Compressed Sensing 3-Dimensional Time-of-Flight Magnetic Resonance Angiography for Cerebral Aneurysms: Optimization...

The aims of this study were to optimize parameters for Nesterov algorithm (NESTA) in reconstruction of 3-dimensional time-of-flight (TOF) magnetic resonance angiography (MRA) at 3 T by performing an exhaustive search and to validate the performance of compressed sensing (CS) by applying it to data from cerebral aneurysms and evaluating diagnostic quality.

Three-dimensional TOF-MRA was obtained using a 3 T MR system with a 32-channel head coil for both healthy volunteers and 10 patients (11 aneurysms). No undersampling was applied for imaging parameters, including parallel imaging or other partial Fourier sampling. In the first step, the experimental setup was for healthy subjects to optimize CS parameters of NESTA and the undersampling mask pattern, so 24,696 different reconstruction conditions were surveyed for sampling rates of 8.0X and 5.0X. Mean square error (MSE) was calculated for each image reconstructed with the undersampling pattern and CS parameter sets. Evaluation was by normalized MSE, edge sharpness for MRA reconstructed using fully sampled data (MRA-full), zero-filled MRA (ZF-MRA) with Poisson disk undersampling mask, and CS-MRA (5.0X and 8.0X) with iterations of 5, 10, 15, 20, 25, 30, 35, 40, 45, and 50. CS-MRA (5.0X and 8.0X) with 5, 10, and 50 iterations of the sampling pattern and CS parameter set with the lowest MSE were visually inspected by 2 neuroradiologists to check the diagnostic quality.

The sampling pattern and CS parameter set with the lowest MSE were identical for both CS-MRA 5.0X and CS-MRA 8.0X. At the initial 5 to 15 iterations, MSE of both sampling rates greatly decreased from that of ZF-MRA. For subsequent iterations, the decrease in MSE was relatively small. For CS-MRA, sharpness greatly increased from that of ZF-MRA within the initial 5 to 15 iterations, followed by slight increases with further iterations. Two neuroradiologists graded most aneurysms as excellent, with the exception of 1 to 4 aneurysms recognized as good by 1 observer in CS-MRA (8.0X).

Optimization of NESTA in the reconstruction of 3-dimensional TOF-MRA was conducted, and the parameters and undersampling mask with the lowest MSE were determined. Caliber measurement should be performed with CS (5.0X) with 25 or 30 iterations. Most cerebral aneurysms were sufficiently recognized using CS-MRA (5.0X) or CS-MRA (8.0X) with 10 iterations.