(A) neural network based artifact correction method for echo planar imaging

Abstract

We proposed a new correction method scheme for Nyquist ghost artifacts and susceptibility induced artifacts in Echo Planar Imaging (EPI) using a neural network with unsupervised learning. The overall scheme is composed of two parts: the first part is for Nyquist ghost artifact correction and the other is for susceptibility induced artifact correction. Each part uses an individual neural network. First, the Nyquist ghost artifact correction neural network estimates the phase error in k-space of obtained EPI scans by using a ghost generation function and correlation coefficient loss. The ghost generation function uses the obtained EPI scan and estimated phase error as inputs and gives Nyquist ghost corrected image as outputs. Two distorted images obtained with dual-polarity phase-encoding gradients are the inputs of the first neural network, producing two Nyquist ghost corrected images. The second neural network takes in the two output images from the previous neural network. The second neural network is trained with an unsupervised learning method that minimizes the L1 loss between the two input images of opposite polarities, which corresponds to finding the frequency-shift map. Then an MR image generation function utilizes the trained neural network and Nyquist ghost corrected images of opposite polarities to obtain the susceptibility induced artifact corrected images. We proved that the neural network trained without MR data successfully corrects the distortion of in-vivo MR data acquired with EPI with dual-polarity phase-encoding gradient.