Parallel MRI Reconstruction Algorithm Implementation on GPU

Abstract

Magnetic resonance imaging (MRI) is a safe, non-ionizing and powerful diagnostic imaging modality and has a large number of variable contrast mechanisms. There is a fundamental limit in MRI data collection time which can be overcome by using parallel imaging algorithms, e.g., SENSE. Graphical processing units (GPUs) using compute unified device architecture have great potential to reduce the scan time by exploiting the inherent parallelism present in parallel imaging algorithms for MR image reconstruction. This work implements SENSE algorithm using GPU and compares the results with multi-core CPU implementation of SENSE. The inversion of the encoding matrix (formed from the under-sampled data) is a key process in SENSE. The encoding matrix is usually rectangular because the number of receiver coils need to be greater than the acceleration factor. This paper implements the inversion of the rectangular matrix on GPU using Left Inverse Method. All the scripts are written by the authors for this implementation of SENSE on GPU. The results show that GPU attains approximately 7× ~ 28× reduction in SENSE reconstruction time as compared to CPU while maintaining the image quality.