Fat‐Water Swap Artifact Correction in MRI‐Based Fat Quantification Using Physics‐Informed Deep Learning With Branch Constrained Optimization

Commonly occurring fat-water swap artifacts in multi-echo MRI hinder accurate estimation of the clinical biomarker, proton density fat fraction (PDFF). This study evaluates a novel physics-informed deep learning (PI-DL) model to correct these swap artifacts. The proposed PI-DL model, the "Swap-Net," utilizes a branch-constrained optimization method to explore different bifurcate regions in the residual function optimized, thereby increasing the probability of reaching global minima, as the swap artifacts are caused by the optimization routine settling into a local minimum. Performance evaluation was conducted using multi-echo MRI of 37 (n = 37) subjects, and this included a cohort of 23 (n = 23) high-resolution MRIs. The proprietary mDixon Quant sequence, with optimized parameter settings for accurate liver fat quantification, served as the clinical standard for benchmarking performance. The results of the quantitative evaluations revealed that the average liver PDFF% from the proposed Swap-Net demonstrated excellent agreement with the clinical standard mDixon Quant (correlation coefficient R of 0.98 in the high-resolution cohort). The qualitative evaluations also showed that the fat-water maps from the Swap-Net contained no swap artifacts. In contrast, other state-of-the-art methods exhibited swap artifacts, particularly in high-resolution MRI. The proposed Swap-Net corrected swap artifacts in MRI-based fat quantification, particularly in high-resolution MRI, enabling high-resolution fat-water imaging.