Reference-free Correction for the Nyquist Ghost in Echo-planar Imaging using Deep Learning

Abstract

Echo-planar imaging suffers from Nyquist ghost (i.e., N/2 ghost) because of the imperfection of the gradient system and gradient delays. The phase mismatch between even and odd echoes can be eliminated by an extra reference scan without the phase encoding. However, due to the non-linear and time-varying local magnetic field changes or movement of the patients, the reference-based methods may have incorrect correction results. Other correction methods like parallel imaging reconstruction may suffer from the image noise amplification and signal-to-noise ratio penalty. In this study, a deep learning method is proposed to eliminate the phase error in k-space and correct the mismatch between even and odd echoes without reference scan and SNR penalty. The Fourier transform layer is introduced into the conventional U-Net structure, and the distortion-free images are directly reconstructed from the k-space EPI data. Turbo spin echo data and single-shot EPI data are tested using this network. The results show that this method has a good performance in ghost correction, and the ghost-to-signal ratio is effectively reduced compared to other state-of-the-art correction methods. The proposed deep learning method is reference-free and effective to correct Nyquist ghost in EPI, and can also combine with parallel imaging to achieve additional acceleration.