Pulmonary emphysema quantification at low dose chest CT using Deep Learning image reconstruction

Purpose: Quantitative analysis of emphysema volume is affected by the radiation dose and the CT reconstruction technique. We aim to evaluate the influence of a commercially available deep learning image reconstruction algorithm (DLIR) on the quantification of pulmonary emphysema in low-dose chest CT. Methods: We performed a retrospective study of low dose chest CT scans in 54 patients with chronic obstructive pulmonary disease (COPD). Raw data were reconstructed using FBP, iterative reconstruction (ASIR-V 70%) and deep learning based algorithms at high, medium and low-strength (DLIR-H,-M,-L). Filtered FBP images served as reference. Pulmonary emphysema volume (proportion of voxels below-950 UH) was measured on each reconstruction dataset and visually assessed by a chest radiologist. Quantitative image quality was assessed by placing 3 regions of interest in the trachea, in air and in a paraspinal muscle. Signal to noise ratio was also measured. Results: The mean CDTIvol was 2.38 +/- 0.68 mGy. Significant differences in emphysema volumes between the filtered FBP reference and ASIR-V, DLIR-H, DLIR-M or DLIR-L were observed, (p < 10(-3)) for all. A strong correlation between filtered FBP volumes and DLIR-H was reported (r = 0.999, p < 10(-4)), a 10% overestimation with DLIR-H being observed. Noise was significantly reduced in DLIR-H volumes compared to the other reconstruction methods. Signal to noise ratio was improved when using DLIR-H (p < 10-6). Conclusion: There are significant differences regarding emphysema volumes between FBP, iterative reconstruction or deep learning-based DLIR algorithm. DLIR-H shows the closest correlation to filtered FBP while increasing SNR.