Abstract
The proposed tiny golden-angle shuffling (TGAS), MRF with optimized spiral-projection trajectory and subspace reconstruction enables high-resolution quantitative mapping in ultrafast acquisition time.
Introduction
Quantitative Magnetic Resonance Imaging (qMRI) is often described as the "holy grail" of neuroimaging. Unlike traditional weighted imaging, which provides qualitative contrast, qMRI yields absolute physical values such as $T_1$ and $T_2$ relaxation times. However, the clinical adoption of qMRI has been hampered by its slow acquisition speeds. Magnetic Resonance Fingerprinting (MRF) was introduced to address this by using pseudo-random acquisitions and pattern matching to accelerate the process. Yet, scaling MRF to 3D whole-brain volumes at high resolution remains challenging due to long scan times, low signal-to-noise ratios (SNR), and artifacts from magnetic field ($B_0$) inhomogeneities.
This paper presents TGAS-SPI-MRF, a framework that combines an optimized "shuffled" spiral trajectory with advanced subspace reconstruction to achieve rapid, high-resolution quantitative brain imaging.
Methodology
The authors introduce two primary innovations to the MRF pipeline to break the trade-off between speed and image quality:
1. Tiny-Golden-Angle-Shuffling (TGAS) Trajectory
Building upon Spiral Projection Imaging (SPI), the researchers developed the TGAS trajectory. While standard SPI methods often follow predictable rotation sequences, TGAS "shuffles" in-plane and through-plane rotations. This technique significantly increases spatiotemporal incoherence, causing undersampling artifacts to appear as pseudo-random noise. This noise-like interference is much more effectively managed by modern reconstruction algorithms than coherent aliasing.
2. Subspace Reconstruction with Locally Low Rank (LLR) Regularization
To manage the massive data volume of MRF, the framework projects the acquisition data into a low-dimensional temporal subspace using Singular Value Decomposition (SVD), specifically utilizing five temporal bases. This reduces the computational complexity while preserving the essential "fingerprint" dynamics. To further suppress noise, the authors applied Locally Low Rank (LLR) regularization, which exploits the fact that small image patches are highly correlated over time.
Additionally, the pipeline incorporates Multi-Frequency Interpolation (MFI) to correct for $B_0$ off-resonance blurring and utilizes pre-scanned $B_1^+$ maps to correct for flip angle inhomogeneities.
Experimental Validation
The study validated the TGAS-SPI-MRF framework using in-vivo data from five healthy volunteers. To demonstrate the method's reproducibility across different hardware environments, scans were performed on both a Siemens 3T Prisma (using 32- and 64-channel coils) and a GE 3T Premier (using a 48-channel coil). The researchers also utilized simulated data to optimize the subspace bases and LLR parameters.
Key Results
The framework demonstrated significant improvements in both speed and spatial resolution:
- Ultrafast 1-mm Isotropic Mapping: The system achieved 1-mm isotropic whole-brain $T_1$ and $T_2$ maps in just 1 minute and 55 seconds. Even at higher acceleration factors ($R=6$), which reduced scan time to approximately one minute, the resulting maps remained highly usable.
- Sub-millimeter Resolution: The authors successfully pushed the resolution to 0.66-mm isotropic in 4 minutes and 19 seconds. This high resolution allowed for the clear visualization of subtle anatomical structures, such as the claustrum and small sulci.
- Accuracy and Robustness: The quantitative values for White Matter (WM) and Gray Matter (GM) were consistent with gold-standard methods and were reproducible across different scanner vendors. Furthermore, the MFI correction effectively removed distortions in regions with high $B_0$ inhomogeneity (up to 150 Hz).
- Hardware Efficiency: The study found that utilizing higher gradient slew rates (150 T/m/s vs. 100 T/m/s) allowed for faster $k$-space traversal, reducing scan times by an additional 25% without compromising image quality.
Discussion and Limitations
While the acquisition is remarkably fast, the current reconstruction process remains a bottleneck. Reconstructing the high-resolution maps takes 3–5 hours on a high-end server and requires approximately 200GB of RAM. The authors suggest that future implementations using deep learning or stochastic optimization could drastically reduce this time.
Additionally, while the MRF scan itself is rapid, the necessary $B_0$ field map currently requires an additional minute of scan time. Future work may focus on "field fingerprinting" or simultaneous calibration to integrate this step into the main sequence. Despite these hurdles, TGAS-SPI-MRF represents a major step toward making high-resolution quantitative whole-brain imaging a practical reality for clinical use.
Summary updated