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

Area of Science:

• Magnetic Resonance Imaging (MRI)
• Medical Imaging
• Artificial Intelligence in Medicine

Background:

• Echo Planar Imaging (EPI) is prone to Nyquist ghosts due to eddy currents and other factors.
• Deep learning shows promise for EPI ghost correction, but lacks large, labeled datasets, especially for under-sampled data.
• Existing ghost correction algorithms have imperfections, limiting their effectiveness for under-sampled EPI data.

Purpose of the Study:

• To develop a multi-coil, synthetic-data-based deep learning method for Nyquist ghost correction.
• To enable accurate reconstruction of under-sampled EPI data.
• To address the challenge of limited labeled datasets in EPI ghost correction.

Main Methods:

• A deep learning network trained exclusively on synthetic data.
• Generation of training labels by combining public MRI data with coil sensitivity maps.
• Synthesis of input data through under-sampling and addition of phase errors between even and odd echoes.
• Inclusion of linear and non-linear 2D phase errors in training data to bridge the synthetic-real data gap.

Main Results:

• The proposed method significantly outperformed mainstream approaches in ghost correction.
• Achieved average ghost-to-signal ratios of 0.51% (fully-sampled) and 0.42% (under-sampled) in vivo.
• Successfully corrected higher-order and 2D phase errors in sagittal EPI, and outperformed reference methods on motion-corrupted data.
• Demonstrated consistent reliability across different phase errors in simulation experiments (37.6/38.3 dB PSNR).

Conclusions:

• The developed method achieves excellent ghost correction and parallel imaging reconstruction.
• No calibration information is required, simplifying its application.
• The method is adaptable to various EPI-based applications.