In Vivo Evaluation of a Novel Deep Learning-based MR Image Reconstruction for Liver Fat Quantification

Deep learning (DL)-based MR reconstruction methods show promise to reduce image noise compared to conventional image reconstruction while maintaining quantitative accuracy. The purpose of this work is to apply and evaluate the performance of DL reconstruction to chemical shift-encoded MRI for quantification of proton density fat-fraction (PDFF) in the liver. We compared quantitative PDFF results, test-retest repeatability, and standard deviation within regions of interest in nine patients with a wide range of PDFF (1-31%), for different levels of DL denoising. PDFF between reconstructions showed excellent agreement and constant test-retest repeatability. Standard deviation decreased with increasing DL denoising levels.