Deep learning corrects artifacts in RASER MRI profiles

A newly developed magnetic resonance imaging (MRI) approach is based on "Radiowave amplification by the stimulated emission of radiation" (RASER). RASER MRI potentially allows for higher resolution, is inherently background-free, and does not require radio-frequency excitation. However, RASER MRI can be "nearly unusable" as heavy distortions from nonlinear effects can occur. In this work, we show that deep learning (DL) reduces such artifacts in RASER images. We trained a two-step DL pipeline on purely synthetic data, which was generated based on a previously published, theoretical model for RASER MRI. A convolutional neural network was trained on 630'000 1D RASER projections, and a U-net on 2D random images. The DL pipeline generalizes well when applied from synthetic to experimental RASER MRI data.