[Paper Review] Deep Learning Methods for Parallel Magnetic Resonance Image Reconstruction

Florian Knöll, Kerstin Hammernik+5 more|arXiv (Cornell University)|Apr 1, 2019

Advanced MRI Techniques and ApplicationsMedicine75 references37 citations

TL;DR

This paper reviews how deep learning and neural networks are applied to parallel MRI reconstruction, covering image-domain and k-space approaches, with emphasis on multi-coil data and learned regularizers/navigation of undersampling artifacts.

ABSTRACT

Following the success of deep learning in a wide range of applications, neural network-based machine learning techniques have received interest as a means of accelerating magnetic resonance imaging (MRI). A number of ideas inspired by deep learning techniques from computer vision and image processing have been successfully applied to non-linear image reconstruction in the spirit of compressed sensing for both low dose computed tomography and accelerated MRI. The additional integration of multi-coil information to recover missing k-space lines in the MRI reconstruction process, is still studied less frequently, even though it is the de-facto standard for currently used accelerated MR acquisitions. This manuscript provides an overview of the recent machine learning approaches that have been proposed specifically for improving parallel imaging. A general background introduction to parallel MRI is given that is structured around the classical view of image space and k-space based methods. Both linear and non-linear methods are covered, followed by a discussion of recent efforts to further improve parallel imaging using machine learning, and specifically using artificial neural networks. Image-domain based techniques that introduce improved regularizers are covered as well as k-space based methods, where the focus is on better interpolation strategies using neural networks. Issues and open problems are discussed as well as recent efforts for producing open datasets and benchmarks for the community.

Motivation & Objective

• Explain the fundamentals of multi-coil parallel MRI and its reconstruction challenges.
• Survey linear and nonlinear traditional methods (SENSE, GRAPPA, SPIRiT) and their limitations with undersampling.
• Introduce and evaluate machine learning frameworks for image-domain and k-space parallel MRI reconstruction.
• Highlight open problems, datasets, and benchmarks to accelerate ML-driven parallel imaging.

Proposed method

• Discuss image-domain vs. k-space formulations of parallel MRI reconstruction.
• Describe nonlinear regularization and compressed sensing as a basis for ML-based methods.
• Present neural-network-inspired unrolled iterative schemes for image-domain reconstruction (variational networks, learned regularizers).
• Explain scan-specific and database-driven k-space interpolation methods (RAKI, DeepSPIRiT, etc.).
• Summarize low-rank/Hankel-based and kernel-based learning approaches (SAKE, LORAKS, ALOHA, etc.).

Experimental results

Research questions

• RQ1 What machine learning strategies can improve parallel MRI reconstruction beyond classical regularizers?
• RQ2 How do image-domain and k-space ML approaches compare in handling undersampling and coil data?
• RQ3 Can learned models achieve artifact suppression with comparable or better quality than CG-SENSE/GRAPPA under various sampling schemes?
• RQ4 What are the practical considerations (data requirements, calibration, computation) for ML-based parallel imaging?
• RQ5 What open datasets and benchmarks exist to evaluate ML methods in parallel MRI?

Key findings

• ML-based image-domain reconstructions can outperform traditional CG-SENSE and TGV-based methods in artifact suppression and feature preservation (higher SSIM in reported comparisons).
• K-space learned interpolation methods (e.g., RAKI, DeepSPIRiT) reduce noise amplification compared with GRAPPA/SPIRiT, though scan-specific calibration introduces trade-offs.
• Unrolled neural networks mirror classic iterative reconstruction, enabling end-to-end learning with data fidelity and learned regularizers.
• Database-trained and scan-specific CNNs can interpolate missing k-space lines without full calibration data, enabling flexible handling of diverse acquisitions.
• Learning-based methods can leverage multi-coil data to achieve improved reconstruction quality and potentially faster inference once trained.