Diagnostic value of a second-generation super-resolution deep learning-based reconstruction combined with a metal artifact reduction algorithm for pelvic CT.
Authors
Affiliations (2)
Affiliations (2)
- Department of Radiology, The University of Tokyo Hospital, 7-3-1 Hongo, Bunkyo-Ku, Tokyo, 113-8655, Japan.
- Department of Radiology, The University of Tokyo Hospital, 7-3-1 Hongo, Bunkyo-Ku, Tokyo, 113-8655, Japan. [email protected].
Abstract
To evaluate the impact of combining a second-generation super-resolution deep learning-based reconstruction (DLR2) with a metal artifact reduction (MAR) algorithm in CT images of patients with metal hip implants by assessing both quantitative metrics and qualitative reader ratings. This retrospective study included 40 patients (30 females; age range, 54-93 years) with metal hip implants. Images were reconstructed using DLR2, a first-generation DLR (DLR1), and a conventional hybrid iterative reconstruction (HIR), each combined with MAR. Images without MAR were also reconstructed using DLR2 (DLR2-only). Standard deviations (SDs) of attenuation in the regions of interest (ROI) over the bladder, gluteus maximus muscle, and gluteal fat were recorded. Artifact indices for muscle (AI<sub>muscle</sub>) and fat (AI<sub>fat</sub>) were also calculated. Three radiologists independently assessed the depiction of pelvic structures (femoral artery, bladder, rectum, uterus/prostate), artifact reduction, and diagnostic usability with 5-point scores. For statistical analysis, the Wilcoxon signed-rank test was used with the Holm correction for multiple comparisons. DLR2 + MAR showed significantly lower SDs than DLR1 + MAR, HIR + MAR, and DLR2-only across all three ROIs (p < 0.01-0.02). For both AI<sub>muscle</sub> and AI<sub>fat</sub>, DLR2 + MAR had the lowest values, though differences with DLR1 + MAR were not significant (p = 0.80, 0.11). DLR2 + MAR showed significantly higher scores for the depiction of the femoral artery and rectum compared with other + MAR reconstruction methods (p < 0.01-0.04), with no significant differences in the remaining categories (p = 0.08-0.91). Inter-rater agreement ranged from 0.67 to 0.77, indicating substantial agreement. Combining DLR2 and MAR improves image quality and visualization of pelvic structures.