Recent studies have shown that cerebrospinal fluid (CSF) flow is strongly affected by respiration, which may potentially be used for future diagnosis and treatment follow-up in diseases such as normal pressure hydrocephalus and congenital malformations. However, quantitative measurements of respiratory effects on CSF flow in small diameters are not currently available. Therefore, this abstract shows a phantom validation of flow measurement using a radial golden-angle real-time flow sequence, reconstructed using compressed sensing. Results show that mean velocities can be quantified with a small underestimation, suggesting that the protocol is promising for future study of respiratory effects on CSF flow.
Cerebrospinal fluid (CSF) flow is critical to the health of the brain and plays a central role in autoregulation of cerebral blood flow, immunological protection and mechanical cushioning. The cerebral aqueduct (Aq), with a diameter of 2-3 mm, is a central pathway for CSF circulation. Cardiac-gated aqueduct CSF flow has previously been investigated in diseases such as congenital malformations1 and hydrocephalus.2 However, recent results have shown that respiration is the main driving force of CSF circulation. Accurate real-time measurement of aqueduct CSF flow may therefore provide better understanding of several diseases.
Real-time quantification of aqueduct CSF flow presents a significant technical challenge, due to the small diameter and rapid cardiac pulsations. Therefore, the aim of this study was to investigate the performance of a real-time flow imaging protocol based on golden angle radial acquisition and compressed sensing reconstruction in a phantom setup.
Phantom setup
A pump was used to generate pulsatile flow in two plastic tubes, with diameters 3 and 4 mm (Figure 1). The tubes were submerged in a water container (1.5 liters) with approximately one tablespoon of table salt added to improve B1 homogeneity. The pump was run at 32, 46 and 58 beats per minute (bpm).
Imaging protocol
Flow imaging was performed on a Philips Achieva 7T MRI system (Philips Healthcare, Best, The Netherlands) using a dual-transmit 32-channel-receive head coil (Nova Medical, Wilmington, MA, USA). Data was acquired during 60 seconds using a 2D radial gradient-echo sequence with a golden angle increment (111.246°). Each spoke was acquired twice with opposite polarity of the flow encoding gradient (VENC=10 cm/s). The field of view was 240×240 mm, spatial resolution 0.6×0.6×5 mm (matrix size 400×400), TR/TE/flip = 10.5/5.1 ms /4°, and bandwidth 208 Hz/pixel.
For reference, a conventional 2D flow sequence was used, gated to the pump with a temporal resolution of 20 frames per cycle (Figure 1). Field of view was 210x210 mm, spatial resolution 0.3×0.3×3 mm, TR/TE/flip = 11/3.6 ms /7°, and bandwith 421 Hz/pixel.
Reconstruction
Radial raw data, cropped to 20 seconds to speed up reconstruction, was corrected for gradient delays and incoherent phases at the center of k-space. The data was then binned into timeframes with 5, 8 or 13 spokes per frame, corresponding to temporal resolutions of 105, 168 and 273 ms and acceleration factors of R=126, R=79 and R=48, respectively. Coil compression to 8 virtual channels was performed. Compressed sensing reconstruction was performed using the Berkeley Advanced Reconstruction Toolbox3 (BART) v0.4.03 using the Parallel Imaging and Compressed Sensing (PICS) module for the problem
$$ \min_x \left\| F_r S x - d \right\|_2^2 + \lambda \left\| T x\right\|_1. \qquad \textrm{(Eqn. 1)}$$
The first term describes data consistency, and the second term the temporal total variation sparsity constraint. The weighting parameter λ was varied in nine steps: 1.00×10-4, 1.78×10-4, 3.16×10-4, 5.62×10-4, 1.00×10-3, 1.78×10-3, 3.16×10-3, 5.62×10-3, and 1.00×10-2. The number of iterations was set to 500 to ensure convergence.
Reconstructions were also performed without the temporal sparsity constraint using a conjugate gradient SENSE (CG-SENSE) method4, also with 5, 8 and 13 spokes per frame.
Data analysis
Phase background was corrected by the average velocity in a ROI surrounding each vessel (Figure 2). Data was analyzed with respect to a) average velocity and b) velocity range. For the gated 2D flow data, average velocity was computed over the single reconstructed flow period. In the real-time data, average velocity was calculated over the full 20-second reconstructed dataset. Velocity range was computed as the difference between the 5th and 95th percentiles of all data points for both gated and real-time data.