Deep learning image reconstruction for optimizing image quality of low-energy spectral monochromatic CT and detecting liver small low-contrast lesions

Objective To investigate the feasibility of deep learning image reconstruction (DLIR) for optimizing image quality of low-energy spectral monochromatic CT and improving detection of liver small low-contrast lesions. Methods Thirty patients with 58 hepatic lesions who underwent upper abdominal portal-venous-phase enhanced CT were enrolled. Monochromatic images with energy levels ranging from 40 to 70 keV (10 keV increment) were reconstructed using DLIR and hybrid model-based adaptive statistical iterative reconstruction V (ASIR-V), respectively. The contrast-to-noise ratio (CNR) of liver, portal vein and hepatic lesions, also image noise were evaluated, the overall image quality, lesion conspicuity and diagnostic confidence were subjectively scored, and the outcomes were compared among different images. Results At the energy levels of 40-70 keV, compared with ASIR-V images, CNRliver, CNRportal veinand CNRhepatic lesionof DLIR images significantly increased (all P<0. 05), while the image noise significantly reduced (all P<0. 05). At the energy levels of 40-60 keV, the overall image quality, lesion conspicuity and diagnostic confidence of DLIR images were higher than those of ASIR-V images (all P<0. 05). Conclusion DLIR technique could reduce noise of low-energy monochromatic images, improve image quality and detectability of liver small low-contrast lesions.

Duke Scholars

Author Daniele Marin Radiology, Abdominal Imaging