Image quality improvement in low-dose chest CT with deep learning image reconstruction.

Abstract

To investigate the clinical utility of deep learning image reconstruction (DLIR) for improving image quality in low-dose chest CT in comparison with 40% adaptive statistical iterative reconstruction-Veo (ASiR-V40%) algorithm. This retrospective study included 86 patients who underwent low-dose CT for lung cancer screening. Images were reconstructed with ASiR-V40% and DLIR at low (DLIR-L), medium (DLIR-M), and high (DLIR-H) levels. CT value and standard deviation of lung tissue, erector spinae muscles, aorta, and fat were measured and compared across the four reconstructions. Subjective image quality was evaluated by two blind readers from three aspects: image noise, artifact, and visualization of small structures. The effective dose was 1.03±0.36mSv. There was no significant difference in CT values of erector spinae muscles and aorta, whereas the maximum difference for lung tissue and fat was less than 5HU among the four reconstructions. Compared with ASiR-V40%, the DLIR-L, DLIR-M, and DLIR-H reconstructions reduced the noise in aorta by 11.44%, 33.03%, and 56.1%, respectively, and had significantly higher subjective quality scores in image artifacts (all p <0.001). ASiR-V40%, DLIR-L, and DLIR-M had equivalent score in visualizing small structures (all p >0.05), whereas DLIR-H had slightly lower score. Compared with ASiR-V40%, DLIR significantly reduces image noise in low-dose chest CT. DLIR strength is important and should be adjusted for different diagnostic needs in clinical application.