Compressed sensing in hyperpolarized³He Lung MRI

Abstract

In this work, the application of compressed sensing techniques to the acquisition and reconstruction of hyperpolarized 3 He lung MR images was investigated. The sparsity of 3 He lung images in the wavelet domain was investigated through simulations based on fully sampled Cartesian two-dimensional and three-dimensional 3 He lung ventilation images, and the kspaces of 2D and 3D images were undersampled randomly and reconstructed by minimizing the L1 norm. The simulation results show that temporal resolution can be readily improved by a factor of 2 for two-dimensional and 4 to 5 for three-dimensional ventilation imaging with 3 He with the levels of signal to noise ratio (SNR) (~19) typically obtained. Hyperpolarized (HP) helium-3 gas MRI takes advantage of the nonequilibrium polarization achieved by optical pumping to provide high-resolution images of lung ventilation and function (1-3). The non-renewable longitudinal magnetization is depleted with application of radiofrequency (RF) excitations, and the relationship between the SNR, k-space sampling pattern, and the number of RF pulses is highly flip-angle dependent. Moreover, lung imaging requires rapid sequences to capture dynamic gas flow in the airways and ventilation volume during a breath hold (20 sec). Hence, HP 3 He MRI is a good candidate for undersampling schemes. In previous work, parallel RF encoding (4), radial (5), and spiral (6) methods have been used in human HP gas MRI of the lungs to accelerate the acquisition. These methods can be demanding on the hardware in that they require multiple receivers and implementation of robust non-Cartesian sequences. Recently, compressed sensing (CS) techniques have been applied to MRI acquisition and reconstruction. These methods were developed in the field of information theory by Donoho (7) and Candes et al. (8) and were recently applied to proton MRI (9) and spectroscopic imaging of HP 13 C (10) by Lustig and co-workers. The idea behind the theory is to reconstruct a subset of linear measurements, much smaller than the actual full data set, using a nonlinear method. With CS algorithms, the sparsity of MR images in a specific transform domain can be exploited in order to reconstruct images from undersampled k-space (9). The reconstruction relies on L1-norm minimization in a sparse transform space and the quality of reconstruction relies on the sparsity of the data.In this work, CS theory was applied to the reconstruction of subsampled Cartesian-encoded HP 3 He gas images of the lungs. The potential advantages and limitations of the method were investigated in the context of a potentially faster acquisition time with fewer RF pulses. First, the sparsity of lung images in the wavelet transform domain was investigated in order to validate the possibility of applying the CS method as a means of subsampling. Then, simulations were performed to investigate the feasibility of two-dimensional (2D) and three-dimensional (3D) Cartesian CS undersampling for 3 He lung MRI. The effect of reduction factor upon the quality of...

Mapping of 3He apparent diffusion coefficient anisotropy at sub‐millisecond diffusion times in an elastase‐instilled rat model of emphysema

“…As hyperpolarized 3 He image signal-to-noise ratio is unaffected by reducing repetition time, and imaging speed is determined primarily by hardware (e.g., gradients), improvements will likely be possible in the future by speeding up imaging to obtain multiple b value images in a single breath-hold. For example, parallel imaging (33) can give an acceleration up to a factor of six with no significant loss in signal-to-noise ratio, perhaps allowing multiple b values in a single breath hold (34).…”

Iterative reconstruction of radially-sampled 31 P bSSFP data using prior information from 1 H MRI

“…Ultra-high field MRS could improve those investigations conspicuously [11][12][13][14]. Also the feasibility of direct 31 P mapping has been demonstrated although the spatial resolution was limited. To overcome the low in vivo SNR per unit time, several approaches were proposed including turbo spin echo (TSE) [15][16][17][18], fast low angle shot (FLASH) [19,20], or balanced steady-state free precession (bSSFP) [21,22] sequences.…”

“…However, promising results have been obtained by means of subsampling methods such as compressed sensing [29,30] applied in recent non-proton MRI studies [31,32]. In MRI, compressed sensing is employed for reconstructing undersampled images while reducing Magnetic Resonance Imaging 37 (2017) 147-158 incoherent undersampling artifacts.…”

“…Conceptual workflow to obtain iteratively reconstructed31 P images. (A) Contrasts of 31 P acquisitions obtained by 3D density-adapted radially-sampled bSSFP sequences.…”

A flexible 32‐channel receive array combined with a homogeneous transmit coil for human lung imaging with hyperpolarized 3He at 1.5 T

“…The reduction in acquisition time provided by accelerated imaging techniques which use reduced radiofrequency (RF) encoding, such as parallel imaging (2)(3)(4)(5) and compressed sensing (6,7), is of particular interest for hyperpolarized gas MRI, as the finite lifetime of hyperpolarization and the duration of breath-hold pose fundamental temporal limits to a hyperpolarized gas pulse sequence acquisition.…”