A k-space-to-image reconstruction network for MRI using recurrent neural network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Neuroscience

Background:

• Magnetic Resonance (MR) image reconstruction from undersampled k-space data is an ill-posed inverse problem.
• Traditional methods often struggle with aliasing artifacts and require iterative processing.

Purpose of the Study:

• To propose a novel method for direct MR image reconstruction from k-space data using a recurrent neural network.
• To develop a unified solution for undersampled k-space data reconstruction applicable to various scanning trajectories.

Main Methods:

• A novel neural network architecture, ETER-net, was developed, integrating bi-recurrent neural networks (bi-RNNs) and convolutional neural networks (CNNs).
• The network performs domain transformation and de-aliasing for image reconstruction.
• Model optimization, cross-validation, and network pruning were conducted using in-house and public datasets (FastMRI).

Main Results:

• ETER-net demonstrated accurate image reconstruction from undersampled k-space data.
• For the in-house dataset (R=4), nMSE was 1.09% and SSIM was 0.938.
• For the FastMRI dataset, nMSE was 1.05% (R=4) and 3.12% (R=8), with SSIM of 0.931 (R=4) and 0.884 (R=8).
• A pruned model maintained performance up to 70% pruning ratio.

Conclusions:

• The proposed method is an end-to-end MR image reconstruction technique based on recurrent neural networks.
• It directly maps k-space data to reconstructed images, serving as a unified solution.
• The method is applicable to diverse scanning trajectories, showcasing its versatility.