Cardiac MRI compressed sensing image reconstruction with a graphics processing unit

Abstract

Compressed sensing (CS) magnetic resonance imaging (MRI) reconstruction 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 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 3D steady-state free precession MRI images from 5 patients, we compared the speed and output quality of CS reconstruction using central processing unit (CPU), CPU with OpenMP parallelization, and graphics processing unit (GPU) platforms. Mean reconstruction time was 13.1 ± 3.8 minutes for the CPU, 11.6 ± 3.6 minutes for the CPU with OpenMP parallelization, and 2.5 ± 0.3 minutes for the CPU with OpenMP plus GPU. GPU and CPU reconstructed image quality as assessed by image subtraction were comparable. Additional developments needed for implementation of rapid CS image reconstruction in the clinical environment are discussed.

Estimation of Spatiotemporal Sensitivity Using Band-limited Signals with No Additional Acquisitions for <i>k−t</i> Parallel Imaging

“…To reduce the reconstruction time for CS, a graphics processor unit (GPU) implementation has also been evaluated. 22 A common LRM solver is a nuclear norm optimizer that uses a scheme such as iterative singular value thresholding. For CS/LRM, an iterative solver such as a conjugate gradient solver is generally used for evaluation of the data consistency of PI.…”

A survey of GPU-based acceleration techniques in MRI reconstructions

“…Their experiments employed 3D steady-state free precession MRI images from five patients, and compared the speed and recon image quality on different parallel platforms, such as CPU, CPU with OpenMP, and GPU. Their recon results showed that the mean reconstruction time was 13.1±3.8 minutes on the CPU platform, 11.6±3.6 minutes on the CPU platform with OpenMP, and 2.5±0.3 minutes for the CPU platform with OpenMP plus GPU (40). And their image qualities estimated by image subtraction were very similar, which are comparable on different parallel architectures.…”

“…Because the NVIDIA CUDA library is more and more perfect to support for GPU computing, the complex sparse reconstruction methods are more easily implemented on GPUs without considering the hardware constrains on GPUs. Actually, there are recently some papers studying CS MR recon methods on GPU architectures (35)(36)(37)(38)(39)(40)(41)(42). For example, Zhuo et al presented a GPU-accelerated regularization reconstruction method with compensations for susceptibility-induced field inhomogeneity effects, which are incorporating a quadratic regularization term (35).…”

“…Otherwise, there are also some other papers studying CS MR recon methods on several parallel architectures of multicore CPUs and multi-core GPUs (39)(40)(41). They presented the huge potential speed-up ability on the architectures.…”

“…All their experiments showed that their CS algorithm could bring huge benefits from those throughput-oriented architectures. Apart from that, Sabbagh et al studied to accelerate the non-linear CS reconstruction problem in cardiac MRI solved by iterative optimization algorithms and to facilitate the migration of CS reconstruction in the clinical applications (40). Their experiments employed 3D steady-state free precession MRI images from five patients, and compared the speed and recon image quality on different parallel platforms, such as CPU, CPU with OpenMP, and GPU.…”