Learning from synthetic data for reference-free Nyquist ghost correction and parallel imaging reconstruction of echo planar imaging.

ABSTRACT

BACKGROUND: Echo planar imaging (EPI) suffers from Nyquist ghost caused by eddy currents and other non-ideal factors. Deep learning has received interest for EPI ghost correction. However, large datasets with qualified labels are usually unavailable, especially for the under-sampled EPI data due to the imperfection of traditional ghost correction algorithms. PURPOSE: To develop a multi-coil synthetic-data-based deep learning method for the Nyquist ghost correction and reconstruction of under-sampled EPI. METHODS: Our network is trained purely with synthetic data. The labels of the training samples are generated by combining a public magnetic resonance imaging dataset and a few pre-collected coil sensitivity maps. The input is synthesized by under-sampling (for the accelerated case) and adding phase errors between the even and odd echoes of the label. To bridge the gap between synthetic data and data from real acquisition, linear and non-linear 2D phase errors are considered during the training data generation. RESULTS: The proposed method outperformed the existing mainstream approaches in several experiments. The average ghost-to-signal ratios of our/3-line navigator-based methods were 0.51%/5.36% and 0.42%/8.64% in fully-sampled and under-sampled in vivo experiments, respectively. In the sagittal experiments, our method successfully corrected higher-order and 2D phase errors. Our method also outperformed other reference-based methods on motion-corrupted data. In the simulation experiments, the peak signal-to-noise ratios were 37.6/38.3 dB for 2D linear/non-linear simulated phase errors, indicating that our method was consistently reliable for different kinds of phase errors. CONCLUSION: Our method achieves superb ghost correction and parallel imaging reconstruction without any calibration information, and can be readily adapted to other EPI-based applications.