Effect of deep learning reconstruction on arm-induced artifacts compared with hybrid iterative reconstruction and filtered-backprojection in abdominal CT.

Abdominal computed tomography (CT) is normally performed with patients raising their arms over abdominal region to prevent arm-induced artifacts that degrade image quality. This study aimed to evaluate the effects of deep learning-based image reconstruction (DLIR) on arm-induced artifacts and image quality in abdominal CT with arms-down positioning, compared to adaptive statistical iterative reconstruction-Veo (ASIR-V) and filtered-backprojection (FBP). A liver nodule phantom with arms from a PBU-60 phantom was scanned in three arms-down positions: alongside the torso, across the abdomen, and crossed over the pelvis. Abdominal CT images of 10 patients in arms-alongside-torso position were also included. Images were reconstructed using DLIRs (L-low, M-medium, and H-high), ASIR-Vs (50% and 100%), and FBP. Phantom images were assessed for artifact strength (location parameter of the Gumbel distribution and standard deviation), signal-to-noise ratio, and contrast-to-noise ratio. Two radiologists qualitatively evaluated patient images for noise, artifacts, sharpness, and overall quality. DLIR-H significantly reduced streak artifacts by 37% in location parameters and by 43% in SD, while improving SNR by 28% and CNR by 29% compared to ASIR-V50%. DLIR-M performed significantly better than ASIR-V50% in all quantitative metrics, except in the arms-alongside-torso position. FBP performed worst, although sharpness was comparable. DLIR-H received the best qualitative scores (low noise and artifacts, minimal blurring, and excellent overall image quality), although ASIR-V100% had lower subjective noise. DLIR outperformed ASIR-V and FBP in arm-induced artifact reduction and image quality and is a preferable reconstruction method for arms-down abdominal CT.