Compressed sensing based magnetic resonance image reconstruction using deep learning

Abstract

Magnetic resonance imaging (MRI) is an increasingly versatile diagnostic tool for various medical purposes. It is a non-invasive imaging modality that is not harmful for the patients unlike X-rays, Computed Tomography (CT) etc. The data acquisition speed of MRI is an important issue. In MRI, the samples are acquired along a trajectory called k-space and the image is reconstructed using an inverse Fast Fourier transform (IFFT) of the acquired k-space. Some methods have been recently proposed to accelerate the data acquisition time in MRI e.g. Parallel MRI (pMRI), Compressed Sensing (CS) etc. Parallel MR Imaging(pMRI) is a robust method for accelerating MRI data acquisition. It involves the use of several receiver coils with distinct spatial sensitivities to acquire a reduced amount of k-space data.Compressed Sensing is another promising technique for fast MR image reconstruction from the highly under-sampled data. It is beneficial in MRI because MR images are sparse or can be made sparse in a known transform domain. The aim of this research is to used combined applications of compressed sensing and deep learning to accelerate and reconstruct high quality MR data from the under-sampled k-space. Variational Networks(VNs) (a Convolutional Neural Network), which is a combination of variational model and deep learning, are used for reconstruction of undersampled MRI data. Mathematical structure of VN comprised of compressed sensingbased reconstruction algorithm embedded in iterative gradient descent scheme. VNs are trained model based on framework of optimization algorithm known as incremental proximal gradient methods. VN has ability to learn whole reconstruction process and parameters of complex MR data including filter kernels, activation function and regularization terms are learned through offline training process. Reconstruction through VN outperforms traditional reconstruction techniques and offers high reconstruction speed.