GPU acceleration of nonlinear diffusion tensor estimation using CUDA and MPI

Diffusion MRI is a non-invasive magnetic resonance technique and has been increasingly used in imaging neuroscience. It is currently the only method capable of depicting the complex structure of white matter of the brain in vivo. One of the most popular techniques is diffusion tensor imaging (DTI) which is commonly used clinically to produce in vivo images of biological tissues with local microstructural characteristics such as water diffusion. Diffusion tensor maps are usually computed on a voxel-by-voxel basis by fitting the signal intensities of diffusion weighted images as a function of their corresponding data acquisition parameters (b-matrices). This processing is computation-intensive and time-consuming which can constraint the clinical practice of DTI. This study presents the application of using high performance GPU clusters in addition to CPUs for diffusion tensor estimation by accelerating the multivariate non-linear regression. The results are tested using both simulated and clinical diffusion datasets and show significant performance gain in tensor fitting. The proposed GPU implementation framework can significantly reduce the processing time of DTI data especially for the datasets with high spatial and temporal resolution, and hence further promote the clinical use of DTI. It also can be used to accelerate statistical analysis of DTI where Monte Carlo simulations are employed, be readily applied to quantitative assessment of DTI using bootstrap analysis, robust diffusion tensor estimation and should be applicable to other MR imaging techniques that use univariate or multivariate regression to fit MRI data to a model.