Domain transformation learning for MR image reconstruction from dual domain input

Area of Science:

• Medical Imaging
• Machine Learning
• Biomedical Engineering

Background:

• Magnetic Resonance Imaging (MRI) relies on k-space data reconstruction for image generation.
• Undersampling k-space data accelerates MRI acquisition but requires advanced reconstruction techniques.
• Traditional inverse Fourier Transform is insufficient for subsampled k-space data.

Purpose of the Study:

• To develop a sophisticated image reconstruction method for undersampled k-space data in MRI.
• To enhance the stability and accuracy of MRI reconstruction from randomly subsampled k-space data.
• To accommodate diverse k-space trajectories and acceleration factors in MRI.

Main Methods:

• A dual-input deep learning network, ETER-net, was trained to learn domain transforms for direct image generation from undersampled k-space data.
• Folded images were incorporated as supplementary inputs to improve reconstruction stability with random subsampling.
• Modifications were made to bi-RNN inputs to handle non-fixed k-space trajectories.

Main Results:

• The dual-input ETER-net demonstrated superior performance in image reconstruction.
• Quantitative metrics including Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Visual Information Fidelity (VIF) showed significant improvements.
• Effective reconstruction was achieved across acceleration factors of 4 and 8, using both regular and irregular sampling trajectories.

Conclusions:

• The dual-input ETER-net is an effective deep learning approach for MRI reconstruction from undersampled k-space data.
• This method enhances image quality and stability, accommodating various sampling patterns and acceleration factors.
• ETER-net offers a promising solution for accelerating MRI acquisition while maintaining diagnostic image quality.