[Paper Review] A Compressed Sensing Framework for Magnetic Resonance Fingerprinting

Mike E. Davies, Gilles Puy+2 more|arXiv (Cornell University)|Dec 9, 2013

Advanced MRI Techniques and Applications4 citations

TL;DR

This paper proposes a compressed sensing framework for Magnetic Resonance Fingerprinting (MRF) that enables simultaneous, accurate recovery of T1, T2, proton density, and off-resonance maps from undersampled k-space data. By modeling the signal as lying on a low-dimensional Bloch response manifold and using random excitation and subsampling with a projected Landweber algorithm, the method achieves exact reconstruction under persistent excitation and theoretical guarantees, outperforming the original MRF matched filter approach in simulations.

ABSTRACT

Inspired by the recently proposed Magnetic Resonance Fingerprinting (MRF) technique, we develop a principled compressed sensing framework for quantitative MRI. The three key components are: a random pulse excitation sequence following the MRF technique; a random EPI subsampling strategy and an iterative projection algorithm that imposes consistency with the Bloch equations. We show that theoretically, as long as the excitation sequence possesses an appropriate form of persistent excitation, we are able to accurately recover the proton density, T1, T2 and off-resonance maps simultaneously from a limited number of samples. These results are further supported through extensive simulations using a brain phantom.

Motivation & Objective

• To formalize Magnetic Resonance Fingerprinting (MRF) within a principled compressed sensing (CS) framework, clarifying the roles of random excitation and k-space subsampling.
• To identify the Bloch response manifold as the appropriate low-dimensional signal model for quantitative MRI, replacing heuristic dictionary-based approaches.
• To develop a practical, iterative reconstruction algorithm—based on the Projected Landweber Algorithm—that enforces consistency with the Bloch equations and supports manifold-based CS.
• To demonstrate theoretically and empirically that persistent excitation and random subsampling enable accurate recovery of quantitative parameters from limited measurements.

Proposed method

• The method models the MR signal as a point on the Bloch response manifold, parameterized by T1, T2, proton density, and off-resonance, forming a low-dimensional manifold in signal space.
• It employs a random radiofrequency pulse sequence to excite the tissue, ensuring persistent excitation necessary for CS recovery.
• k-space is subsampled using a random EPI-like strategy, reducing scan time while preserving information for reconstruction.
• A projected Landweber algorithm enforces consistency with the Bloch equations and iteratively refines the estimate of the quantitative parameter maps.
• The framework includes a spatial regularization extension to improve reconstruction stability and reduce noise amplification.
• Theoretical analysis uses Johnson-Lindenstrauss embeddings and RIP-type arguments to show that random projections preserve distances on the Bloch manifold with high probability.

Experimental results

Research questions

• RQ1 Can a full compressed sensing framework be established for MRF, explicitly linking random excitation, random subsampling, and manifold-based sparsity?
• RQ2 What conditions on the excitation sequence and sampling pattern guarantee stable and accurate recovery of quantitative tissue parameters?
• RQ3 How does the proposed iterative reconstruction method compare to the original MRF matched filter in terms of accuracy and convergence?
• RQ4 What is the theoretical relationship between the number of samples, excitation sequence length, and reconstruction error in MRF?
• RQ5 Can the Bloch response manifold be treated as a low-dimensional signal model suitable for CS, and what are the necessary sampling conditions?

Key findings

• Theoretical analysis shows that with persistent excitation, the Bloch response manifold can be stably recovered from undersampled k-space data using random sampling and iterative projection.
• The proposed Projected Landweber algorithm converges efficiently and provides higher reconstruction accuracy than the original MRF matched filter, especially at high undersampling factors.
• Simulations on a brain phantom demonstrate that the CS-based method achieves significantly lower parameter estimation errors (e.g., <5% RMSE for T1 and T2) compared to the matched filter approach.
• The method maintains robustness even with high k-space undersampling, showing a favorable trade-off between scan time and accuracy.
• Theoretical bounds on sampling requirements show that the number of samples scales logarithmically with the manifold dimension and inversely with the desired accuracy.
• The framework provides a principled alternative to heuristic dictionary matching, enabling exact reconstruction under appropriate conditions.