Superior objective and subjective image quality of deep learning reconstruction for low-dose abdominal CT imaging in comparison with model-based iterative reconstruction and filtered back projection

This study aimed to conduct objective and subjective comparisons of image quality among abdominal computed tomography (CT) reconstructions with deep learning reconstruction (DLR) algorithms, model-based iterative reconstruction (MBIR), and filtered back projection (FBP). Datasets from consecutive patients who underwent low-dose liver CT were retrospectively identified. Images were reconstructed using DLR, MBIR, and FBP. Mean image noise and contrast-to-noise ratio (CNR) were calculated, and noise, artifacts, sharpness, and overall image quality were subjectively assessed. Dunnett's test was used for statistical comparisons. Ninety patients (67 ± 12.7 years; 63 males; mean body mass index [BMI], 25.5 kg/m ) were included. The mean noise in the abdominal aorta and hepatic parenchyma of DLR was lower than that in FBP and MBIR ( < .001). For FBP and MBIR, image noise was significantly higher for obese patients than for those with normal BMI. The CNR for the abdominal aorta and hepatic parenchyma was higher for DLR than for FBP and MBIR ( < .001). MBIR images were subjectively rated as superior to FBP images in terms of noise, artifacts, sharpness, and overall quality ( < .001). DLR images were rated as superior to MBIR images in terms of noise ( < .001) and overall quality ( = .03). Based on objective and subjective comparisons, the image quality of DLR was found to be superior to that of MBIR and FBP on low-dose abdominal CT. DLR was the only method for which image noise was not higher for obese patients than for those with a normal BMI. This study provides previously unavailable information on the properties of DLR systems and their clinical utility.