Method and system for low-field MRI denoising with a deep complex-valued convolutional neural network

A system for denoising a magnetic resonance (MR) image, comprising:

• one or more processors;
• a memory device coupled to the one or more processors and storing instructions which, when executed by the one or more processors, cause the system to perform functions that include:
• acquiring complex magnetic resonance (MR) image data of an area of interest of a subject, wherein the image data comprises complex noisy input images of multi-coil MR image data, and wherein the complex noisy input images comprise noise data and noise free image data;
• for each of the noisy input images:
• calculating complex single coil data from the complex noisy input images of multi-coil MR image data;
• running an iteration of a complex convolution to a converging sequence with a complex de-noising convolutional neural network (C-DnCNN), wherein running an iteration of the complex convolution comprises running the complex convolution on a real part of the complex single coil data and an imaginary part of the complex single coil data in a single calculation;
• in each iteration, updating parameter settings used in calculating a series of image feature sets with the C-DnCNN, the image feature sets predicting a respective residual image from the complex noisy input images, relative to a corresponding latent clean image;
• and
• removing a final latent clean image from the noisy input images.

A non-transitory computer-readable medium storing instructions thereon which, when executed by one or more processors, cause a computer to perform functions for denoising a magnetic resonance (MR) image that include:

• acquiring complex magnetic resonance (MR) image data of an area of interest of a subject, wherein the image data comprises complex noisy input images of multi-coil MR image data, and wherein the complex noisy input images comprise noise data and noise free image data;
• for each of the noisy input images:
• calculating complex single coil data from the complex noisy input images of multi-coil MR image data;
• running an iteration of a complex convolution to a converging sequence with a complex de-noising convolutional neural network (C-DnCNN), wherein running an iteration of the complex convolution comprises running the complex convolution on a real part of the complex single coil data and an imaginary part of the complex single coil data in a single calculation;
• in each iteration, updating parameter settings used in calculating a series of image feature sets with the C-DnCNN, the image feature sets predicting a respective residual image from the complex noisy input images, relative to a corresponding latent clean image;
• and
• removing a final latent clean image from the noisy input images.