Combination of Iterative Reconstruction and CNN-Based Denoising for Non-Uniform Noise for Parallel Imaging in MRI

Abstract

Parallel imaging in magnetic resonance imaging produces non-uniform noise when the acceleration factor is high. To reduce non-uniform noise, we developed a method that combines iterative reconstruction with soft-thresholding and CNN-based denoising. Non-uniform noise is reduced by iterative reconstruction, and further reduced by combining iterative reconstruction with CNN-based denoising. We applied this method to parallel imaging of the heads of six volunteers and confirmed that our method was able to reduce non-uniform noise in reconstructed images with an acceleration factor of 3, where the g-factor is high. Compared with a conventional reconstruction with CNN-based denoising for an acceleration factor of 3, our method improved the structural similarity index measure values relative to the reference images for T2-weighed, T1-weighted, and FLAIR images from 0.922/0.929/0.864 to 0.932/0.938/0.886, respectively (P < 0.05). Therefore, our method contributes to improving image quality with a shorter scan time.