Deep Learning-based Method for Denoising and Image Enhancement in Low-Field MRI

Deep learning has proven successful in a variety of medical image processing applications, including denoising and removing artifacts. This is of particular interest for low field Magnetic Resonance Imaging (MRI), which is promising for its affordability, compact footprint, and reduced shielding requirements, but inherently suffers from low signal-to-noise ratio. In this work, we propose a method of simulating scanner specific images from publicly available, 1.5T and 3T database of MR images, using a signal encoding matrix incorporating explicitly modeled imaging gradients and fields. We apply a stacked, U-Net architecture to reduce noise from the system and remove artifacts due to the inhomogeneous B0 field, nonlinear gradients, undersampling of k-space and image reconstruction to enhance low-field MR images. The final network is applied as a post-processing step following image reconstruction to phantom and human images acquired on a 60-67mT MR scanner and demonstrates promising qualitative and quantitative improvements to overall image quality.