Deep Learning-Driven Estimation of Hepatic Proton Density Fat Fraction from Routine MRI Sequences as a Substitute for Quantitative Dixon Imaging

Motivation: Quantitative Dixon imaging (qDixon) provides precise Proton Density Fat Fraction (PDFF) and R2* maps for accurate liver fat quantification. However, it is not widely accessible for the prolonged scan time. Goal(s): To accurately estimate PDFF with deep learning from routine MRI sequences, such as two-point eDixon and DWI (b=800). Approach: TotalSegmentator was used for liver segmentation and a U-Net model was trained on eDixon and DWI images to predict PDFF values. Model performance was evaluated both in the whole map and the liver region. Results: The generated PDFF was highly consistent with reference qDixon PDFF, and inclusion of DWI improved the PDFF accuracy. Impact: This approach could make PDFF more accessible using standard MRI sequences, enabling liver fat quantification without the need for qDixon imaging. It also facilitates the early detection and monitoring of liver steatosis in routine clinical settings.