HighRes‐Net: Physics‐Informed Deep Learning Framework for High‐Resolution Fat–Water Imaging

ABSTRACT Fat–water separation in 3T scanners is challenging due to the rapid signal loss due to faster dephasing, pronounced field inhomogeneities, and short echo‐spacing requirements that limit high‐resolution fat–water imaging. This study evaluates a novel deep learning (DL) framework, HighRes‐Net, for the challenging task of fat–water separation in high‐resolution single‐shot multi‐echo MRIs acquired at 3T. The proposed HighRes‐Net has DL networks and a biophysical equation–based loss function that are combined specially for handling fat–water separation in low signal‐to‐noise ratio (SNR) high‐resolution multi‐echo MRIs. A novel DL architecture for handling phase errors and the regularized and accurate estimation of field inhomogeneity maps is also introduced. The HighRes‐Net utilizes the ad hoc reconstruction technique using the biophysical equation–based loss that accounts for phase errors. Novel echo‐wise weighting is introduced in this biophysical equation–based loss to account for signal loss in farther echoes. Unlike conventional DL methods, HighRes‐Net employs a subject‐specific ad hoc reconstruction strategy, requiring no large training dataset and avoiding generalizability issues. Efficacy of the HighRes‐Net is evaluated using three types of test datasets, including a high‐resolution dataset acquired at 3T. The liver proton density fat fraction% (PDFF%) values obtained from the test datasets are compared with those obtained from a mDixon Quant sequence run in low‐resolution settings with parameters optimized for accurate fat quantification. The quantitative results demonstrated that the PDFF% values from the HighRes‐Net agreed with those from the mDixon Quant sequence, even in the high‐resolution dataset. The qualitative results demonstrated that the HighRes‐Net produced fat–water maps without swap artifacts. HighRes‐Net enables reliable fat–water separation in high‐resolution 3T multi‐echo MRI, addressing the SNR limitations that hinder conventional techniques and expanding the potential of clinical fat quantification.