Cardiac magnetic resonance imaging (MRI) has become a crucial part of monitoring patients with congenital heart diseases. An important limitation of cardiac MRI using the prominent 3D steady-state free precession
(3D-SSFP) sequence is its long scan time. Compressed sensing (CS) algorithm reduces the scan time by undersampling the data but increases the image reconstruction time because a non-linear optimization problem must be iteratively solved to estimate the missing data and reconstruct the images.
The growing demand for reducing the examination time in cardiac MRI led us to investigate opportunities to accelerate this non-linear optimization problem to facilitate the migration of CS into the clinical environment.
Using undersampled 3D-SSFP datasets acquired from five patients, we compared the speed and output quality of CS image reconstruction algorithm using a Central Processing Unit (CPU), a CPU with OpenMP parallelization, and two different Graphics Processing Unit (GPU) platforms. Reconstruction time had a mean of 13.1 minutes with a standard deviation of 3.8 minutes for the CPU, a mean of 11.6 minutes with a standard deviation of 3.6 minutes for the CPU with OpenMP parallelization, a mean of 2.5 minutes with a standard deviation of 0.3 minutes for the CPU with OpenMP plus NVIDIA k20m GPU, and a mean of 1.7 minutes with a standard deviation of 0.3 minutes for the CPU with OpenMP plus NVIDIA k40m GPU. The quality of images reconstructed on GPU and on CPU, as assessed by image subtraction, was comparable. Furthermore, necessary steps for implementation of rapid CS image reconstruction in the clinical environment are discussed.
Advisor: Professor Miriam Leeser
Professor Miriam Leeser
Professor Mehdi H. Moghari (Co-Advisor)
Professor Dana Brooks
Dr. Andrew J. Powell