Abstract
In this survey, we provide a detailed review of recent advances in the recovery of continuous domain multidimensional signals from their few non-uniform (multichannel) measurements using structured low-rank matrix completion formulation. This framework is centered on the fundamental duality between the compactness (e.g., sparsity) of the continuous signal and the rank of a structured matrix, whose entries are functions of the signal. This property enables the reformulation of the signal recovery as a low-rank structured matrix completion, which comes with performance guarantees. We will also review fast algorithms that are comparable in complexity to current compressed sensing methods, which enables the application of the framework to large-scale magnetic resonance (MR) recovery problems. The remarkable flexibility of the formulation can be used to exploit signal properties that are difficult to capture by current sparse and low-rank optimization strategies. We demonstrate the utility of the framework in a wide range of MR imaging (MRI) applications, including highly accelerated imaging, calibration-free acquisition, MR artifact correction, and ungated dynamic MRI.
Introduction: The Challenge of MRI Speed
Magnetic Resonance Imaging (MRI) is fundamentally constrained by the sequential nature of data collection in the Fourier domain (k-space). While Compressed Sensing (CS) has significantly advanced the field by enabling reconstruction from undersampled data, it often relies on discrete sparsity models. These models can suffer from discretization errors and struggle with complex multi-channel or dynamic motion data. This paper reviews the Structured Low-Rank (SLR) framework, which addresses these limitations by operating in the continuous domain.
The Structured Low-Rank Framework
The core of the SLR approach is the concept of "Lifting." Instead of processing the signal directly, it is transformed into a large, structured matrix (e.g., Hankel or Toeplitz). This framework exploits the fundamental duality between a signal's compactness (sparsity) and the rank of its lifted matrix.
Lifting and Matrix Completion
In the lifted domain, signal recovery is reformulated as a structured matrix completion problem. Since the lifted matrix is low-rank, missing k-space data can be recovered using techniques like nuclear norm minimization.
Annihilation Relations and FRI
The sparsity of a signal in the spatial domain translates to "annihilation relations" in the Fourier domain. Specifically, there exists a filter that, when convolved with the signal's Fourier samples, results in zero. This Finite Rate of Innovation (FRI) property allows for "off-the-grid" recovery, reducing artifacts associated with fixed pixel grids.
Computational Efficiency
To handle large-scale 3D or 4D datasets, the authors highlight fast algorithms like GIRAF (Generic Iterative Reweighted Annihilating Filter). These methods avoid the memory overhead of explicitly constructing giant lifted matrices by performing operations in the original signal domain using Fast Fourier Transforms (FFTs).
Applications and Clinical Impact
The SLR framework has been validated across diverse MR applications:
- Static Brain MRI: Demonstrating recovery from highly undersampled radial and Cartesian grids.
- Multi-shot Diffusion MRI: Correcting phase errors and "ghosting" artifacts (e.g., via the MUSSELS algorithm) caused by physiological motion.
- Dynamic Cardiac MRI: The SToRM algorithm enables high-quality reconstruction from free-breathing, ungated acquisitions by mapping images onto a smooth manifold.
Theoretical Insights and the Machine Learning Connection
A key insight of this survey is the mathematical equivalence between spatial sparsity and low-rankness in lifted k-space. Furthermore, the authors establish a link to modern AI, suggesting that Deep Convolutional Neural Networks (CNNs) can be viewed as combinatorial versions of SLR framelet representations. The ReLU activation function acts as a "basis selector," providing a theoretical basis for the effectiveness of deep learning in adaptive image reconstruction.
Conclusion and Future Directions
While SLR offers powerful, calibration-free recovery with theoretical efficiency comparable to CS, challenges remain. Sensitivity to sampling patterns and robustness to high noise levels are areas for future research. Additionally, bridging the gap between the well-defined convex optimization landscape of SLR and the complex landscape of deep learning models remains a significant open direction.
Summary updated