection of pneumonia and image quality
Objective To investigate the effect of different image reconstruction algorithms,including filtered back projection,adaptive statistical iterative reconstruction V and deep learning image reconstruction(DLIR)on AI-assisted quantitative analysis of pneumonia(uAI-Discover-NCP)and image quality under ultra-low dose chest CT scanning conditions.Methods Fourty-three patients undergoing follow-up for pneumonia from July to December 2023 at the Affiliated Hospital of Shaanxi University of Chinese Medicine were included in this study.Each patient underwent personalized ultra-low dose CT scanning.The raw data were reconstructed using filtered back projection,40%intensity adaptive statistical iterative reconstruction V,various intensities of DLIR(DLIR-M,DLIR-H),and DLIR-H with additional E2 edge enhancement(DLIR-H+E2),resulting in five groups of images.Measurements were taken for CT values and noise values in the ROIs-air,lung tissue,thoracic aorta,left subscapularis muscle,and thoracic vertebra 10.The signal-to-noise ratio was calculated.Two physicians subjectively rated the overall quality and pneumonia presentation of the five groups of images on a 5-point scale.The images were independently analyzed using CT image-assisted pneumonia quantitative analysis software,recording pneumonia index,pneumonia volume and its percentage,and pneumonia quality and its percentage.Repeated measures ANOVA or Friedman's rank-sum test were used to compare quantitative parameters and subjective scores among groups.Results There was no significant difference in CT values of the lung parenchyma,thoracic aorta,left subscapularis muscle,and thoracic vertebra 10 across the five image groups(P>0.05).However,differences in noise values and signal-to-noise ratio among tissues were statistically significant(P<0.05),with the DLIR-H group demonstrating the lowest image noise and highest signal-to-noise ratio,significantly outperforming the others four groups(P<0.05).The consistency of subjective scoring by the two physicians was high(Kappa=0.811-0.894).There was a significant difference in overall image quality and pneumonia presentation scores across the five groups(P<0.001),with the highest scores in the DLIR-H and DLIR-H+E2 groups,although the difference between these groups was not significant(P>0.05).There was no significant difference in the overall variation of the five pneumonia quantitative parameters among the groups(P>0.05).Conclusion Under ultra-low dose scanning conditions,AI-assisted quantitative analysis of pneumonia is not affected by the reconstruction algorithm.Compared to filtered back projection and 40%adaptive statistical iterative reconstruction V,high-intensity DLIR significantly reduces image noise and noticeably improves image quality,offering substantial clinical diagnostic advantages.
deep learning image reconstructionimage qualitypneumoniaartificial intelligencebody lever photography
万方数据
维普
