Motivation: The application of deep learning (DL) to chemical shift-encoded MRI (CSE-MRI) to generate proton density fat fraction (PDFF) maps has the potential to reduce scan time, but its feasibility and quantitative consistency have yet to be evaluated. Goal(s): We aimed to analyze the linearity of DL-accelerated CSE-PDFF (DL-PDFF) versus standard PDFF (STD-PDFF) and MR spectroscopy-derived PDFF (MRS-PDFF). Approach: Measurements from automated circular ROI, volumetric ROI, and manual circular ROI from 9 liver segments were compared using linear regression analysis and Bland-Altman plots. Results: DL-PDFF showed excellent linearity and low bias compared to STD and MRS-PDFF, while reducing the scan time to 10 seconds. Impact: DL-accelerated CSE-PDFF showed near perfect linearity with conventional PDFF and MRS-PDFF while reducing the scan time to 10 s, highlighting its role in patients with limited breath-hold capacity and in screening settings where time efficiency is a priority.
Kim et al. (Tue,) studied this question.