Automatic hepatic steatosis quantification using low-dose CT with deep learning-based noise reduction and CT fat...

To evaluate the accuracy of CT-derived fat fraction (CDFF) software for quantifying hepatic steatosis at various radiation doses, using MRI-derived proton density fat fraction (MRI-PDFF) as the reference standard, and examines the impact of deep learning (DL)-based noise reduction on CDFF accuracy in low-dose CT (LDCT) scans.

We conducted a retrospective analysis of 125 living liver donor candidates who underwent non-contrast CT and MRI between July 2016 and December 2017. CDFF was measured on full-dose and simulated LDCT scans at 50%, 25%, and 10% radiation doses. DL-based denoising reconstruction (DLDR) was applied to LDCT scans for CDFF recalculation. The accuracy of CDFF was compared with MRI-PDFF using Pearson correlation coefficients and receiver operating characteristic (ROC) curve analysis, focusing on the effects of radiation dose and DLDR.

Of the 125 participants (mean age 38 ± 10 years; 77 males), 29 (23%) had hepatic steatosis (MRI-PDFF ≥5%). Full-dose CDFF showed moderate correlation with MRI-PDFF (r = 0.728; P < .001). Correlation decreased with lower doses (r = 0.684-0.725) but improved with DLDR (r = 0.725-0.736). ROC AUC for diagnosing hepatic steatosis was 0.82 for full-dose CDFF, with similar performance across other doses except 10%.

CDFF accuracy declines at lower radiation doses, but DLDR enhances accuracy, improving alignment with MRI-PDFF, especially at reduced doses.

DLDR significantly enhances the accuracy of CDFF accuracy at lower radiation doses, enabling high diagnostic performance for hepatic steatosis while potentially reducing patient radiation exposure.