Deep learning improves accuracy of proton-density fat fraction estimation from In-phase and out-of-phase T1-weighted MRI

Proton density fat fraction (PDFF) is an established quantitative-imaging-biomarker for hepatic -fat quantification, but typically requires specialized confounder-corrected chemical-shift-encoded (CSE) magnetic-resonance-imaging (MRI) pulse sequences. We developed and assessed the feasibility of deep learning to infer hepatic PDFF maps from conventional T1-weighted-in-and-opposed-phase (T1w-IOP) MRI. Using PDFF maps reconstructed from CSE-MRI as reference, we trained a convolutional-neural-network (CNN) to infer voxel-wise PDFF maps from T1w-IOP MRI. The CNN was evaluated using both internal and external test datasets. Participant-level median CNN-inferred-PDFF were compared with reference CSE-MRI using linear regression, intraclass correlation, and Bland-Altman analysis. Median CNN-inferred PDFF agreed closely with reference CSE-MRI PDFF.