Addressing Bias and Precision in Low SNR Chemical Shift Encoded MRI Proton Density Fat Fraction Estimation using a Deep Learning Reconstruction

Low bias and high precision are important for accurate diagnosis, staging, and treatment monitoring of chronic liver disease using chemical shift-encoded (CSE)-MRI. However, CSE-MRI proton density fat fraction (PDFF) measurements are often biased by an asymmetric noise distribution present in PDFF maps acquired with low/moderate signal-to-noise ratio (SNR). This work investigates the use of deep learning de-noising to mitigate this bias in phantoms and in vivo. Results demonstrate that deep learning reconstruction removes or reduces noise-related PDFF estimation bias while maintaining the expected noise distribution characteristic of PDFF.